关于鸟的脑神经系统的研究资源
A resource for brain researchers
Profile: Erich Jarvis
The work of neuroscientist Erich Jarvis demonstrates the power of open-mindedness in the lab.
http://www.pbs.org/wgbh/nova/sciencenow/3214/03.html
Profile: Erich JarvisLinksThe Jarvis Lab Avian Brain Nomenclature Exchange The Life of Birds Bird Brains Modern Bird Anatomy Bird Mag Dot Com Inside the Animal Mind BooksBird Brains: The Intelligence of Crows, Ravens, Magpies, and Jays The Parrot's Lament: And Other True Tales of Animal Intrigue, Intelligence, and Ingenuity Inside the Animal Mind: A Groundbreaking Exploration of Animal Intelligence Recent Article"Minds of Their Own: Birds Gain Respect" http://www.jarvislab.net/index.html
|
发布于12月10日 0:00 | 评论数(15) 阅读数(1743) | 我的文章
"Birdbrain" No Longer Means "Stupid," Asserts Scientific Consortium
 |  | |
| | |
|
Note: An accompanying video interview with Erich Jarvis can be viewed here. DURHAM, N.C. -- An international consortium of 29 neuroscientists has proposed a drastic renaming of the structures of the bird brain to correctly portray birds as more comparable to mammals in their cognitive ability. The scientists assert that the century-old traditional nomenclature is outdated and does not reflect new molecular, genetic and behavioral studies that reveal the brainpower of birds. For example, they identified behavioral studies demonstrating that pigeons can discriminate cubist from impressionistic styles of painting; that crows can make useful tools and pass on their skills to other birds, and that parrots can not only learn human words but use them to communicate with humans. The researchers emphasize that the old view of evolution as progressive and linear is outdated, pointing out that so-called "primitive" animals such as birds evolved some 50 to 100 million years after mammals. The Avian Brain Nomenclature Consortium published a report on the rationale for the proposed revised nomenclature in the February 2005 issue of Nature Reviews Neuroscience. A technical report detailing the revisions was published in the May 2004 issue of the Journal of Comparative Neurology. The consortium's efforts were supported by the National Institutes of Health and the National Science Foundation, including the NSF's Waterman Award for young researchers to the Nature Reviews Neuroscience paper's first author, Duke University Medical Center neurobiologist Erich Jarvis. "We believe that names have a powerful influence on the experiments we do and the way in which we think," wrote the consortium members in their paper. "For this reason, and in the light of new evidence about the function and evolution of the vertebrate brain, the international consortium of neuroscientists has reconsidered the traditional 100-year-old terminology that is used to describe the avian cerebrum. "Our current understanding of the avian brain -- in particular the neocortex-like cognitive functions of the avian pallium -- requires a new terminology that better reflects these functions and the homologies between avian and mammalian brains." The consortium members asserted that the old terminology -- which implied that the avian brain was more primitive than the mammalian brain -- has hindered scientific understanding. They concluded that "The inaccurate evolution-based terminology for the vertebrate brain that was used throughout the twentieth century became a severe impediment to the communication of scientific discoveries and the generation of new insights." The consortium's revision of the nomenclature for avian brains is aimed at replacing the century-old system developed in the 19th century by Ludwig Edinger, considered the father of comparative neuroanatomy. Edinger's system was based on the then-common practice of combining Darwin's recent theory of evolution and Aristotle's old concept that there exists a natural "scale" of creatures from lowest to highest. The result were the views that evolution was progressive from organisms with "lower" intelligence to those with "higher" intelligence and that evolution had a purpose -- the generation of humans. The resulting nomenclature used prefixes such as palaeo- ("oldest") and archi- ("archaic") to designate structures in the avian brain and neo- ("new") to designate supposedly new structures, particularly in the mammalian brain. "According to this theory, the avian cerebrum is almost entirely composed of basal ganglia, the basal ganglia is involved only in instinctive behavior, and the malleable behavior that is thought to typify mammals exclusively requires the so-called neocortex," wrote the researchers. However, said Jarvis, "We have to get rid of the idea that mammals -- and humans in particular -- are the pinnacle of evolution. We have to stop using words like 'lower vertebrates' and 'higher vertebrates.' We also have to understand that evolution is not linear, but an intricate branching process. So, we can't automatically expect to track a structure in the human brain back to other current vertebrate species." According to Jarvis, new research "debunks the theory that the brain evolved in stages, like the laying down of geological sediments layer by layer. There is no evidence to show that there was a primordial brain structure to which so-called higher brain structures were systematically added." In the Nature Reviews Neuroscience paper, the authors described studies by other researchers and their own studies demonstrating that the so-called "primitive" regions of avian brains were actually sophisticated processing regions homologous to those in mammals. Those studies, which included tracing of neural pathways and behavioral studies, showed that such avian brain regions carried out sensory processing, motor control and sensorimotor learning just as did the mammalian neocortex. Also, wrote the scientists, molecular studies have shown that the avian and mammalian brain regions are comparable in their genetic and biochemical machinery. The neocortex and related areas in the mammalian brain are derived from a region in the embryonic cerebrum called the pallium, which means mantle or covering. Edinger thought, however, that most of this region in the bird cerebrum was part of the basal ganglia. Accordingly, he gave them names that ended in the basal ganglia term "-striatum", a practice he also employed in naming the parts of the mammalian basal ganglia. As a result of the recent studies, the consortium has recommended such changes as renaming the avian brain region called the "archistriatum" as the "arcopallium," (arched pallium); and renaming the region that includes part of the true basal ganglia in birds, the "palaeostriatum primitivum" and the "ventral palaeostriatum" which sits below the pallium as the "pallidum" (pallidal or pale domain). The consortium's work began in 1997 and was organized by Jarvis, Anton Reiner of the University of Tennessee Health Science Center in Memphis, Martin Wild of the University of Auckland in New Zealand, and other neurobiologists, dubbing themselves the ThinkTank. Jarvis recalled that "there were people in the field of avian neurobiology who knew the real structures behind these names and knew the names were wrong. And as a member of the younger generation of neurobiologists, I just felt that it was against my conscience to continue to use terminology that I knew was wrong and would mislead scientists." For example, said Jarvis, researchers not familiar with the growing body of scientific literature demonstrating the sophistication of the avian brain could not understand how birds could exhibit sophisticated cognitive abilities with brains that held only what the nomenclature designated as the equivalent of the human basal ganglia. The result of the scientists' objections led to a seven-year effort, which steadily recruited new participants. This effort culminated in an intensive international scientific forum at Duke in 2002, in which the new nomenclature was developed. "We knew that we were doing something that may have an impact, not only on the immediate conduct of research in neuroscience, but on neuroscience for the next hundred years," said Jarvis. "And, this nomenclature will help people understand that evolution has created more than one way to generate complex behavior -- the mammal way and the bird way. And they're comparable to one another. In fact, some birds have evolved cognitive abilities that are far more complex than in many mammals." Besides Jarvis, other co-authors of the paper were
| ||||||||||||||||||
http://news.mc.duke.edu/news/article.php?id=8401
Bird Talk: Probing the Avian Instrument
Tweet Mystery of Life
(Originally published in the July-August 1994 issue)
To fathom the intricacies of bird song, a Duke zoologist has concocted experiments using paraphernalia that range from the most sophisticated--soundproof chambers, videotape recorders, audiotape players, and computers--to the humblest--lollipop sticks, rubber bands, and helium.
Steve Nowicki is about to become a father. Before dawn on a spring morning, he and his graduate students rise from warm beds in the rustic cabin, shivering and sipping hot coffee in the chilly darkness. They don layers of clothing against the cold and sleepily pull on stiff hip waders, emerging from the cabin as the sun rises dimly over a still lake. They drive bouncing and swerving along rutted back roads in the damp Pennsylvania woods until they reach the swamp. Swatting at breakfast-bent mosquitoes, bracing themselves against the cold water, they wade into the slough, binoculars at the ready, ears cocked for the faintest sound.
With luck, over the next hours they will spot an evanescent flash of brown feathers or hear a tell-tale twitter that marks their quarry--a swamp sparrow carrying food. If they are luckier still, they will pinpoint a sparrow's nest of newly hatched baby birds. Gently, the will ease both nest and birds into a cloth bag and return to their cabin. There, Nowicki and his students will become the birds' fathers and mothers, every half-hour faithfully feeding the baby birds "meat glop"--a health-giving mix of ground sirloin, tofu, baby carrots, vitamins, and minerals.
Thus does Nowicki, a Duke associate professor of zoology, obtain the fascinating animals whose complex trills, warbles, and chirps, and exquisitely pure tones he seeks to decipher. He and his colleagues concoct experiments using paraphernalia that range from the most sophisticated--soundproof chambers, videotape recorders, audiotape players, and computers--to the humblest--lollipop sticks, rubber bands, and helium. From those experiments have come intriguing insights into the talents of nature's most accomplished musicians.
What's more, Nowicki's own fine-tuned teaching abilities have brought students flocking to his rigorous undergraduate courses on neuroscience and animal communication. An ex-trombonist and adept juggler, he has developed his own brand of scientific street theater to jump-start students' minds and drive home intellectual points.
To illustrate a discussion of auditory processing, he's had a rock band perform. To make points about evolution and the mind, he's picked arguments with actors dressed as Darwin, Freud, and a giant frog who invaded his class. Rumor has it that, to illustrate the fallibility of perception, he even showed up for class once dressed in drag. (He points out proudly that his wife, Susan Peters, also a professional biologist, expertly sewed the frog's head. He is mum about where he got his dress.) Because of his determination to communicate, his clear compelling lectures, his imaginative exams, and his commitment to his students, he was awarded the 1992-1993 Robert B. Cox Trinity College Distinguished Teaching Award.
His baby birds even like him. After all, they luxuriate in a sort of sparrow spa in his laboratory. They get free food, a comfortable soundproof chamber, and free singing lessons via tape recordings. In fact, other scientists' studies of the solitary birds find that they experience lower stress and longer lives in captivity than in the wild. (Incidentally, nor are the swamp-sparrow parents particularly upset at the lost of the nest and its nestlings. They typically immediately rebuild the next and lay more eggs.)
To Nowicki, songbirds represent stunningly complex examples of animal communication. A typical bird song, lasting about two seconds, is a rapid-fire aria of fifty or more "notes," each as short as ten thousandths of a second. The bird can rattle off these notes up to five times faster than a human can speak syllables. Besides their songs, birds may have a repertoire of five to twenty calls--a collection of bird war whoops, alarm calls, love songs, and lullabies to their offspring. Even weirder is that birds are fully capable of singing duets with themselves. Their vocal organ, the syrinx, has two vibrating membranes that can somehow produce and modulate two independent tones at the same time.
Such extraordinary abilities can teach humans about their own speech abilities, says Nowicki, which is one reason his studies are funded by the National Institutes of Health. What's more, he says, bird song can yield powerful insight into the intricate mysteries of animal communication--the process by which an animal encodes and transmits a "thought" across space to another animal, where it is decoded and transformed back into thought.
If bird song seems remarkable on its surface, Nowicki's deeper studies of this tweet mystery of life have revealed it to be even more amazing. Until Nowicki's work, most scientists believed that the syrinx, located just beneath the bird's breastbone, was the only part of the bird's "instrument" that figured in its song-production. To Nowicki, such a theory was like arguing that the only important part of a clarinet was the reed. He believed that birds use their beak and throat to change the resonances of their song, much like humans singers use their mouths and throats to control their song harmonics. To test his theory, Nowicki had his captive songbirds sing solos in a harmless helium atmosphere. If only the syrinx mattered, helium wouldn't affect how it vibrated because the birds' song would remain unchanged. But if the windpipe were important, the song would resonate differently, just humans chattering away with a lungful of helium sound like Alvin the Chipmunk.
In fact, the helium did change the birds' song, causing new harmonic overtones to appear. The discover marked a critical new understanding that a bird's instrument consists of practically its whole breathing apparatus. "That discovery had very important implications for how a bird's songis wired up neurobiologically, how the bird learns his song," says Nowicki. "It adds a level of complexity to the problem of motor control that was previously unexplored." But like all good science, the experiment raised even more questions. "Having, I think, demonstrated that something is going on, we are now left in the position of trying to figure out what exactly it is," he says.
So, Nowicki and his students began high-speed videotaping of their birds singing, attempting to understand how a bird alters its head, throat, and beak to create its song. They also began experiments in which they fit the birds with "braces" to understand the beak's role. For brief periods at a time, they insert a small lollipop stick in a bird's bill and hold it in place with a rubber band. Their object is to fix the bird's beak open at a certain angle. "We think the bird does use its beak to change the effective length of its vocal tract, thereby changing its natural frequencies," says Nowicki. The birds remain unflappable during the procedure, quickly commencing to sing with their braces on.
So far, the researchers have found that the beak is, indeed, a critical part of the bird's song. The next step will be the daunting technical challenge of understanding specifically how the bird changes the shape of its vocal tract to control its sound.
While such studies probe the nature of the avian instrument, Nowicki is also trying to understand the new meaning behind bird song. "Ever since Darwin, bird song has been cited as an example of an exaggerated male trait, like the peacock's tail," he says. "Darwin proposed a distinct form of selection to account for these exaggerated traits--sexual selection as opposed to a natural selection." This sexual selection has to do with traits that evolve either to better attract mates or otherwise increase reproductive success. Whether such traits be peacock tails or ram horns, bigger is better up to a point, says Nowicki. "Now, the interesting question is what is 'bigger' in bird song?" he asks. Perhaps, scientists believe, the most successful male birds sweep females off their skinny bird feet by singing more kinds of songs--a correlation that they have found in some bird species but not others.
But the most fascinating discovery by Nowicki and graduate student Jeffrey Podos is that songbirds basically "wing it" when they sing: "If you listen to a song sparrow, you'll hear an individual sing about eight or twelve basic song types. But if you record those song types and analyze them closely, you'll realize that almost every time a bird sings a song, even of the same type, it does something slightly different." Thus, like human cabaret singers belting out old hits, birds add a little variety each time they sing even a standard song. This variety is very likely an important spice of bird life, says Nowicki. He and his students have discovered that their isolated baby birds learn to sing a multitude of song variations, to see how accomplished the birds can become.
But to really explore the meaning of this song complexity, Nowicki decided to "ask" the birds in the field what they hear when songs vary. "You know, we can measure in the lab until the cowbirds come home, but ultimately the question has to be validated perceptually by the birds," says Nowicki. So this summer, Nowicki and University of Miami zoologist William Searcy mounted an expedition to Pennsylvania whose aim was, basically, to mildly annoy male birds. His equipment: wooden marking poles, binoculars, a tape recorder, a loudspeaker, and an infinite amount of patience. His technique involves first pinpointing a male songbird's territory and installing a loudspeaker at its center. Then the scientists play a variety of songs, measuring how the male bird reacts.
At the first chirp of a recorded song, the bird will aggressively fly close to the loudspeaker. But it soon grows used to the song, wandering away until it renews its threat when the zoologists begin to play a subtly different version. By measuring the bird's response to different variations, the scientists can begin to understand what the bird perceives.
Nowicki and Searcy's summer expedition was also meant to make bird love, not just war. The scientists planned to capture adult female birds, give them a bit of hormone to put them in a loving mood, and play them recorded songs of avian amour. If the male songs are alluring, the female will fluff her feathers and adopt a ready-for-love "precopulatory" posture. "So, with these experiments, we'll get to see if this kind of variation has functional significance for courtship," says Nowicki.
Nowicki brings the same studious observation to his teaching as he does to his field work. He does something slightly embarrassing but highly useful the first week of his popular undergraduate course in neuroscience: Before class begins, he stands in front of the room with pictures of all 100 or so students spread in front of him. As student enter, he points to them, reciting their names, with the goal of eventually learning all of them. "I do it partly because it just gives me an excuse to stare at the students a bit," says Nowicki. "I try to get a sense pretty quickly who are the students who have a lot of scientific background, who are the students who are just intrinsically sharp, who are the students who are going to need more help. I also try to look into their eyes when I lecture."
As his class has grown from forty to seventy to 115, such individual attention has become more difficult, but he has persevered. His course's popularity has skyrocketed not because of its easiness. It's a rigorous semester that spans the breadth of neurobiology from molecule to neuron, and from jellyfish to the human consciousness.
Says zoology major David Finley, "It's an extremely interesting class, but it's not easy at all. The information he gives you, he doesn't expect you to just spit back on an exam. For example, on tests, he invents new organisms and asks you to use what you know to tell him what experiments you could do to answer questions about them." But it's not just the message that attracts students, says Finley, it's the medium of Nowicki's delivery. "He created a class dynamic that was just amazing," says Finley. "He commanded your complete attention; nobody flipped through the Chronicle or did crossword puzzles. It was how vivacious and dynamic he was in the classroom that caused you to be interested in what he said."
Nowicki's ambitious aim in the course is to help students begin to grasp the vast intellectual realm that attempts to explain how mind arises from body. "We want to do no less than help students try to understand the fullness of our own mental experiences, from their feelings about their mother to their feelings about Beethoven's Fifth Symphony. Of course, science is nowhere near that goal, but we don't want to forget it's the ultimate goal of neuroscience."
Nowicki's course also emphasizes how mind evolved from primitive forms. "We take the perspective that the higher-level processes that we're doing in class--teaching and learning--are in some way related to a toad catching a worm."
Certainly, Nowicki's street theater assures that nobody falls asleep. "I want to break down their intellectual complacency," he says of his thespian efforts. "I want them on the edge of their chairs, wonder what will happen next." Along with that suspense comes learning, he says. "The point isn't just to have a good time; the point is to also try to hammer home the complexity of some of these issues." Only once, he says, has a bit of theater fallen really flat. For a discussion of body structure, he had a fake Hollywood "wound" installed on his arm under fake skin, so he could dramatically rip off the skin to reveal the tendons and arteries beneath. The moment came, he ripped off the skin, the sight was satisfyingly yucky, and--laughter. "I was expecting fainting or worse," says a disappointed Nowicki. Perhaps, he theorizes, the MTV generation is too used to gore in its entertainment.
Besides his large neurobiology class, Nowicki teaches a smaller seminar on animal communication. He and University of North Carolina zoologist Haven Wiley have informally joined their Duke and UNC classes to explore scientific literature on animal communication. The students read assigned scientific papers and then report on and discuss them in class. A typical class might range through reports on communications in frogs, toads, damsel flies, warblers, and crickets.
Nowicki finds that the smaller class leads to more personal involvement by the student. "When this class works at its best, the students get very motivated and they get very excited. In fact, the students are just hard to shut up," he says with obvious pleasure. "The students get very passionate about whether animals can think or not. And they get passionate about our concern for animals and animal rights; because if animals can think, then what are our responsibilities and obligations to them? And they get passionate about human evolution and the origin of our own sense of self-awareness."
Nowicki sees the zoology department as a gateway for students into science, as well as into intellectual inquiry in general. "I really think that this department is an asset to Duke, partly because we happen to have many faculty who think broadly, even artistically, in a way that could become very highly integrated into a liberal education." Department chair Fred Nijhout, who studies butterfly wing patterns, is also an artist; professor Stephen Wainwright, who studies animal structure, is also a sculptor.
"This department also really values teaching, which makes me feel good about working on trying to be a good teacher," says Nowicki. "And what also excites me is just the wonder of knowing about things, and ours is a department where that kind of wonder is valued."
发布于12月9日 23:46 | 评论数(0) 阅读数(1150) | 我的文章
针灸改善大脑血液循环有科学依据
据《日本经济新闻》近日报道，日本东京老人综合研究所专家最近通过动物实验发现针灸治疗改善脑血流的机理。http://www.bioon.com/biology/advance/neuroscience/200510/164668.html
报道说，针灸给人带来疼痛和热的刺激，经神经传递，可改善大脑的血液流动。
人上了年纪，大脑血液流动衰减，记忆力下降，接受针灸治疗后脑
血流有所改善的病例很多，但此前人们并没有找到相关科学依据。东京老人综合研究所专家通过对实验鼠实施针灸，研究针灸如何促进脑血流的变化。
研究人员用针刺实验鼠面部，发现实验鼠大脑血流增加１０％至２０％。这种血流改善的状况可持续１分钟。用针刺入实验鼠腿部，效果基本相同。而切断与脊髓相连的神经，再进行相同的实验，实验鼠大脑血流没有改善。
研究人员在研究脑血流增加的实验鼠时发现，大脑皮质分泌的作为神经传递质的乙酰胆碱约增加两倍。研究人员认为，这是因为腿部和面部神经受到刺激产生的兴奋传递给脑神经，促使乙酰胆碱分泌。
发布于12月8日 16:04 | 评论数(0) 阅读数(939) | 我的文章
几个脑研究的报道
加拿大研究发现：男女大脑的确不同http://www.bioon.com/biology/advance/neuroscience/200512/167366.html
加拿大艾伯塔大学的一个研究小组日前宣布，科学证明，男人的大脑和女人的大脑的确不同。
据美国《华盛顿时报》２日报道，艾伯塔大学的研究小组给２３名男子和１０名女子做了核磁共振成像，结果发现，即使面对完全一样的任务，男人和女人的大脑也可能使用不同的区域。
精神病医师、研究报告的作者彼得·西尔弗斯通博士说：“这项研究表明，我们也许越来越会发现，男性和女性的大脑‘硬件’确有不同。”
西尔弗斯通博士希望上述发现能带来治疗抑郁和其他精神疾病的新方法；但有朝一日，上述研究成果或许也能为某些长久存在的行为模式提供解释。比如，同样是开车旅行，为什么男人拒绝问路而女人忙着看地图和路标？再比如，同是一部电影，为什么女人泪流满面而男人呼呼大睡？会不会是“硬件”不同？
研究小组成员埃米莉·贝尔说：“研究结果令我们非常意外。有时，女性和男性在做同一件事时却表现出不同的大脑活动。有时他们做不同的事却表现出同样的大脑活动。”
美国斯坦福大学医学院的精神病学家曾于１１月７日宣布，男人和女人的幽默感也不一样。他们在１０名男子和１０名女子看报纸漫画时利用核磁共振成像技术监测他们的大脑活动，结果发现，大脑对幽默的反应也存在性别差异。比如，男性期待着画龙点睛的那一句。女性则对语言有更好的欣赏力，期待也较少。但是，如果笑话中的确出现非常精彩的一句，女性会获得更大的满足感。
澳大利亚科学家找到紧张致病根源
http://www.bioon.com/biology/advance/neuroscience/200512/167674.html
澳大利亚悉尼加文医学研究所的研究人员宣布，他们已从科学上证实了情绪紧张与多种疾病之间存在联系。研究人员发现，人在紧张时释放的神经肽Ｙ（ＮＰＹ）会削弱肌体的免疫系统，使人患病。
５日出版的《实验医学杂志》月刊发表了这项研究成果。研究所的赫伯特·赫佐格说，ＮＰＹ对血压和心率的影响已为医学界所知，但发现它对免疫系统的影响为治疗某些疾病打开了新的大门。
赫佐格在澳大利亚广播公司说：“当你患有某种疾病时，如伤风感冒，情绪紧张会使你免疫力减弱，在更严重的情况下——如患有癌症，这会使你的病情加重。”研究人员认为，与情绪紧张相关的疾病还包括风湿性关节炎、多种硬化症、狼疮、Ⅰ型糖尿病等。
该研究所的另一位研究人员法比耶娜·麦凯强调说，研究针对ＮＰＹ的药物可能需要若干年，短期内最好的办法是患者自已缓解紧张。她说：“最好的办法是改变我们的生活方式，练瑜珈、最大限度地放松，消除生活中的压力。”
默契的脑神经机制（稽古轩主按：原文题目与内容不符合，据内容草拟）
http://www.bioon.com/biology/advance/neuroscience/200512/167725.html
明年，关于镜像神经元的研究有望取得重大突破。
镜像神经元是一种特别的神经细胞，通过研究这类细胞，科学家可能会发现，大脑如何让我们领会他人的想法。我们会弄明白，为什么在足球场上和舞场上，搭档们彼此能够心领神会。更重要的是，通过研究这些镜像神经元，科学家可以确定，人类的语言并非从说话开始，而是起源于姿势和模仿。
镜像神经元的故事始于１９９５年。当时，意大利帕尔马大学的奥亚科莫·里佐拉蒂实验室正在测算短尾猿大脑运动前区皮质脑细胞的电活性。研究人员发现，当短尾猿捡花生时，一些特别的神经细胞变得活跃起来。当短尾猿注意到一名研究人员伸手捡花生时，它们的这些细胞再度活跃起来。
这一研究显示，当我们看到某人在做某件我们要做或做过的事时，我们大脑中的同一区域也被激活，就像我们自己正在做这件事一样，这就是关键所在：我们其实不需要思索和分析，只需要激活我们大脑中的同一区域，就可以实时领会他人的思想。
直接领会同伴思想的这种能力，将灵长类同其他动物区分开来。当然，与人类相比，猿类和猴类的这种能力，最多只能算是“入门”水平。通过研究镜像神经元，科学家可能会揭开人类撒谎、欺骗和模仿他人等行为的秘密。
镜像神经元在心理学上的意义，就像ＤＮＡ在生物学上的意义一样重大，它将帮助科学家解开一些谜团。这些研究有望在２００６年取得重大突破。
产生意识觉察的时机
http://www.bioon.com/biology/advance/neuroscience/200509/160817.html 当受试者要求分别在半秒钟的时间内观看两张不同的图像或单词时，他们很难用意识觉察到第二张图像或第二个单词。一项发表在10月份出版的《自然—神经科学》上的新研究指出：受试者越早对第一个图像产生意识觉察，他就越有可能用意识感觉到第二个图像。
当受试者在确定自己是否看见了快速呈现的两个目标中的第二个时，Claire Sergent和同事记录下了他们的脑电活动。他们发现一个在每个目标出现后的早期活动波，无论受试者是否报告说看见过这个目标。相反地，在每个目标出现后产生了一个持续约300毫秒的晚期活动波，但这活动波只有在目标被意识觉察到后才会出现，这个波扩散到包括意识区域在内的许多大脑领域。与第一个目标相关的晚期活动波出现得越快，第二个目标被意识觉察到的可能性越大。新发现为区分意识觉察和无意识觉察的大脑过程的序列和时间提供了进一步的信息。
www.nature.com/neuro/journal/v8/n10/full/nn1549.html -
熟睡时失去意识的原因 大脑会“短路”
http://www.bioon.com/biology/advance/neuroscience/200510/162100.html 据美国媒体10月3日报道，我们在熟睡时常常对周边发生的事没有半点感觉或者是感觉模糊，这是为什么呢？美国科学家日前研究发现，原来人类大脑中的细胞是通过经常交换电子信号来进行信息交流，当人们熟睡时大脑中的一些区域会出现电子信号交换中断，这导致人们在睡醒时发生意识减退的现象。
此前，为什么意识会在熟睡中减退一直困扰着科学家，因为科学家很长时间之前就知道人们睡觉时大脑仍然是活跃的。美国威斯康星大学的精神病学家朱利奥-托诺尼带领一个研究小组对此展开了研究。他们使用了一种新技术——穿颅磁刺激，它可以精确并且无伤害地激活大脑中的小块区域。另外，他们让受试验者戴上一顶电极帽，它可以监测受试验者大脑中的电子活动情况。
在大脑中，信息都是通过神经元细胞传递的。托诺尼和他的小组发现，当受实验者清醒时，信息会沿着神经元网络传递到大脑的不同终端，然而他们熟睡时，这种信息传递突然中断，细胞不再进行电子交换。研究小组在一份报告中称，这显示在没有进入梦乡的睡眠中出现的意识消退可能是因为大脑皮层的不同区域发生功能性信号传递中断。大脑皮层主要负责协调感知、思考和行为。
研究小组还发现，受实验者在前半夜醒来时都会发现只有很少的意识或者是完全没有，而在后半夜，特别是在清晨，实验人员可以生动地描述他们做的梦，这意味着睡眠的最后一个阶段是有意识的。
托诺尼称，睡眠是最常见的意识状态变化，每个人每天都会睡觉，当熟睡时他们的意识通常就会消退，而当意识消退时，大脑就像是变成了无数的小岛，它们之间不会再发生交谈。
这一新的发现非常重要，因为它提供了大脑在睡眠中如何改变意识状态的第一手线索。意识目前在科学界仍是一个“黑暗”领域，因为科学家很少研究大脑如何维持并改变精神的各种状态。托诺尼是几名探索意识前沿领域的科学家之一，他已经建立了理论化的思想，他认为意识取决于大脑整合信息的能力，也就是大脑各区域进行信息沟通的能力。
发布于12月8日 15:49 | 评论数(0) 阅读数(1029) | 我的文章
关于脑机接口的最新研究进展和有关实验室
神奇帽子“人脑—电脑界面”用意念控制行动http://www.bioon.com/biology/advance/neuroscience/200509/160862.html
奥地利科技大学的计算机专家在上周开幕的2005伦敦科技展上，向大家展示了一种具有特殊功能的帽子，戴上它，计算机就能读懂你的思维。
这种神奇的帽子被称为“人脑———电脑界面”，它能探测到人脑中特定的运动区域的神经活动，然后在计算机的虚拟世界中用电子信号来模拟演示那个运动。神奇帽子的发明者Gert Pfurtscheller教授表示，这项技术有朝一日能帮助瘫痪的病人移动机器人手臂，或是帮助他们在虚拟的键盘上打字。
这项技术并非是个全新的概念，它是通过电极捕捉神经细胞的活动，然后通过电子信号传递给计算机。但是以往的技术需要通过手术将电极植入病人的脑中，而Pfurtscheller教授的帽子使用起来非常简单，可以免除手术之苦。
在展示会上，测试者戴着这顶帽子和一副三维视镜，根据计算机的指令行事。如果他按照计算机的要求想象着正在走路，大脑发出的运动信号若是被帽子成功解读的话，虚拟的人物就会按照他的指令开始前行，若是失败的话，虚拟人物则是站着不动。但是，不是所有的人随随便便就能完成测试的，要想成功地用意念来控制计算机，至少需要5个小时的专门训练。
稽古轩主按：
Gert Pfurtscheller教授的研究机构：http://www.dpmi.tu-graz.ac.at/
发布于12月8日 15:36 | 评论数(0) 阅读数(1203) | 我的文章
Methodological Issues in Event-Related Brain Potential and Magnetic Field Studies
Methodological Issues in Event-Related Brain Potential and Magnetic Field Studies
Walton T. Roth, Judith M. Ford, Adolf Pfefferbaum, and Thomas R. Elbert
Psychiatry in its search for the roots of abnormal thoughts, feelings, and behavior has again turned its attention to the human brain and is trying to apply the methods of the many scientific disciplines that have cast light on normal brain functioning-disciplines such as neuroanatomy and histology, biochemistry and molecular biology, and electrophysiology. This chapter concentrates on ways of maximizing what can be learned from noninvasive electrophysiology, a technique that is singular in its ability to record millisecond-by-millisecond changes in the brain following repeated external or internal events. Although the triggering events are often simple sensory stimuli, the cognitive processes that follow them and leave their trace in fluctuating voltage or magnetic fields can be quite complex. In the last decade competing noninvasive techniques such as positron emission tomography (PET) have challenged the preeminence of electrophysiology, particularly in spatial localization of brain processes. This challenge has stimulated a number of technological and methodological developments in acquiring, analyzing, and presenting brain electrical and magnetic data. But before we review these developments, we remind you of some basic principles and give examples of their relevance to psychiatry (see also A Critical Analysis of Neurochemical Methods for Monitoring Transmitter Dynamics in the Brain, Electrophysiology, and Pharmacology and Physiology of Central Noradrenergic Systems for related discussion).
Nerve cells generate extracellular current flow by fluctuations in the slower changing membrane potentials of dendrites and cell bodies. Postsynaptic potentials cause an outflow of negative (excitatory) or positive (inhibitory) ionic charges into extracellular fluid, which are then pumped back into the cell. This current flow, when summated, results in volume-conducted potentials recorded at the scalp as the electroencephalogram (EEG). Event-related potentials (ERPs) are EEG changes that are time-locked to sensory, motor, or cognitive events. They have provided a way to evaluate brain functioning in mental disorders and the effects of psychoactive drugs. Recent conceptual and technical developments have greatly expanded our capability to understand and document the mechanisms underlying surface recordings. Particular attention has been paid to identifying the location, orientation, and distribution of current dipoles (pairs of opposite charges) that may be the sources of scalp-recorded electrical activity.
Nerve cells also generate intracellular current flow from dendrites to cell body. This flow results in a magnetic field that can be detected at the scalp as a magnetoencephalogram (MEG), even though it is a billionfold less intense than the earth's magnetic field. Event-related magnetic fields (ERFs) can be elicited and time-locked to specific events and are analogous to ERPs. Magnetoencephalograms and ERFs convey different information than EEG and ERPs. This is because voltage fields on the surface of a sphere, which the skull enclosing the brain approximates, are produced equally well by dipoles oriented radially and tangentially with respect to a radius of the sphere. In contrast, 90% of the magnetic field at the skull can be ascribed to tangential dipoles alone. This is a consequence of the geometrical orientation of masses of nerve cells and of magnetic sensors. Fig. 1 illustrates how dipole orientation can be either correlated or random for different gyri and sulci. Parallel dipoles lying tangentially on sulcal walls contribute much more to the MEG than random dipoles or dipoles lying radially along the crowns of gyri.
Why are the methodological issues that this chapter addresses relevant to psychiatrists and psychologists? First, ERPs and ERFs are theoretically relevant because they provide ways of testing theories of abnormal brain functioning that no other methods can offer. For example, unlike ordinary behavioral tests of cognitive processing, ERPs give an index of the processing of task-irrelevant events, distracting stimuli, or events subjects have been told to ignore. The topographic distribution of ERPs and ERFs gives clues as to what parts of the brain are active during a particular cognitive activity. Second, ERPs and to a less extent ERFs have been demonstrated empirically to be relevant. ERP abnormalities have been repeatedly observed in psychiatric disorders, notably in the P300 and P50 components. The P or N signifies positive or negative and the number is the mean peak latency in milliseconds. Thus, the P300 component is a positive potential that occurs approximately 300 msec after a stimulus that is infrequent and in some way relevant. The most venerable and consistent psychiatric ERP finding is that of reduced P300 amplitude in schizophrenics (60), although this is not specific to schizophrenia (see refs. 59 and 22 for reviews). For instance, a longitudinal study demonstrated that lower P300 amplitude at age 15 was predictive of poorer global personality functioning at age 25 (66). Latency at P300 is generally greater in patients with dementia than in normals or in patients with schizophrenia or depression (28, 54). Recently, psychiatric attention has been directed to P50, an ERP component to auditory stimuli whose amplitude is suppressed if the eliciting stimulus is paired with another that precedes it by one-half second. Schizophrenics show less P50 suppression than controls (25) as indicated by smaller amplitude ratios (P50 to the second stimulus of a pair divided by P50 to the first), although again this finding is not limited to schizophrenia (4).
Abnormalities of ERPs in psychiatric patients can be interpreted in light of a considerable amount of knowledge that has accumulated about the significance of certain ERP components in normal human information processing. For example, P300 is known to reflect the categorization of events, depending jointly on stimulus probability, stimulus significance, and the information value of the event (36). Probably, P300 has multiple, partially asynchronous generators (58). Components occurring 60 to 100 msec after onset of auditory stimuli, including N100, have been shown to reflect selective attention to auditory stimulus channels (42). In contrast, auditory ERPs with latencies less than 10 msec are insensitive to attention effects but give a unique assessment of the intactness of brainstem circuitry (32).
The literature on ERFs in normal subjects is quite extensive although magnetic recording techniques have been available only a relatively short time. Much of that literature has documented the existence of ERF components that parallel those established by invasive and noninvasive ERP recording. However, to date, most clinical MEG studies have been done in neurological rather than psychiatric patients, although that is likely to change in the near future. Reite et al. (57) recorded ERFs in six medicated, paranoid schizophrenic patients and six normal controls. The M100 component (analogous to the N100 of the ERP) showed less interhemispheric asymmetry in schizophrenics and had different source orientations in the left hemisphere. Tiihonen et al. (68) compared the M100 component in two schizophrenic patients when they were experiencing auditory hallucinations and when they were not. During hallucinations, M100 peaked approximately 20 msec later, an effect similar to that of external masking noise in normals.
We now turn to methodological trends that are transforming ERP and ERF research. Specific topics include data acquisition, signal averaging, ocular artifact, choice of reference electrodes, digital filtering, measuring components including dipole modeling, and statistical and diagnostic considerations.
Electroencephalogram Systems
Older electroencephalographic tube-based amplifiers have been completely replaced with high impedance solid-state amplifiers with electronically controlled amplification and filter settings. In many laboratories, pen-chart recorders have been replaced with electronic data storage and display systems, but paper records are still widely used for visual analysis of diagnostic EEGs and sleep. Laboratory computers are constantly evolving toward faster, cheaper, and more powerful models. New storage media based on tape or magnetic or optical disks permit archiving of data from many subjects in an easily retrievable form. As welcome as these advances have been, they have generated difficult new choices for researchers. Should they buy commercial EEG and ERP hardware and software systems or develop their own? Which commercial systems or routes to laboratory-program development are satisfactory? Commercial systems tend to be limited in flexibility, details of data analysis may be a trade secret (which is unacceptable scientifically), and access to raw data for special analyses may be difficult. Laboratory-developed systems require deciding among manifold hardware and software possibilities, and then allocating many hours to programming. As will be learned from this chapter, methodologically up-to-date ERP analysis requires much more than eye-movement artifact rejection and signal averaging.
Whereas the conventional 10–20 system of Jasper (35) used 19 electrodes with a typical distance of 6 cm between them, some investigators have greatly expanded the electrode arrays in order to record more of the spatial detail present in the EEG. Thus arrays of 124, or even 256 electrodes, which yield interelectrode distances of 2.25 and 1.6 cm, are now being advocated (27) and have been shown to enhance localization. The application of multiple electrodes is a lengthy, labor-intensive process, which requires care in scalp preparation and accuracy in electrode placement. For localization studies relating EEG or MEG data to brain structures visualized by magnetic resonance imaging (MRI), it is important that electrodes be aligned correctly according to skull landmarks, and fiducial markers visible in MRI scans are used. (Vitamin E capsules are easily available and the right size.)
Electrode application entails a potential health risk to both subject and technician if the intactness of the scalp is compromised by procedures to reduce electrical resistance between electrode and scalp or by skin lesions. Acquired immunodeficiency syndrome and hepatitis B can both be transmitted by this route, so it is absolutely essential that proper precautions be taken. Putnam et al. (56) give recommendations for disinfecting reusable electrodes and for protecting the technician.
Magnetoencephalogram Systems
The recording of the MEG has been made practical by the development of superconducting quantum interference devices (SQUIDs) that are sensitive to minute magnetic fields. The MEG technology is much more expensive than the EEG technology. Not only are the SQUIDs themselves expensive, but they require provision for liquid helium at 4.2°K to cool them, and a recording room shielded with a high-permeability material against magnetic fields and with aluminum against eddy currents. The liquid helium is kept in a vacuum-insulated container called a dewar. Locating magnetic sources requires recording from multiple sites, preferably simultaneously. Otherwise, separate stimulation runs must be made, moving sensors from one location to another between runs. More runs take more recording time and increase the likelihood that the subject's mental state will change, altering the sources. A MEG system with over 30 channels costs approximately $3,000,000, 100 times more than the same number of EEG channels. Because MEG prices reflect the cost of research and development more than construction of the apparatus, the price per unit would drop if more units were sold. In one system, 37 sensors are placed 2.2 cm apart to cover a single hemisphere (12).
An advantage of MEG sensors is that they do not touch the head, so transmission of infectious agents is of less concern. Fixation of head position is critical so that sensors can be aligned according to skull landmarks. Modern SQUID technology allows recording of signals that vary slowly over a minute, undisturbed by electrode drift. A new method for recording even slower or static magnetic fields converts such fields to more rapidly changing fields by having the subject lie on a mechanically driven platform that executes a circular movement of a few centimeters at 0.2 Hz (26). Auditory and visual stimulation cannot be given by conventional earphones or CRT displays because of their magnetic properties. Instead, sounds have to be delivered from outside the testing chamber through hollow tubes and visual stimuli projected through a window in the magnetic shield or delivered fiber optically.
Both ERPs and ERFs benefit greatly from signal averaging to enhance their signal-to-noise ratio (SNR). Data are generally digitized at a fixed rate to fill a data array, and a stimulus or other synchronizing event defines the time epoch of interest within this array. The event is repeated (each repetition is called a trial), and a time-locked signal (ensemble) average is calculated across trials epochs for each time point of the epoch. If Xj(t) is the electrical potential (voltage) or magnetic field strength at some electrode or sensor location at time t and trial j, the signal average is defined as
If Xjt is considered the sum of true signal mt and random noise Njt (background EEG and measurement error), signal averaging improves the SNR. Unbiased estimates of signal power 
, noise power 
, and SNR can be calculated as follows (71).
One of the assumptions of signal averaging is that the signal is invariant across trials. This assumption is violated when the amplitude of the ERP component of interest habituates or when its latency varies from trial to trial, as is clearly the case for components related to certain cognitive processes, such as the P300. One way of dealing with component latency variability is to locate the signal on each trial and align the trials on these signals rather than on the eliciting stimulus. Woody (75) proposed an iterative procedure (an adaptive filter) that located the signal on each single trial by moving a template (initially the signal average) by time increments along the trial to find the latency of maximum correlation. A new average was then formed by aligning trials on the identified signal latencies, and the new average was used as a new template. If the SNR is too low, this procedure produces results that simply reflect random noise. Gratton et al. (31) tested the procedure with simulated signals and background EEG noise and demonstrated that iterations (up to three) were important only when the original template had a wavelength on the order of two times longer than the signal.
Roth et al. (56) used this procedure to analyze ERPs elicited from schizophrenics and controls performing an auditory choice reaction time paradigm in order to test whether P300 amplitude reduction in schizophrenics could be attributed to latency variability. They found that individual trial P300 latency was indeed more variable in schizophrenics but that schizophrenic P300 amplitude was still smaller than control amplitude after latency adjustment. To reduce distortions due to noise, Pfefferbaum and Ford (53) modified the procedure by only including trials whose covariance is greater in the part of the epoch where signal is expected than in the part where noise is expected, and whose correlation with the template (initially a half-sine wave) exceeds a set threshold. Using this modified procedure, Ford et al. (23) replicated the Roth et al. (61) finding that schizophrenic P300 remained smaller. Furthermore, schizophrenics had more trials that did not pass the covariance–correlation screen than controls. Trials that did not qualify for latency adjustment had longer reaction times, showing that they were deviant behaviorally as well as electrophysiologically. In addition, Ford et al. calculated for each subject the covariance of P300 signal average across trials with that subject's EEG in single signal epochs and in single nonsignal epochs. The ratio of mean signal covariance to mean noise covariance was significantly smaller in the schizophrenics. Because trials were filtered with a bandpass of 0.5 to 4.4 Hz, noise was EEG activity in the frequency range of P300 rather than higher frequency like a, b, or muscle activity.
Another assumption of signal averaging is that background EEG noise is random noise. This is only an approximation to the truth, as a study of event-related spectral perturbation indicates (41). In normal subjects, auditory tone pips reliably produced momentary increases in spectral power in the 2- to 8-Hz and 10- to 40-Hz bands.
Eye movement and blinks produce electrical potentials and magnetic fields that are often much larger than those deriving from brain sources. The magnetic fields are more restricted to the vicinity of the eye than are the electrical fields and for this reason are less troublesome if unsynchronized with events of experimental interest. Synchronized eye artifact can cause major errors in peak measurement or source localization. Attempts to control this artifact by instructing subjects to fixate their gaze on a point or not to blink are often ineffective, particularly if the subject is psychotic or cognitively impaired. Thus methods for removing eye artifact from the ERP or ERF need to be applied. Many are based on determining the coefficients Ak in the equation
V(k,t) = Ak * EOG(t) + EEG(k,t)
where V(k,t) is the voltage observed in lead k at time t, and EOG(t) and EEG(k,t) are the true EOG and EEG voltage contributions at that time.
Spatial-temporal dipole models of eye movements and blinks make it clear that the same correction cannot be used for both (6). Thus eye-correction procedures should include at a minimum the following steps: (a) Separate blinks from movements on the basis of their temporal properties, (b) calculate separate linear regressions for the propagation of artifacts from each, and (c) correct EEG leads by the amount predicted by the regression coefficients. Gratton et al. (29), whose method has been used by a number of investigators, adds an additional step of subtracting signal averages from individual trials to avoid distortions resulting from ERP effects in both EEG and EOG records. A computerized implementation of this procedure that adjusts for both a vertical and a horizontal EOG channel, has been developed (43). Although certain technical issues in implementing EOG corrections remain unresolved—the proper number and position of EOG electrodes, the error attendant upon assuming a linear relationships between the EOG signal and EEG artifacts, the implications of the presence of EEG artifacts in EOG leads, how to deal with overlapping eye movement and blinks, and instability of individual propagation factors between sessions and even between tasks within a session (19)—the use of such off-line procedures have greatly increased the number of trials available for analysis in clinical studies.
Whereas MEG sensors detect the absolute magnetic field at a given location in space and need no reference in the body, the EEG must be measured as voltage differences between two points on or in the organism. Ideally one point should be close to the biological voltage source under investigation, and the other should be a reference point with constant voltage or at least a voltage not correlated with the source voltage. Traditional references for human ERP have been linked mastoids, linked ears, or the nose; unfortunately none of these is unaffected by brain sources. Special disadvantages of linked ear references include the possibility that shorting can reduce asymmetry if resistance is low, and the possibility that artifactual spatial asymmetry will result if resistances at the two ears are not equal (48). Shorting is not a serious consideration as long as skin-electrode resistance at each ear is greater than 5 kW, because in that case scalp path resistance is reduced less than 5% (44). Resistance at the two ears can be balanced with a potentiometer, or one ear (say A1) can be used as a reference and recorded as a separate channel. Then a linked ear reference for say Cz, a scalp electrode in the 10–20 system, can be created algebraically, (Cz - A1) - (A2 - A1)/2 = Cz - (A1 + A2)/2.
To avoid active reference electrodes on the head, some investigators have turned to noncephalic (e.g., sternovertebral) electrodes (67). Unfortunately these electrodes are liable to pick up heart activity even when adjusted to be at right angles to the main vector of voltage during the cardiac cycle, since cardiac depolarization and repolarization vectors do not maintain a perfectly constant direction over the cycle.
Another solution is to use an average reference. At each time point, an average reference defines zero over C electrodes in a data array A as
A limitation of the average reference is that when electrodes are not densely and equally spaced around the brain, for example, there are none at the bottom of the head (69), the sum in the formula above is generally different from true zero. For example, Desmedt et al. (16) have shown that P14 of the somatosensory evoked response, which is present with a linked ears reference, disappears when a zero reference based on 27 scalp electrodes is applied, becoming surrounded by "ghost" negativities. A linked-ear reference reflects more accurately the medial lemniscal volley that is the presumed basis of P14. In addition, local changes can be mistaken for global changes with a zero reference. These distortions are less likely to affect tangential than radial dipoles.
In conclusion, there is no perfect reference for all cases. As a general principle, a known local source should be referred to an electrode distant from it.
Before measurements are made on ERPs or ERFs, it is useful to apply SNR-enhancing filters that incorporate assumptions about frequency, timing, and spatial distribution of the component of interest. For example, the ERP P300 component may be expected from experiments in the literature to have a frequency lower than 2 Hz (30), to peak in a range of 280 to 400 msec (in a simple auditory choice reaction time task in young adults) and to be maximal at Pz, another electrode in the 10–20 system. Though signal averaging attenuates unsynchronized noise at every frequency as it improves SNR, frequency filters are commonly applied prior to component measurement. These filters are useful whenever the frequency of the noise is different from that of the signal.
Digital Filters
Digital frequency filters (11) have the advantage over analog filters of being able to operate without introducing distorting phase shifts into the signal. The most commonly used digital filter has been the moving average or boxcar filter, in which each point of the signal is replaced by an average of that point and a certain number of prior and subsequent points. This is only possible for stored data, because it makes use of future time points to calculate current output. Farwell et al. (20) have shown that a simple moving average filter does not prepare average and single-trial waveforms as well for P300 peak-picking as does a filter designed by an optimizing algorithm. Such an algorithm determines a set of weights that are able to reduce deviations (ripple or ringing) in the passband and stopband of the filter. Optimized filters have less tendency to reduce P300 amplitude or distort shape and, in the case of averages, gave more stable latency measurements. For P300, the authors recommend that the optimum filter have a passband cut-off frequency of 6 Hz, a stopband cut-off frequency of 8 or 8.5 Hz, and use 490/n points, where n is the sampling interval in milliseconds. It should be emphasized that analog filters still have a place in data acquisition prior to digital filtering—a low-pass analog filter with a half-power frequency below but close to half the sampling rate prevents aliasing, and, for P300 recording, a high-pass analog filter with a half-power frequency of less than 0.16 Hz minimizes irrelevant baseline shifts (20).
Spatial Filters
Current source density maps (also called surface Laplacian or radial current estimate maps) act as spatial filters emphasizing localized components with a high spatial frequency. For this to work well of course, electrodes must be placed with a high spatial frequency. Maps can be made of unaveraged activity such as epileptic spikes or of signal averages. Sensory ERP components show a more localized distribution using this approach than in voltage maps. For example, Nagamine et al. (46) compared voltage and current source density maps on the scalp ERPs obtained by tibial nerve stimulation. The results for a single subject presented in Fig. 2 demonstrate better localization for P40, N50, and P60 for the current source density map. The equation for calculating current source density is I = r(d2V/dx2 + d2V/dy2), where V is the voltage, x and y the surface location on the x–y plane, and r the charge density. In addition, r = k * d2, where d is the distance between electrodes and k is a constant for all electrodes within a subject. The Laplacian operator can give limits for finding equivalent dipoles. It has a physical interpretation—local radial current flow from the brain into the scalp and vice versa—but it is different from dipole modeling (described below) and is free of dipole modeling's ambiguities.
In the Laplacian calculation, surface contours can be generated by a method called spherical spline interpolation, which is based on physical principles for minimizing the deformation energy of a thin sphere constrained to pass through known points (51). This produces a smooth surface running through the data values and filling in between them, even when electrodes are irregularly placed on the scalp. Spherical splines have advantages over plate splines, which are based on deformation of an infinite thin plate. As might be expected from the fact that interpolated values at any point are derived from data from other locations, coherence (a measure of covariation) is inflated by interpolation. Nearest-neighbor interpolations are less smooth and inferior for locating extrema (peaks and troughs must lie on an electrode site) but do not inflate coherence.
Gevins et al. (27) have demonstrated a method of current source density mapping they call finite element model deblurring that they believe is superior to the Laplacian method. Mathematically, it is a less computationally demanding version of dipole modeling known as spatial deconvolution, which assumes that all dipoles are located on a cortical surface. Gevins et al. use the subject's head MRI to provide information about conducting volumes between scalp and cortical surfaces.
A simpler spatial filter, the vector filter (30), has been used for component measurement. Its output is the weighted sum of data points at different electrodes. Conceptually, measuring a component at one lead is the same as applying a vector filter with weight 1 assigned to values at that lead and weight 0 to values at all other leads. Vector filtering assumes that the distribution of the component to be measured is constant despite changes in amplitude or latency. The crux of the procedure is how to specify the weights: using three 10–20 system scalp electrodes, Fz, Cz, and Pz, weights of 0.15 for Fz, -0.53 for Cz, and 0.83 for Pz were found to produce optimal discrimination in an oddball paradigm between rare trials, which contain substantial P300s, and frequent trials, which do not (30). Thus, optimum weights do not necessarily correspond to component distribution, because P300 is larger at Cz than at Fz. Dipole modeling, which is described below, can act as both a spatial and temporal filter.
Measurement Methods
A component can be defined as electrical or magnetic activity associated with a specific neurological or psychological process, for example, a motor act such as moving one's finger, a sensory process such as the reaction to a light flash, or a cognitive process such as categorizing a stimulus as target or nontarget. In a statistical sense a component explains experimental variance. The details of the experimental method are part of the operational definition of a component. As more experiments are done, theoretical expectations about components develop into generalizations. For example, many experiments in which subjects performed a fixed foreperiod reaction time task have resulted in a parietal–central negative shift prior to the button press. A natural generalization is that the parietal–central shift represents preparation for a motor act. Furthermore, because the source of the recorded data is a physical location within the brain, the ultimate description of a component must include reference to the specific brain structures activated. Some leads or sensors will pick up activity from those structures better than others, particularly when sources are multiple with overlapping influences. In the case of ERPs, the choice of voltage reference influences how electrical activity from a source appears in the EEG recording.
Measurement procedures include peak picking, area measurement, waveform subtraction, principal components analysis, template correlation, and dipole modeling. Peak picking means finding maxima or minima in specified latency ranges and determining peak latency and amplitude with respect to a prestimulus baseline. This is the simplest method of component evaluation, but can be biased when latency ranges are selected after an inspection of the data, and is perhaps unduly restricted in that it considers only peaks among other waveform features. In addition, it is often based on only one point, which may be influenced by noise or overlapping components. With multiple leads, another limitation of peak picking becomes obvious: what appears by shape to be a single component has maxima at different time points in different leads, and it is not clear how best to resolve the discrepancies. Furthermore, the choice of reference electrodes can determine when peaks and troughs appear.
Area measurement is sometimes used when the component is believed to be more rectangular than peaked. Area is measured in a specified latency range, and is thus based on multiple points, but area measurement, like peak picking, can be biased and is influenced by overlapping components.
Waveform subtraction can be used before peak picking or area measurement to reduce the effects of component overlap. For example, consider a paradigm where tones of two pitches are given in an unpredictable sequence and one occurs less frequently and is designated as the target of some task. The ERP to the rare tone can be considered a combination of the sensory effects of the tone and the cognitive effects of the tone being a rare target. By subtracting the ERP to the frequent tones from the ERP to the infrequent tones, the sensory effects are removed leaving behind the cognitive effects. This assumes that the sensory responses to the two tones are identical and that cognitive and sensory effects are additive, an assumption that is not always warranted. For example, frequency-specific temporal recovery of the auditory N100, a noncognitive effect, makes the response of N100 to frequents smaller than the response of N100 to rares.
Principal components analysis (PCA) is another approach to ERP component measurement, which uses the time points on waveforms from different subjects, different electrodes, and different experimental conditions to define components. In statistical terms PCA identifies orthogonal axes of maximal variance in a multidimensional space defined by the variables. Generally these axes are rotated according to the varimax procedure. Less arbitrary than peak picking, PCA makes no assumption about the latency range in which specific components will be found but only that they have a fixed latency across conditions and subjects. It has some ability to separate overlapping components. However, PCA is not completely free from arbitrariness. First, PCA solutions are not unique. Many rotations of the factors are possible. Second, results depend to a certain extent on what experimental conditions are chosen and how many leads are included. Variance from electrodes, subjects, conditions, and correlated noise are all treated the same. Furthermore, each experiment gives slightly different factor structures, and there is no established criterion for deciding whether these differences are significant or not. Thus, it is uncertain how many statistical components to interpret, and how to identify these components with ones previously described.
Template correlation assesses the similarity of a template of the component to the waveform to be evaluated. The template may be based on prior knowledge of the component shape or on signal averages (see the iterative Woody filter procedure described above). The template is usually compared to waveforms at specified intervals over a designated latency range to identify the latency of maximum correlation (or in one variation, maximum covariance). This time point is defined as the peak. The sum of cross products at this time point or the difference between amplitude at this point and a baseline can define amplitude.
Interpreting latency data under different experimental conditions can be difficult when multiple leads are involved. Latency may vary at different leads and topography may vary under different conditions, implying different components whose latency cannot be compared. To solve these problems, Brandeis et al. (8) spatially generalized the Woody filter procedure using an average reference map, and applying a measure they call global field power (GFP) defined by the following formula for an array A consisting of data from C electrodes:
Further, global dissimilarity (GD) is defined as the root mean square (rms) power of the difference maps calculated by subtracting two normalized GFP maps. The procedure is as follows: (a) Grand averages are used to form template GFP maps, from which component model maps at single latencies near 100, 200, and 400 msec are derived, corresponding to P1, N1, and P3 (see ref. 8 for details). (b) Component model maps are moved in specified latency ranges around the latency of each model's component. The minimum of GD multiplied by sequential dissimilarity (GD between current and previous map: a stability constraint) is calculated, and the minimum of this function (best fit) is defined as the map latency for that component. (c) In an iteration, the average of all normalized maps at their latencies of best fit is used as a new model, and the search window is set around the new mean latency. The results show that components can be identified by topography alone, without respect to amplitude or time. However, this method does not take into account possible overlapping components and would fail if such components influenced topographies. Furthermore, average references for P300, which is widely distributed on the top of the head, may be inferior to a noncephalic reference.
Dipole modeling is a method for reducing data from multilead EEG or multisource MEG by deducing the dipole sources that may have produced them. Although the forward problem (calculating scalp distribution from known dipoles) has a unique solution whose accuracy is limited only by the approximations of skull geometry and conductivities, the inverse problem has multiple mathematically valid solutions as was pointed about by Helmholz more than a century ago (33). The reason is that a single scalp distribution can be produced by different numbers of dipoles in different combinations of locations and orientations. Thus, various constraints on the number of sources allowed and their approximate location must be applied to reach a solution. Sometimes these constraints are so severe as to specify that the source be a single dipole located somewhere in the brain.
At an abstract level, dipole modeling is like PCA in that an equation U = C * S must be solved where U is an array of k electrodes at t times that represents the linear superimposition of the array S of m sources at t times multiplied by C weighing coefficients at k electrodes for m sources (62). Whereas PCA determines C and S from mathematical constraints, dipole modeling assumes that C depends on volume conduction from j dipoles at certain locations, assuming Ckj = f(rj,oj,ek ), where f is a nonlinear function of the electrode location vector ek and of the geometry of the source and the head. The dipole has a location vector rj and the orientation vector oj. Equations defining a 3-shell sphere model of the head with differing conductivities for scalp, skull, and brain are found in the appendix to this chapter. Using these equations to model dipoles at various depths, Pfefferbaum (52) demonstrated how increasing the thickness of the superficial extrasulcal subarachnoid layer of cerebrospinal fluid (CSF) or skull thickness might affect scalp ERP amplitudes and topographic distributions.
One procedure for the dipole modeling of ERPs was developed by Scherg and Berg (64). Their software is available commercially as brain electrical source analysis (BESA, from Neuroscan, Inc.). It models a window of points, assuming a finite number of equivalent dipoles with fixed location and orientation. In its recent version, it does not assume a parametric dipole magnitude function (like the decaying sinusoid of ref. 70) but computes a varying magnitude function over the window of points for each dipole. The BESA model is applied iteratively, calculating at each step the residual variance (percentage of recorded data not explained by the model). The first step looks for the inverse solution by calculating parameters of a plausible dipole from an EEG or MEG data map. Then forward solutions calculate resultant EEG or MEG maps from those dipoles. Hundreds of iterations may take place, stopping when the change in residual variance is less than some criterion, such as 0.001%. When more than one dipole is modeled, some may be fixed in position (but not in amplitude) while a new dipole is optimized. The results of these procedures depend among other things on the starting location and other parameters of a dipole. An iterative procedure may find topographically local optima that would not be optima if all locations and orientations were tested. Scherg and Berg (64) explained that multiple-source solutions are less arbitrary if spatial and temporal constraints are added. For example, two sources may be required to have a symmetry between hemispheres, radial and tangential dipoles, or lie in the supratemporal plane. How this method works is illustrated in Fig. 3 and Fig. 4, Figure 3 shows ERPs to clicks and resultant dipoles that were inferred from these ERPs. Figure 4 shows how well four models account for the data. The model that explains the greatest amount of the variance (99.4%) and corresponds best to anatomic reality assumes six dipoles: one central, two bilaterally symmetrical pairs, and one unilateral, coming from the postauricular muscle. Of course, some of the 99.4% may be noise rather than signal.
Other procedures are possible. Turetsky et al. (70) developed a method called the dipole components model, which simultaneously fits multilead data from a time window in multiple averages, pooling noise estimates. It assumes that the component shape is a decaying sinusoid and that the skull is a sphere of homogeneous conductivity. Turetsky et al. (70) applied it to P300 elicited in an auditory oddball paradigm and found four dipoles in two dimensions, three of which varied with experimental conditions. Cardenas et al. (9) applied it to the P50 suppression paradigm in the reliability study described below.
A single dipole modeled at brief intervals can mathematically generate a moving trajectory of loci. The two main alternatives to single dipole modeling are multiple dipole models and fully distributed models (34). For the second, a probability density is generated for widely distributed current sources. In addition, cylinders rather than points may be modeled. A distinction between a point source and a region can only be made if the region is of a size comparable to the distance between sensors.
For a MEG, it is not necessary to employ a layer model because magnetic permeability is unaffected by variations in conductivity. In practice only the radial component of the field is measured because it is convenient to place pickup coils parallel to the scalp (reviewed in ref. 39). Although generally it is assumed that the source is composed of similarly oriented and concurrently active neurons, this simplification is clearly wrong in certain cases, such as the folds of the visual cortex, which are better modeled by a cross-shaped arrangement of dipoles. The strength of the resultant dipole detected at the scalp depends very much on the symmetry. Synchronization (as with the appearance of a waves) may actually be a periodic breaking of symmetry of activation of the component dipoles (39).
Often, ERF analyses use peak data to model the dipole, because the SNR is likely to be highest there. An example of a dipole analysis in which the results were coordinated with MRI scan data is the work of Pantev et al. (50). They analyzed ERFs elicited by auditory tones of varying pitches and based on at least 96 trials for each pitch from each of 60 measuring positions at the M100 peak (in this case at 88 msec) for a single current dipole source. Fig. 5 shows isofield contour plots of the ERF at 88 msec for a single subject and the positions of the dipoles associated with each pitch. Fig. 6 shows the coronal MRI section with the dipole locations for that subject. They lie just below the surface of the transverse temporal gyrus (Heschl), the assumed location of the primary auditory cortex, and are ordered in depth by pitch. Modeling before and after the peak may give somewhat different dipoles, but it is hard to exclude the possibility that they are spurious. To accentuate the onset of activation of weak secondary dipoles, Moran et al. (45) calculated dipoles associated with auditory ERFs on the basis of differences in magnetic fields in 4-msec intervals between 0 and 300 msec. This interval selects for components of a frequency high enough to change during it. Using this method, the authors found evidence for a source spatially separate from N1m but coactive with it. A distributed source in Heschl's gyrus and adjacent areas could also produce such a result.
The number of sensors (SQUIDs or electrodes) is important. For ERFs we need to know n * 5 parameters if n is the number of sources and, for ERPs, n *6 (65). For ERPs, this means a minimum of (n * 6) + 1 electrodes. Thus, the conventional 19 electrodes allow only 1 or 2 generators to be determined. The results of dipole modeling can be ambiguous in that substantially different models provide only trivially inferior fits. It is more important to analyze the number of sources and their gross location than their exact location. A good initial approximation escapes local minima in residual (unexplained) variance but begs the question. Noise, particularly if it is spatiotemporally organized, can distort solutions by creating local minima. Achim et al. (1) created simulations and used a variety of procedures to analyze them. By using several initial approximations (rather than simply reinitializing with a previous solution) and a multiplicity of optimizations, they managed largely to escape local minima. Precise localization is prevented by the presence of background EEG noise. Errors ranged from 2.5% to 13% of sphere radius. The authors developed a residual orthogonality test for testing the presence of signal in residues after modeling.
Sensory ERPs and ERFs are more likely to be amenable to dipole modeling than more complex cognitive ones. Witt et al. (74) applied dipole modeling to brainstem auditory evoked potentials (BAEPs). These authors recorded simultaneously from 12 electrodes constituting three three-channel bipolar montages. Data from all montages were transformed to fit the same central dipole. The authors concluded that a tetrahedral montage equivalent to Einthoven's Triangle for the EKG is adequate for clinical work, although it is slightly inaccurate because the dipole is known to move over time.
In an investigation of a nonsensory ERF, Elbert et al. (17) measured the magnetic field prior to button response in a go–no-go reaction time task. With this task, the EEG shows a negative shift prior to the button press called the contingent negative variation (CNV). The magnetic equivalent, which they called the contingent magnetic variation (CMV), was larger for go than no-go conditions, but a moving single dipole model accounted for less than 80% of the variance in four of eight subjects. The authors conclude that the later parts of the CMV are particularly dependent on distributed sources in motor, sensory, and association areas. Another component that is likely to have multiple sources is P300 (37). Turetsky et al. (70) applied their dipole model to model electrical P300 in 18 subjects using data from the oddball two-tone choice reaction time task. Using four dipoles in the midsaggital plane, they could explain approximately two-thirds of the total variance across subjects, conditions, and electrodes.
An unsolved problem with dipole estimation is how to decide if dipoles are equivalent. For example, experimenters may want to statistically compare dipoles modeled from individual subjects to draw general conclusions valid for a group, yet each dipole will vary somewhat in its location and orientation from every other. In the approach of Turetsky et al. (70), a single solution encompasses all subjects and conditions in an experiment, but it is still important to be able to compare dipoles between experiments. A related problem is how many of the multiple component dipoles generated in a given application of a model should be considered valid. This is analogous to the problem of how many PCA components to accept in a given analysis.
Measurement Reliabilities
The reliability and accuracy of certain computerized methods for measuring P300 has been assessed for averages and single trials. Reliability of automated measurement is a function of two factors that are often difficult to untangle: the stability of the underlying component being measured over time and the effects of electrical sources other than the component (background EEG and muscle and eye artifact). Whatever its cause, unreliability reduces a measure's usefulness.
Recent parametric studies have illuminated some of the variables underlying unreliability. Fabiani et al. (19) found that P300 latency estimates of averages had split-half reliabilities between 0.63 and 0.88, and in most paradigms was rather similar for peak picking and template correlation. Amplitude estimates of P300 were most reliable (between 0.90 and 0.96) when based on covariance with a full-cycle 2-Hz cosinusoidal wave. Making measurements at Pz alone was almost as good as using the output of a vector filter based on Fz, Cz, and Pz. Subtracting averages of frequent trials from infrequent trials led to more reliable measurement of the probability effect than when the two types of trials were measured separately. Test–retest reliabilities of both amplitude and latency were lower between than within sessions, probably because of changes in P300 over time. Gratton et al. (31) did a simulation study of P300 single-trial latency estimation, embedding known signals in noise from actual EEG records adjusted to give various SNRs. Peak picking and several methods of template correlation were compared after data were prepared by frequency filtering with various lowpass parameters (in some comparisons, 6.29 to 2.38 Hz) and sometimes by vector filtering. Accuracy of latency estimation increased exponentially with the template SNR. Regardless of the SNR, template cross-correlation was better than peak picking. Vector filtering helped, but with lower lowpass frequencies, the differences were rather small (in one comparison, the optimum lowpass cutoff was 1.76 Hz). Vector filtering was most useful when overlapping components of different distributions were simulated.
P50 is a more difficult component to measure than P300 because its amplitude is 10% to 25% that of P300. Typically measurements have been made by human observers picking peaks from averages of 32 trials. The ratio of P50 amplitudes to paired conditioning (S1) and testing (S2) stimuli is calculated. Ratios are less reliable than measurement of the numerator or denominator alone because ratios combine the statistically independent noise of both measures (3, 9). Two studies have found the reliabilities of P50 amplitude ratios to be less than 0.15 (7, 38). Freedman (24) has emphasized the importance of using only moderate intensity clicks and recording with the subject in a supine position for minimizing muscle artifact. Cardenas et al. (9) showed that the reliability of S2/S1 could be improved by applying the dipole modeling method of Turetsky et al. (70) to averages of 110 to 120 trials filtered with a 10- to 50-Hz bandpass. Even though reliability for peak picking was only 0.27 (interclass correlation of 6 repetitions), it was 0.63 for a model that fit a single source simultaneously to P50s evoked by S1 and S2. One caveat about reliabilities from dipole modeling is that complex computational methods can achieve results that turn out to be artifactual in simulations. These checks have yet to be made.
Accuracy of Source Localization
To accurately locate brain sources a number of known error sources must be controlled. Electrodes or magnetic sensors must be accurately placed in relation to the skull. A precise alignment of dipole and structural brain images must be made and the SNR must be enhanced. Assumptions of mathematical models for computing the dipole must be met, including assumptions about sphericity, conductivity (in the case of an EEG), and the temporal stability of sources. The size of the error made by the assumption of a spherical head shape was explored by Law and Nunez (40). Using a three-dimensional digitizer, they located 62 positions on an electrode cap. An ellipsoidal shape fit the electrode positions better than a sphere. Law and Nunez described a method for determining by tape measure the three axes of the shape conforming best to the head of an individual subject.
One presumed advantage of a MEG over an EEG was that the former affords more precise localization of sources. Controversy about this point, stimulated by a report by Cohen et al. (10), reached the news section of the magazine Science (12). Cohen et al. created an artificial source by passing subthreshold current through depth electrodes implanted in three patients for seizure monitoring. The exact locations of the electrodes could be determined from roentgenographs, and these locations were compared to those calculated for dipoles based on MEG and EEG recordings, each from 16 head locations. The average error for a MEG was 8 mm and for an EEG, 10 mm, thus showing no significant advantage for the MEG. In a follow-up study from the same research group, Cuffin et al. (13) calculated additional EEG dipoles using the same method and found an average localization error of 11 mm.
The studies above used artificial sources. Baumann et al. (5) tested the between-session reliability of dipole parameters from the P1m (50 msec), N1m (100 msec), and P2m (165 msec) components of an auditory ERF. Spatial parameters had an absolute difference of 3 to 10 mm. Errors were attributed to changes in attention, SNR, and local asymmetries in head shape. The sizes of sources detected by MEG after sensory stimulation have been estimated by Williamson and Kaufman (73) to be between 40 and 400 mm2. These are intermediate in size between macrocolumns of the visual cortex and a full sensory area, which can be several square centimeters.
A consensus statement by a group of scientists (2) pointed out that EEG and MEG should be considered complementary, because their different sensitivity to dipoles of different direction and depth gives valuable information about neural organization. The MEG is most sensitive to activity in fissures of the cortex where currents flow tangentially and to superficial sources, whereas the EEG is sensitive to both radial and tangential currents and is more sensitive than the MEG to deep sources, since in the MEG there is minimal magnetic field spreading by volume condition. The MEG has the advantage of being independent of inhomogeneities in concentric conductivities, whereas localization by an EEG depends on how accurately these conductivities can be approximated. Information from MRI and models of the real geometry of the head are needed. Additional advantages of the MEG are that it requires no electrode placement and permits very slow frequencies to be measured. On the other hand, it is not portable and is sensitive to ambient noise. Until recently, the MEG has had a limited number of channels, and its sensors have been relatively large, with diameters of 3 cm or more positioned at least 1 cm from the scalp.
EEG and MEG localization is comparable to the best 15O positron emission tomography (PET) resolution (6 to 100 mm), but both EEGs and MEGs have certain advantages over PET: the sample time of O15PET is 45 to 60 sec in contrast to the millisecond resolution of an EEG or MEG, PET requires administration of radioactive materials, and PET facilities are much more expensive than even MEG facilities (27). In addition, important neural events may not be concentrated enough to increase blood flow regionally. For example, Eulitz et al. (18) had subjects respond to nouns every 6 sec by silently articulating related verbs. Subjects repeated the task during separate sessions of MEG recording and PET imaging. In two regions, one in Wernicke's area and one in Broca's area, cerebral blood flow was increased on PET. Analysis of the MEG showed that during the first 200 msec of the 6-sec interval, a single current dipole was present in the primary cortex, but thereafter multiple dipoles appeared that were not confined to the regions of increased blood flow. Of course, it is somewhat misleading to cast PET and EEG/MEG as direct competitors because the two methods are most valid in different realms. Only PET assesses blood flow, disturbances of which are often the primary cause of brain dysfunction.
A framework for combining from EEG, MEG, and MRI data has been provided by Dale and Sereno (15). Such a combination of data makes possible the identification of plausible multiple cortical sources with a spatial resolution as good as PET but with a much finer temporal resolution. When available, PET and functional MRI data, can be added to the reconstruction.
Modern multichannel EEG and MEG recording have expanded many fold the amount of data recorded from each subject, leading to problems of statistical inference. This can be seen graphically, for example, when the probability of statistical difference between two groups is plotted across electrode sites (this has been called significance probability mapping). Groups usually differ by at least one electrode, and if they differ at one electrode, they tend to differ at adjacent electrodes, creating regions of significant difference. Of course, because there are multiple electrodes and because data at adjacent electrodes tend to correlate, the extent of significant difference often appears greater than it is. For correct statistical inference, the number of variables must somehow be reduced. Because data between time points and between topographic locations are often highly correlated, breakdown into components, factors, or dipoles as outlined above is possible. Even then, too many variables may remain for the number of subjects that can be tested.
The best way to avoid type I errors (rejecting the null hypothesis when it is true) is by replication of initial findings on a second data set, distinguishing between exploratory and confirmatory data analysis. In the exploratory phase of research, it would be foolish to limit data collection to a few variables chosen to test definitively a few a priori hypotheses. For clinical studies, the second data set needs to come from an independent clinical sample. Less satisfactory than the two-step approach of confirmation of exploratory findings is the application to a single data set of Bonferroni corrections or leave-one-out (jackknifing) methods. The latter sequentially leaves out one subject from the data set and determines how well a discriminant function based on the other subjects classifies the one. The cost of the Bonferroni correction is high, since it increases the likelihood of type 2 errors (accepting the null hypothesis when it is false). It should be noted that demonstrations of statistically significant replicability do not guarantee that significant neural events have been observed—artifact can be highly replicable too.
The application of evoked MEG and EEG tests to clinical diagnosis has the same requirements as for other clinical tests. To establish the usefulness of a test, well-accepted standards should be used to define the disease, the test should be evaluated on a population different from the one used to derive the test, and the test should have a low false-positive rate, or if it is meant to exclude a diagnosis, a low false-negative rate (49). A few definitions need to be kept in mind: a true positive (TP) is a positive test in a patient with the disease, whereas a false positive (FP) is a positive test in a patients without the disease. A true negative (TN) is a negative test in a person without the disease, and a false negative (FN) is a negative test in a person with the disease. Sensitivity = TP/(TP + FN) and specificity = TN/(TN + FP). Positive predictive power = TP/(TP + FP) and negative predictive power = TN/(TN + FN).
In psychiatric contexts, ERPs have generally been considered a way to investigate cognitive or biological differences between already-diagnosed patients and controls, rather than a way to make a diagnosis. This has been the case even for the most replicable ERP findings such as P300 amplitude reduction in schizophrenia and P300 latency prolongation in dementia. Occasionally, the diagnostic usefulness of ERPs in psychiatry has been debated as in the pair of articles discussing the pros (28) and cons (54) of P300 latency in assessing dementia. Goodin (28) points out that in neurology, brainstem auditory ERPs are very sensitive in diagnosing cerebellopontine angle tumor, with a false-negative rate of less than 3%. The EEG is useful in diagnosing suspected epileptics, although its sensitivity is only 52% because it is 96% specific. However, P300 latency is limited for diagnosing dementia because its sensitivity in some studies is less than 60%, but since its false-negative rate is low, a negative result can give valuable information in some contexts. Of course P300's usefulness presumes that it can be elicited reliably in the population to be tested, which some studies affirm (more than 95% of subjects had adequate P300s) and one denies (less than 20% had adequate P300s) (54).
Pfefferbaum et al. (54) argue that better discrimination between demented and nondemented patients can be made if the effects of age itself are taken into account by regression analysis. They point out that the sensitivity and specificity of a test depends on the cutoff used to define abnormality and the prevalence of the disease in the population. The trade-offs between sensitivity and specificity at various cutoffs can be depicted in a receiver operating-characteristics graph. In the data of Pfefferbaum et al. (54) a statistically optimal cutoff for discrimination between demented and nondemented neurological and psychiatric patients yielded a specificity of 93% and a sensitivity of 38%. Thus, P300 latency is unsuitable for screening because of the low sensitivity, but might be more useful for confirmation of diagnosis because of its higher specificity. In a low-risk population, however, the specificity of P300 is likely to be even lower. A fundamental problem in dementia testing with P300 is that the paradigm used so far to elicit P300 requires the subject to perform a task that severely demented patients may be unable to do, or do in a way that results in P300s with low SNRs. ERP or ERF components less dependent on subject cooperation may play a greater role in clinical assessment in the future.
Ford et al. (21) did a sensitivity–specificity (receiver operating characteristics) analysis of the utility of P300 in diagnosing schizophrenia. Using data originally reported in Pfefferbaum et al. (55) they expressed P300 amplitudes of 20 schizophrenics, 34 depressed, 37 demented, and 9 nondemented patients as age-corrected z-scores based on P300 data from 115 control subjects. Diagnosis of schizophrenia on the basis of P300 amplitude was less successful than the diagnosis of dementia on the basis of P300 latency: a specificity of 90% corresponded to a sensitivity of only 15%. However, P300 amplitude could be used to rule out schizophrenia in certain cases: no patient with a z-score above 1.6 was schizophrenic.
The methodology of evoked brain potential and magnetic field studies is in a phase of rapid technical evolution. A 122-channel MEG system is already on-line in Finland (72). In the future more and more studies will coordinate EEG and MEG data with data from MRI, PET, and SPECT scans. The claims of analysis methods to identify actual brain sources will be tested. Electrical and magnetic localization and other imaging methods will vie with each other in precision. Not just the sources of ERP and ERF components to simple stimuli will be localized, but also those reflecting more complex cognitive processes. The application of these new methods, particularly magnetic field measurement, to psychiatric disorders has hardly begun. We hope and expect that this situation will change in the near future.
To calculate the potential on the surface of a sphere, the following equations must be satisfied (14). For a Px dipole located along the radial projection at distance f from the center of a sphere
}
For a Py dipole located along the radial projection at distance f from the center of a sphere
For a Pz dipole located along the radial projection at distance f from the center of a sphere 
where Pn and {
} are Legendre polynomials and for n = 1 to 30 iterations
and where
ACKNOWLEDGMENTS
Preparation of this chapter was supported by the National Institute of Mental Health, grants MH30854 and MH40052, and by the Department of Veterans Affairs.
We thank Margaret J. Rosenbloom for her critical comments.
发布于12月4日 22:29 | 评论数(0) 阅读数(958) | 我的文章
Methodological Issues in Event-Related Brain Potential and Magnetic Field Studies
Methodological Issues in Event-Related Brain Potential and Magnetic Field Studies
Walton T. Roth, Judith M. Ford, Adolf Pfefferbaum, and Thomas R. Elbert
Psychiatry in its search for the roots of abnormal thoughts, feelings, and behavior has again turned its attention to the human brain and is trying to apply the methods of the many scientific disciplines that have cast light on normal brain functioning-disciplines such as neuroanatomy and histology, biochemistry and molecular biology, and electrophysiology. This chapter concentrates on ways of maximizing what can be learned from noninvasive electrophysiology, a technique that is singular in its ability to record millisecond-by-millisecond changes in the brain following repeated external or internal events. Although the triggering events are often simple sensory stimuli, the cognitive processes that follow them and leave their trace in fluctuating voltage or magnetic fields can be quite complex. In the last decade competing noninvasive techniques such as positron emission tomography (PET) have challenged the preeminence of electrophysiology, particularly in spatial localization of brain processes. This challenge has stimulated a number of technological and methodological developments in acquiring, analyzing, and presenting brain electrical and magnetic data. But before we review these developments, we remind you of some basic principles and give examples of their relevance to psychiatry (see also A Critical Analysis of Neurochemical Methods for Monitoring Transmitter Dynamics in the Brain, Electrophysiology, and Pharmacology and Physiology of Central Noradrenergic Systems for related discussion).
Nerve cells generate extracellular current flow by fluctuations in the slower changing membrane potentials of dendrites and cell bodies. Postsynaptic potentials cause an outflow of negative (excitatory) or positive (inhibitory) ionic charges into extracellular fluid, which are then pumped back into the cell. This current flow, when summated, results in volume-conducted potentials recorded at the scalp as the electroencephalogram (EEG). Event-related potentials (ERPs) are EEG changes that are time-locked to sensory, motor, or cognitive events. They have provided a way to evaluate brain functioning in mental disorders and the effects of psychoactive drugs. Recent conceptual and technical developments have greatly expanded our capability to understand and document the mechanisms underlying surface recordings. Particular attention has been paid to identifying the location, orientation, and distribution of current dipoles (pairs of opposite charges) that may be the sources of scalp-recorded electrical activity.
Nerve cells also generate intracellular current flow from dendrites to cell body. This flow results in a magnetic field that can be detected at the scalp as a magnetoencephalogram (MEG), even though it is a billionfold less intense than the earth's magnetic field. Event-related magnetic fields (ERFs) can be elicited and time-locked to specific events and are analogous to ERPs. Magnetoencephalograms and ERFs convey different information than EEG and ERPs. This is because voltage fields on the surface of a sphere, which the skull enclosing the brain approximates, are produced equally well by dipoles oriented radially and tangentially with respect to a radius of the sphere. In contrast, 90% of the magnetic field at the skull can be ascribed to tangential dipoles alone. This is a consequence of the geometrical orientation of masses of nerve cells and of magnetic sensors. Fig. 1 illustrates how dipole orientation can be either correlated or random for different gyri and sulci. Parallel dipoles lying tangentially on sulcal walls contribute much more to the MEG than random dipoles or dipoles lying radially along the crowns of gyri.
Why are the methodological issues that this chapter addresses relevant to psychiatrists and psychologists? First, ERPs and ERFs are theoretically relevant because they provide ways of testing theories of abnormal brain functioning that no other methods can offer. For example, unlike ordinary behavioral tests of cognitive processing, ERPs give an index of the processing of task-irrelevant events, distracting stimuli, or events subjects have been told to ignore. The topographic distribution of ERPs and ERFs gives clues as to what parts of the brain are active during a particular cognitive activity. Second, ERPs and to a less extent ERFs have been demonstrated empirically to be relevant. ERP abnormalities have been repeatedly observed in psychiatric disorders, notably in the P300 and P50 components. The P or N signifies positive or negative and the number is the mean peak latency in milliseconds. Thus, the P300 component is a positive potential that occurs approximately 300 msec after a stimulus that is infrequent and in some way relevant. The most venerable and consistent psychiatric ERP finding is that of reduced P300 amplitude in schizophrenics (60), although this is not specific to schizophrenia (see refs. 59 and 22 for reviews). For instance, a longitudinal study demonstrated that lower P300 amplitude at age 15 was predictive of poorer global personality functioning at age 25 (66). Latency at P300 is generally greater in patients with dementia than in normals or in patients with schizophrenia or depression (28, 54). Recently, psychiatric attention has been directed to P50, an ERP component to auditory stimuli whose amplitude is suppressed if the eliciting stimulus is paired with another that precedes it by one-half second. Schizophrenics show less P50 suppression than controls (25) as indicated by smaller amplitude ratios (P50 to the second stimulus of a pair divided by P50 to the first), although again this finding is not limited to schizophrenia (4).
Abnormalities of ERPs in psychiatric patients can be interpreted in light of a considerable amount of knowledge that has accumulated about the significance of certain ERP components in normal human information processing. For example, P300 is known to reflect the categorization of events, depending jointly on stimulus probability, stimulus significance, and the information value of the event (36). Probably, P300 has multiple, partially asynchronous generators (58). Components occurring 60 to 100 msec after onset of auditory stimuli, including N100, have been shown to reflect selective attention to auditory stimulus channels (42). In contrast, auditory ERPs with latencies less than 10 msec are insensitive to attention effects but give a unique assessment of the intactness of brainstem circuitry (32).
The literature on ERFs in normal subjects is quite extensive although magnetic recording techniques have been available only a relatively short time. Much of that literature has documented the existence of ERF components that parallel those established by invasive and noninvasive ERP recording. However, to date, most clinical MEG studies have been done in neurological rather than psychiatric patients, although that is likely to change in the near future. Reite et al. (57) recorded ERFs in six medicated, paranoid schizophrenic patients and six normal controls. The M100 component (analogous to the N100 of the ERP) showed less interhemispheric asymmetry in schizophrenics and had different source orientations in the left hemisphere. Tiihonen et al. (68) compared the M100 component in two schizophrenic patients when they were experiencing auditory hallucinations and when they were not. During hallucinations, M100 peaked approximately 20 msec later, an effect similar to that of external masking noise in normals.
We now turn to methodological trends that are transforming ERP and ERF research. Specific topics include data acquisition, signal averaging, ocular artifact, choice of reference electrodes, digital filtering, measuring components including dipole modeling, and statistical and diagnostic considerations.
Electroencephalogram Systems
Older electroencephalographic tube-based amplifiers have been completely replaced with high impedance solid-state amplifiers with electronically controlled amplification and filter settings. In many laboratories, pen-chart recorders have been replaced with electronic data storage and display systems, but paper records are still widely used for visual analysis of diagnostic EEGs and sleep. Laboratory computers are constantly evolving toward faster, cheaper, and more powerful models. New storage media based on tape or magnetic or optical disks permit archiving of data from many subjects in an easily retrievable form. As welcome as these advances have been, they have generated difficult new choices for researchers. Should they buy commercial EEG and ERP hardware and software systems or develop their own? Which commercial systems or routes to laboratory-program development are satisfactory? Commercial systems tend to be limited in flexibility, details of data analysis may be a trade secret (which is unacceptable scientifically), and access to raw data for special analyses may be difficult. Laboratory-developed systems require deciding among manifold hardware and software possibilities, and then allocating many hours to programming. As will be learned from this chapter, methodologically up-to-date ERP analysis requires much more than eye-movement artifact rejection and signal averaging.
Whereas the conventional 10–20 system of Jasper (35) used 19 electrodes with a typical distance of 6 cm between them, some investigators have greatly expanded the electrode arrays in order to record more of the spatial detail present in the EEG. Thus arrays of 124, or even 256 electrodes, which yield interelectrode distances of 2.25 and 1.6 cm, are now being advocated (27) and have been shown to enhance localization. The application of multiple electrodes is a lengthy, labor-intensive process, which requires care in scalp preparation and accuracy in electrode placement. For localization studies relating EEG or MEG data to brain structures visualized by magnetic resonance imaging (MRI), it is important that electrodes be aligned correctly according to skull landmarks, and fiducial markers visible in MRI scans are used. (Vitamin E capsules are easily available and the right size.)
Electrode application entails a potential health risk to both subject and technician if the intactness of the scalp is compromised by procedures to reduce electrical resistance between electrode and scalp or by skin lesions. Acquired immunodeficiency syndrome and hepatitis B can both be transmitted by this route, so it is absolutely essential that proper precautions be taken. Putnam et al. (56) give recommendations for disinfecting reusable electrodes and for protecting the technician.
Magnetoencephalogram Systems
The recording of the MEG has been made practical by the development of superconducting quantum interference devices (SQUIDs) that are sensitive to minute magnetic fields. The MEG technology is much more expensive than the EEG technology. Not only are the SQUIDs themselves expensive, but they require provision for liquid helium at 4.2°K to cool them, and a recording room shielded with a high-permeability material against magnetic fields and with aluminum against eddy currents. The liquid helium is kept in a vacuum-insulated container called a dewar. Locating magnetic sources requires recording from multiple sites, preferably simultaneously. Otherwise, separate stimulation runs must be made, moving sensors from one location to another between runs. More runs take more recording time and increase the likelihood that the subject's mental state will change, altering the sources. A MEG system with over 30 channels costs approximately $3,000,000, 100 times more than the same number of EEG channels. Because MEG prices reflect the cost of research and development more than construction of the apparatus, the price per unit would drop if more units were sold. In one system, 37 sensors are placed 2.2 cm apart to cover a single hemisphere (12).
An advantage of MEG sensors is that they do not touch the head, so transmission of infectious agents is of less concern. Fixation of head position is critical so that sensors can be aligned according to skull landmarks. Modern SQUID technology allows recording of signals that vary slowly over a minute, undisturbed by electrode drift. A new method for recording even slower or static magnetic fields converts such fields to more rapidly changing fields by having the subject lie on a mechanically driven platform that executes a circular movement of a few centimeters at 0.2 Hz (26). Auditory and visual stimulation cannot be given by conventional earphones or CRT displays because of their magnetic properties. Instead, sounds have to be delivered from outside the testing chamber through hollow tubes and visual stimuli projected through a window in the magnetic shield or delivered fiber optically.
Both ERPs and ERFs benefit greatly from signal averaging to enhance their signal-to-noise ratio (SNR). Data are generally digitized at a fixed rate to fill a data array, and a stimulus or other synchronizing event defines the time epoch of interest within this array. The event is repeated (each repetition is called a trial), and a time-locked signal (ensemble) average is calculated across trials epochs for each time point of the epoch. If Xj(t) is the electrical potential (voltage) or magnetic field strength at some electrode or sensor location at time t and trial j, the signal average is defined as
If Xjt is considered the sum of true signal mt and random noise Njt (background EEG and measurement error), signal averaging improves the SNR. Unbiased estimates of signal power , noise power , and SNR can be calculated as follows (71).
One of the assumptions of signal averaging is that the signal is invariant across trials. This assumption is violated when the amplitude of the ERP component of interest habituates or when its latency varies from trial to trial, as is clearly the case for components related to certain cognitive processes, such as the P300. One way of dealing with component latency variability is to locate the signal on each trial and align the trials on these signals rather than on the eliciting stimulus. Woody (75) proposed an iterative procedure (an adaptive filter) that located the signal on each single trial by moving a template (initially the signal average) by time increments along the trial to find the latency of maximum correlation. A new average was then formed by aligning trials on the identified signal latencies, and the new average was used as a new template. If the SNR is too low, this procedure produces results that simply reflect random noise. Gratton et al. (31) tested the procedure with simulated signals and background EEG noise and demonstrated that iterations (up to three) were important only when the original template had a wavelength on the order of two times longer than the signal.
Roth et al. (56) used this procedure to analyze ERPs elicited from schizophrenics and controls performing an auditory choice reaction time paradigm in order to test whether P300 amplitude reduction in schizophrenics could be attributed to latency variability. They found that individual trial P300 latency was indeed more variable in schizophrenics but that schizophrenic P300 amplitude was still smaller than control amplitude after latency adjustment. To reduce distortions due to noise, Pfefferbaum and Ford (53) modified the procedure by only including trials whose covariance is greater in the part of the epoch where signal is expected than in the part where noise is expected, and whose correlation with the template (initially a half-sine wave) exceeds a set threshold. Using this modified procedure, Ford et al. (23) replicated the Roth et al. (61) finding that schizophrenic P300 remained smaller. Furthermore, schizophrenics had more trials that did not pass the covariance–correlation screen than controls. Trials that did not qualify for latency adjustment had longer reaction times, showing that they were deviant behaviorally as well as electrophysiologically. In addition, Ford et al. calculated for each subject the covariance of P300 signal average across trials with that subject's EEG in single signal epochs and in single nonsignal epochs. The ratio of mean signal covariance to mean noise covariance was significantly smaller in the schizophrenics. Because trials were filtered with a bandpass of 0.5 to 4.4 Hz, noise was EEG activity in the frequency range of P300 rather than higher frequency like a, b, or muscle activity.
Another assumption of signal averaging is that background EEG noise is random noise. This is only an approximation to the truth, as a study of event-related spectral perturbation indicates (41). In normal subjects, auditory tone pips reliably produced momentary increases in spectral power in the 2- to 8-Hz and 10- to 40-Hz bands.
Eye movement and blinks produce electrical potentials and magnetic fields that are often much larger than those deriving from brain sources. The magnetic fields are more restricted to the vicinity of the eye than are the electrical fields and for this reason are less troublesome if unsynchronized with events of experimental interest. Synchronized eye artifact can cause major errors in peak measurement or source localization. Attempts to control this artifact by instructing subjects to fixate their gaze on a point or not to blink are often ineffective, particularly if the subject is psychotic or cognitively impaired. Thus methods for removing eye artifact from the ERP or ERF need to be applied. Many are based on determining the coefficients Ak in the equation
V(k,t) = Ak * EOG(t) + EEG(k,t)
where V(k,t) is the voltage observed in lead k at time t, and EOG(t) and EEG(k,t) are the true EOG and EEG voltage contributions at that time.
Spatial-temporal dipole models of eye movements and blinks make it clear that the same correction cannot be used for both (6). Thus eye-correction procedures should include at a minimum the following steps: (a) Separate blinks from movements on the basis of their temporal properties, (b) calculate separate linear regressions for the propagation of artifacts from each, and (c) correct EEG leads by the amount predicted by the regression coefficients. Gratton et al. (29), whose method has been used by a number of investigators, adds an additional step of subtracting signal averages from individual trials to avoid distortions resulting from ERP effects in both EEG and EOG records. A computerized implementation of this procedure that adjusts for both a vertical and a horizontal EOG channel, has been developed (43). Although certain technical issues in implementing EOG corrections remain unresolved—the proper number and position of EOG electrodes, the error attendant upon assuming a linear relationships between the EOG signal and EEG artifacts, the implications of the presence of EEG artifacts in EOG leads, how to deal with overlapping eye movement and blinks, and instability of individual propagation factors between sessions and even between tasks within a session (19)—the use of such off-line procedures have greatly increased the number of trials available for analysis in clinical studies.
Whereas MEG sensors detect the absolute magnetic field at a given location in space and need no reference in the body, the EEG must be measured as voltage differences between two points on or in the organism. Ideally one point should be close to the biological voltage source under investigation, and the other should be a reference point with constant voltage or at least a voltage not correlated with the source voltage. Traditional references for human ERP have been linked mastoids, linked ears, or the nose; unfortunately none of these is unaffected by brain sources. Special disadvantages of linked ear references include the possibility that shorting can reduce asymmetry if resistance is low, and the possibility that artifactual spatial asymmetry will result if resistances at the two ears are not equal (48). Shorting is not a serious consideration as long as skin-electrode resistance at each ear is greater than 5 kW, because in that case scalp path resistance is reduced less than 5% (44). Resistance at the two ears can be balanced with a potentiometer, or one ear (say A1) can be used as a reference and recorded as a separate channel. Then a linked ear reference for say Cz, a scalp electrode in the 10–20 system, can be created algebraically, (Cz - A1) - (A2 - A1)/2 = Cz - (A1 + A2)/2.
To avoid active reference electrodes on the head, some investigators have turned to noncephalic (e.g., sternovertebral) electrodes (67). Unfortunately these electrodes are liable to pick up heart activity even when adjusted to be at right angles to the main vector of voltage during the cardiac cycle, since cardiac depolarization and repolarization vectors do not maintain a perfectly constant direction over the cycle.
Another solution is to use an average reference. At each time point, an average reference defines zero over C electrodes in a data array A as
A limitation of the average reference is that when electrodes are not densely and equally spaced around the brain, for example, there are none at the bottom of the head (69), the sum in the formula above is generally different from true zero. For example, Desmedt et al. (16) have shown that P14 of the somatosensory evoked response, which is present with a linked ears reference, disappears when a zero reference based on 27 scalp electrodes is applied, becoming surrounded by "ghost" negativities. A linked-ear reference reflects more accurately the medial lemniscal volley that is the presumed basis of P14. In addition, local changes can be mistaken for global changes with a zero reference. These distortions are less likely to affect tangential than radial dipoles.
In conclusion, there is no perfect reference for all cases. As a general principle, a known local source should be referred to an electrode distant from it.
Before measurements are made on ERPs or ERFs, it is useful to apply SNR-enhancing filters that incorporate assumptions about frequency, timing, and spatial distribution of the component of interest. For example, the ERP P300 component may be expected from experiments in the literature to have a frequency lower than 2 Hz (30), to peak in a range of 280 to 400 msec (in a simple auditory choice reaction time task in young adults) and to be maximal at Pz, another electrode in the 10–20 system. Though signal averaging attenuates unsynchronized noise at every frequency as it improves SNR, frequency filters are commonly applied prior to component measurement. These filters are useful whenever the frequency of the noise is different from that of the signal.
Digital Filters
Digital frequency filters (11) have the advantage over analog filters of being able to operate without introducing distorting phase shifts into the signal. The most commonly used digital filter has been the moving average or boxcar filter, in which each point of the signal is replaced by an average of that point and a certain number of prior and subsequent points. This is only possible for stored data, because it makes use of future time points to calculate current output. Farwell et al. (20) have shown that a simple moving average filter does not prepare average and single-trial waveforms as well for P300 peak-picking as does a filter designed by an optimizing algorithm. Such an algorithm determines a set of weights that are able to reduce deviations (ripple or ringing) in the passband and stopband of the filter. Optimized filters have less tendency to reduce P300 amplitude or distort shape and, in the case of averages, gave more stable latency measurements. For P300, the authors recommend that the optimum filter have a passband cut-off frequency of 6 Hz, a stopband cut-off frequency of 8 or 8.5 Hz, and use 490/n points, where n is the sampling interval in milliseconds. It should be emphasized that analog filters still have a place in data acquisition prior to digital filtering—a low-pass analog filter with a half-power frequency below but close to half the sampling rate prevents aliasing, and, for P300 recording, a high-pass analog filter with a half-power frequency of less than 0.16 Hz minimizes irrelevant baseline shifts (20).
Spatial Filters
Current source density maps (also called surface Laplacian or radial current estimate maps) act as spatial filters emphasizing localized components with a high spatial frequency. For this to work well of course, electrodes must be placed with a high spatial frequency. Maps can be made of unaveraged activity such as epileptic spikes or of signal averages. Sensory ERP components show a more localized distribution using this approach than in voltage maps. For example, Nagamine et al. (46) compared voltage and current source density maps on the scalp ERPs obtained by tibial nerve stimulation. The results for a single subject presented in Fig. 2 demonstrate better localization for P40, N50, and P60 for the current source density map. The equation for calculating current source density is I = r(d2V/dx2 + d2V/dy2), where V is the voltage, x and y the surface location on the x–y plane, and r the charge density. In addition, r = k * d2, where d is the distance between electrodes and k is a constant for all electrodes within a subject. The Laplacian operator can give limits for finding equivalent dipoles. It has a physical interpretation—local radial current flow from the brain into the scalp and vice versa—but it is different from dipole modeling (described below) and is free of dipole modeling's ambiguities.
In the Laplacian calculation, surface contours can be generated by a method called spherical spline interpolation, which is based on physical principles for minimizing the deformation energy of a thin sphere constrained to pass through known points (51). This produces a smooth surface running through the data values and filling in between them, even when electrodes are irregularly placed on the scalp. Spherical splines have advantages over plate splines, which are based on deformation of an infinite thin plate. As might be expected from the fact that interpolated values at any point are derived from data from other locations, coherence (a measure of covariation) is inflated by interpolation. Nearest-neighbor interpolations are less smooth and inferior for locating extrema (peaks and troughs must lie on an electrode site) but do not inflate coherence.
Gevins et al. (27) have demonstrated a method of current source density mapping they call finite element model deblurring that they believe is superior to the Laplacian method. Mathematically, it is a less computationally demanding version of dipole modeling known as spatial deconvolution, which assumes that all dipoles are located on a cortical surface. Gevins et al. use the subject's head MRI to provide information about conducting volumes between scalp and cortical surfaces.
A simpler spatial filter, the vector filter (30), has been used for component measurement. Its output is the weighted sum of data points at different electrodes. Conceptually, measuring a component at one lead is the same as applying a vector filter with weight 1 assigned to values at that lead and weight 0 to values at all other leads. Vector filtering assumes that the distribution of the component to be measured is constant despite changes in amplitude or latency. The crux of the procedure is how to specify the weights: using three 10–20 system scalp electrodes, Fz, Cz, and Pz, weights of 0.15 for Fz, -0.53 for Cz, and 0.83 for Pz were found to produce optimal discrimination in an oddball paradigm between rare trials, which contain substantial P300s, and frequent trials, which do not (30). Thus, optimum weights do not necessarily correspond to component distribution, because P300 is larger at Cz than at Fz. Dipole modeling, which is described below, can act as both a spatial and temporal filter.
Measurement Methods
A component can be defined as electrical or magnetic activity associated with a specific neurological or psychological process, for example, a motor act such as moving one's finger, a sensory process such as the reaction to a light flash, or a cognitive process such as categorizing a stimulus as target or nontarget. In a statistical sense a component explains experimental variance. The details of the experimental method are part of the operational definition of a component. As more experiments are done, theoretical expectations about components develop into generalizations. For example, many experiments in which subjects performed a fixed foreperiod reaction time task have resulted in a parietal–central negative shift prior to the button press. A natural generalization is that the parietal–central shift represents preparation for a motor act. Furthermore, because the source of the recorded data is a physical location within the brain, the ultimate description of a component must include reference to the specific brain structures activated. Some leads or sensors will pick up activity from those structures better than others, particularly when sources are multiple with overlapping influences. In the case of ERPs, the choice of voltage reference influences how electrical activity from a source appears in the EEG recording.
Measurement procedures include peak picking, area measurement, waveform subtraction, principal components analysis, template correlation, and dipole modeling. Peak picking means finding maxima or minima in specified latency ranges and determining peak latency and amplitude with respect to a prestimulus baseline. This is the simplest method of component evaluation, but can be biased when latency ranges are selected after an inspection of the data, and is perhaps unduly restricted in that it considers only peaks among other waveform features. In addition, it is often based on only one point, which may be influenced by noise or overlapping components. With multiple leads, another limitation of peak picking becomes obvious: what appears by shape to be a single component has maxima at different time points in different leads, and it is not clear how best to resolve the discrepancies. Furthermore, the choice of reference electrodes can determine when peaks and troughs appear.
Area measurement is sometimes used when the component is believed to be more rectangular than peaked. Area is measured in a specified latency range, and is thus based on multiple points, but area measurement, like peak picking, can be biased and is influenced by overlapping components.
Waveform subtraction can be used before peak picking or area measurement to reduce the effects of component overlap. For example, consider a paradigm where tones of two pitches are given in an unpredictable sequence and one occurs less frequently and is designated as the target of some task. The ERP to the rare tone can be considered a combination of the sensory effects of the tone and the cognitive effects of the tone being a rare target. By subtracting the ERP to the frequent tones from the ERP to the infrequent tones, the sensory effects are removed leaving behind the cognitive effects. This assumes that the sensory responses to the two tones are identical and that cognitive and sensory effects are additive, an assumption that is not always warranted. For example, frequency-specific temporal recovery of the auditory N100, a noncognitive effect, makes the response of N100 to frequents smaller than the response of N100 to rares.
Principal components analysis (PCA) is another approach to ERP component measurement, which uses the time points on waveforms from different subjects, different electrodes, and different experimental conditions to define components. In statistical terms PCA identifies orthogonal axes of maximal variance in a multidimensional space defined by the variables. Generally these axes are rotated according to the varimax procedure. Less arbitrary than peak picking, PCA makes no assumption about the latency range in which specific components will be found but only that they have a fixed latency across conditions and subjects. It has some ability to separate overlapping components. However, PCA is not completely free from arbitrariness. First, PCA solutions are not unique. Many rotations of the factors are possible. Second, results depend to a certain extent on what experimental conditions are chosen and how many leads are included. Variance from electrodes, subjects, conditions, and correlated noise are all treated the same. Furthermore, each experiment gives slightly different factor structures, and there is no established criterion for deciding whether these differences are significant or not. Thus, it is uncertain how many statistical components to interpret, and how to identify these components with ones previously described.
Template correlation assesses the similarity of a template of the component to the waveform to be evaluated. The template may be based on prior knowledge of the component shape or on signal averages (see the iterative Woody filter procedure described above). The template is usually compared to waveforms at specified intervals over a designated latency range to identify the latency of maximum correlation (or in one variation, maximum covariance). This time point is defined as the peak. The sum of cross products at this time point or the difference between amplitude at this point and a baseline can define amplitude.
Interpreting latency data under different experimental conditions can be difficult when multiple leads are involved. Latency may vary at different leads and topography may vary under different conditions, implying different components whose latency cannot be compared. To solve these problems, Brandeis et al. (8) spatially generalized the Woody filter procedure using an average reference map, and applying a measure they call global field power (GFP) defined by the following formula for an array A consisting of data from C electrodes:
Further, global dissimilarity (GD) is defined as the root mean square (rms) power of the difference maps calculated by subtracting two normalized GFP maps. The procedure is as follows: (a) Grand averages are used to form template GFP maps, from which component model maps at single latencies near 100, 200, and 400 msec are derived, corresponding to P1, N1, and P3 (see ref. 8 for details). (b) Component model maps are moved in specified latency ranges around the latency of each model's component. The minimum of GD multiplied by sequential dissimilarity (GD between current and previous map: a stability constraint) is calculated, and the minimum of this function (best fit) is defined as the map latency for that component. (c) In an iteration, the average of all normalized maps at their latencies of best fit is used as a new model, and the search window is set around the new mean latency. The results show that components can be identified by topography alone, without respect to amplitude or time. However, this method does not take into account possible overlapping components and would fail if such components influenced topographies. Furthermore, average references for P300, which is widely distributed on the top of the head, may be inferior to a noncephalic reference.
Dipole modeling is a method for reducing data from multilead EEG or multisource MEG by deducing the dipole sources that may have produced them. Although the forward problem (calculating scalp distribution from known dipoles) has a unique solution whose accuracy is limited only by the approximations of skull geometry and conductivities, the inverse problem has multiple mathematically valid solutions as was pointed about by Helmholz more than a century ago (33). The reason is that a single scalp distribution can be produced by different numbers of dipoles in different combinations of locations and orientations. Thus, various constraints on the number of sources allowed and their approximate location must be applied to reach a solution. Sometimes these constraints are so severe as to specify that the source be a single dipole located somewhere in the brain.
At an abstract level, dipole modeling is like PCA in that an equation U = C * S must be solved where U is an array of k electrodes at t times that represents the linear superimposition of the array S of m sources at t times multiplied by C weighing coefficients at k electrodes for m sources (62). Whereas PCA determines C and S from mathematical constraints, dipole modeling assumes that C depends on volume conduction from j dipoles at certain locations, assuming Ckj = f(rj,oj,ek ), where f is a nonlinear function of the electrode location vector ek and of the geometry of the source and the head. The dipole has a location vector rj and the orientation vector oj. Equations defining a 3-shell sphere model of the head with differing conductivities for scalp, skull, and brain are found in the appendix to this chapter. Using these equations to model dipoles at various depths, Pfefferbaum (52) demonstrated how increasing the thickness of the superficial extrasulcal subarachnoid layer of cerebrospinal fluid (CSF) or skull thickness might affect scalp ERP amplitudes and topographic distributions.
One procedure for the dipole modeling of ERPs was developed by Scherg and Berg (64). Their software is available commercially as brain electrical source analysis (BESA, from Neuroscan, Inc.). It models a window of points, assuming a finite number of equivalent dipoles with fixed location and orientation. In its recent version, it does not assume a parametric dipole magnitude function (like the decaying sinusoid of ref. 70) but computes a varying magnitude function over the window of points for each dipole. The BESA model is applied iteratively, calculating at each step the residual variance (percentage of recorded data not explained by the model). The first step looks for the inverse solution by calculating parameters of a plausible dipole from an EEG or MEG data map. Then forward solutions calculate resultant EEG or MEG maps from those dipoles. Hundreds of iterations may take place, stopping when the change in residual variance is less than some criterion, such as 0.001%. When more than one dipole is modeled, some may be fixed in position (but not in amplitude) while a new dipole is optimized. The results of these procedures depend among other things on the starting location and other parameters of a dipole. An iterative procedure may find topographically local optima that would not be optima if all locations and orientations were tested. Scherg and Berg (64) explained that multiple-source solutions are less arbitrary if spatial and temporal constraints are added. For example, two sources may be required to have a symmetry between hemispheres, radial and tangential dipoles, or lie in the supratemporal plane. How this method works is illustrated in 发布于12月4日 22:27 | 评论数(0) 阅读数(682) | 我的文章
欧洲研发出数字手 能让使用者拥有触觉 转载 并附英文
欧洲研发出数字手 能让使用者拥有触觉
--------------------------------------------------------------------------------
http://www.sina.com.cn 2005年12月04日 14:47 南方都市报
让机器手像人手一样灵活
预计两年后面世
义肢能像人手一样灵敏感知外界刺激吗？
如此灵活、完美的机器手臂似乎只可能出现在科幻电影里，但欧洲一家机器人实验室近来研发的一种通过电脑控制的仿人机器手——“数字手”，不但是全球首例能让使用者有“触觉”，将彻底颠覆“钳状机器手”的不雅外观的人工义肢，还具有以假乱真的人类皮肤“外表”。对于截肢者而言，这款仿人机器手可算一大福音。
“数字手”问世
“数字手”是六个科学家团队共同的智慧结晶，成员分别来自意大利、德国、西班牙和丹麦四国。
“数字手”耗费了研究人员3年半的心血，已耗资150万欧元(约合180万美元)。它是全球首例能接收并向使用者大脑传递外界信号的人工义肢。研发小组成员保罗·达里奥介绍，同以往的人工义肢相比，“数字手”最特别之处在于能使使用者“有触觉”。
仿真外观
研究人员估计，如果一切顺利，两年后“数字手”就可以推向市场，供病患安装使用。“数字手”内部为金属骨骼，但在罩上一只贴合金属的“人造手套”后，看起来与正常人手无异。
研究人员乔瓦尼·斯泰林说，很多病人都不太乐意安装外表如钳状的机器义肢，觉得“难为情”。这种“钳状机器手”二战以来一直在广泛使用，但精密程度根本无法同“数字手”相提并论。
“数字手”的问世将彻底颠覆“钳状机器手”的不雅外观。它可以安装在截肢者肘下，上面覆盖几层人造材料制成的“皮肤”，机器手因此触感更柔软，运动更灵活。
不过，研究人员巴顿承认，“数字手”现阶段在技术上仍存在一些问题。比如，制造机器手的材料是否会遭到人体“排斥”还是未知数。另外他们暂时也还不能确定使用者脑部如何适应这一人体“异物”的存在，以及“数字手”靠什么作动力。新华
影响
欧洲开创“机器人经济”
“数字手”是由设立在意大利的“波洛·圣安娜·巴尔德拉研究所”研制的，研发小组成员保罗·达里奥说，“数字手”的研制成功反映了欧洲在机器人研究领域的巨大潜力。
就欧洲在机器人研究方面的长处，达里奥信心十足地预言:“我们有一整套(协作)网络，我们知道如何合作，我们做好了完成(历史)飞跃的准备。”
但是，“数字手”项目若想继续发展，必须克服资金“瓶颈”。它已耗资150万欧元(约合180万美元)，全部由欧洲联盟一项专门赞助新兴科技的基金提供。欧盟委员会官员10月份为“数字手”项目向欧洲各国政府和企业申请更多资金赞助时，盛赞其取得的“巨大成功”。
欧盟委员会官员强调，如果欧洲想挖掘机器人产业的巨大经济价值，并与此领域中的强国如美国、日本、韩国展开竞争，增加投入至关重要。
委员会同时公布了欧洲和其他国家在机器人研究上资金投入的具体金额。每年欧盟委员会与欧盟各国在机器人研究上所花资金为8500万欧元(1000万美元)。日本和韩国花费数目与此相当，但美国研究资金投入高达5亿美元，大部分源于军事机器人的巨大需求。
http://tech.sina.com.cn/d/2005-12-04/1447782171.shtml
Cyberhand Leads Europe‘s Robot Efforts
Staff and agencies
03 December, 2005
By AIDAN LEWIS, 35 minutes ago
PONTEDERA, Italy - The metallic fingers close around yours in near-perfect synchrony, then tighten their grip as you try to pull away. For now, it is a computer that orders "Cyberhand" to greet you at the robotics lab where researchers have spent the past 3 1/2 years creating the first prosthetic hand capable of eliciting natural sensory signals.
Cyberhand would allow the maimed to have "the feeling of touching things," says Paolo Dario, the project‘s coordinator at the Polo Sant‘Anna Valdera institute in this central Italian town.
"We have a network, we know how to work together. We are ready to make a leap ahead," he said.
Financed with $1.8 million from a special European Union fund for emerging technologies, Cyberhand was cited as a success by European Commission officials in October when they appealed to governments and industry to give robotics more financial backing.
Increased funding is essential, they said, if Europe is to exploit robotics‘ vast economic potential and compete with projects in the United States, Japan, and South Korea .
In Dario‘s view, Europe‘s strength in robotics is in a broad approach that is also perhaps more sensitive to the social and ethical issues raised by the increasing use of robots to help humans with everyday tasks.
"They‘ve been pioneers in launching those considerations: what is an acceptable practice for robots, how do we make robots safe, are they safe, psychologically how will they influence people and their behavior?"
The Cyberhand team not only has tried to develop a hand that would provide greater grip and control for an amputee, but it also has been concerned about the hand‘s aesthetics.
Cyberhand would be attached to amputees below the elbow and covered by several layers of synthetic material that would seek to copy the features of a natural hand by making the prosthetic replacement soft, compliant, and flexible.
Though researchers in the United States have covered similar ground, they have not addressed the problems of electrodes, prosthesis, sensory feedback, control, and processing of commands all together, said Silvestro Micera, a Cyberhand researcher.
What remains to be seen, Patton says, is whether the materials used for Cyberhand will be compatible with the human body, how a patient‘s brain will adapt and how the hand can be powered.
Another project touted by European officials is HYDRA, a project coordinated from Denmark that is developing the world‘s first shape-shifting robot. It is made up of modules, each containing its own processors, batteries, sensors and actuators, which can attach and detach from each other so the robot can change its physical form.
Such a robot could be used, for example, in relief efforts after an earthquake, said Henrik Hautop Lund, a professor at the University of Southern Denmark and HYDRA‘s coordinator.
Having driven to a site, the robot could transform into a crawler to climb over debris, a snake to get through a hole, or columns to hold up a collapsed building and protect a survivor.
HYDRA has developed 100 modules, and Lund is looking for industrial partners who would invest in manufacturing the robot and put it to use. The project, begun in 2001, has received $2.1 million — about two-thirds of its total funding — from the EU.
Like Dario, Lund argues that Europe has an advantage in its more integrated approach to robotics. But he also notes the financial constraints.
Member states have failed to agree in recent months on the EU‘s 2007-2013 budget, so researchers still don‘t know how much support they will receive, sparking concern that projects could lose momentum.
"One of the problems Europe has had in its robotics research has been getting it out to market as product," said Ken Young, chairman of the British Automation and Robotics Association.
"While we may have a good research network at (the) academic level, I don‘t see the big industrial players getting involved to the extent they do in Japan and Korea. Ultimately it is these people who will take it to market and make it a success. ... In the EU it strikes me we develop some great technology and then leave it for the rest of the world to pick up and exploit."
___
On the Net:
发布于12月4日 21:53 | 评论数(0) 阅读数(1017) | 我的文章
In Pioneering Study, Monkey Think, Robot Do
In Pioneering Study, Monkey Think, Robot Do
By SANDRA BLAKESLEE
In experiments at Duke University, implants in the monkeys' brains picked up brain signals and sent them to a robotic arm, which carried out reaching and grasping movements on a computer screen driven only by the monkeys' thoughts.
The achievement is a significant advance in the continuing effort to devise thought-controlled machines that could be a great benefit for people who are paralyzed, or have lost control over their physical movements.
In previous experiments, some in the same laboratory at Duke, both humans and monkeys have had their brains wired so they could move cursors on computer screens just by thinking about it. And wired monkeys have moved robot arms by making a motion with their own arms. The new research, however, involves thought-controlled robotic action that does not depend on physical movement by the monkey and that involves the complex muscular activities of reaching and grasping.
The study is being published today in the inaugural issue of The Public Library of Science, a peer-reviewed scientific journal that makes articles available free of charge. The research team was led by Dr. Miguel A. L. Nicolelis, a neurobiology professor and co-director of the Center for Neuroengineering at Duke, in North Carolina. Dr. Nicolelis also did the earlier research on monkeys and robot arms at Duke.
While other laboratories have helped monkeys use thoughts to move robots, using different experimental designs, the Duke findings go furthest in the sense that their robots were mentally assimilated into the animals' brains.
"For nearly completely paralyzed people, this promises to be a fantastic boon," said Dr. Jon Kaas, a psychology professor at Vanderbilt University in Nashville, who is familiar with Dr. Nicolelis's research. "A person could control a computer or robot to do anything in real time, as fast as they can think."
While experts agree that thought-controlled personal robots are many years off, the Duke University team recently showed that humans produce brain signals like those of the experimental monkeys.
"Monkeys not only use their brain activity to control a robot," said Dr. John Chapin, a professor of physiology and pharmacology at the State University of New York Downstate Medical Center in Brooklyn. "They improve their performance with time. The stunning thing is that we can now see how this occurs, how neurons change their tuning as the monkey does different tasks."
Dr. Nicolelis implanted tiny probes called microwires into several brain regions of two rhesus monkeys. At first, each monkey learned to move a joystick that controlled a cursor on a computer screen. When a ball appeared, the animal had to move the cursor to the target to earn a drink of juice. Researchers collected electrical patterns from the monkey's brain as it performed the tasks.
After the monkey became skilled at the exercise, the scientists disconnected the joystick. At first, the monkey jiggled the stick and stared at the screen, Dr. Nicolelis said. Even though the joystick was not working, the monkey's reaching and grasping motor plans were being sent to a computer, which translated those signals into movements on screen.
There was an "incredible moment" when the monkey realized that it could guide the cursor and grasp an object on the screen just by thinking it, Dr. Nicolelis said. The arm dropped. Muscles no longer contracted.
The final step was to divert brain signals to a computer model that controlled the movements of a robot. The monkey continued to think the movements but in doing so it now moved the robot arm directly, without a joystick, which in turn directed movements of the cursor.
Controlling a shaky, jerky robot with thought is not easy, Dr. Nicolelis said. When the robot is first added, the monkey's performance degrades. It takes two days for the animal to learn the mechanical properties of the arm and to incorporate its delays into motor planning areas.
"By the end of training, I would say that these monkeys sensed they were reaching and grasping with their own arms instead of the robot arm," Dr. Nicolelis said. "Every time we use a tool to interact with our environment, such as a computer mouse, car or glasses, our brain assimilates properties of the tool into neuronal space. Tools are appendages which are incorporated into our body schema. As we develop new tools, we reshape our brains," he said.
http://news.mc.duke.edu/filebank/2003/10/41/Robot_arm.swf
发布于12月2日 0:00 | 评论数(0) 阅读数(904) | 我的文章
总共1页 第页
发消息
加为好友
