Natural Language Processing in Biomedicine

188.1

Introduction

188.2

Linguistic Principles

188.3

Applications in Biomedicine

Speech Systems • Lexical Systems • Syntactic and

Stephen B. Johnson

Semantic Systems • Discourse Systems

Columbia University

188.4

Challenges

Introduction

Natural language is the primary means of communication in all complex social interactions. In biomedical areas, knowledge and data are disseminated in written form, through articles in the scientific literature, technical and administrative reports, and hospital charts used in patient care. Much vital information is exchanged verbally, in interactions among scientists, clinical consultations, lectures, and in conference presentations. Increasingly, computers are being employed to facilitate the process of collecting, storing, and distributing biomedical information. Textual data is now widely available in an electronic format, through the use of transcription services, and word processing. Important examples include articles published in the medical literature and reports dictated during the process of patient care (e. g., radiology reports and discharge summaries).

While the ability to access and review narrative data is highly beneficial to researchers, clinicians, and administrators, the information is not in a form amenable to further computer processing, for example, storage in a database to enable subsequent retrievals. At present, the most significant impact of the computer in medicine is seen in processing structured data, information represented in a regular, pre­dictable form. This information is often numeric in nature, e. g., measurements recorded in a scientific study, or made up of discrete data elements, e. g., elements selected from a predefined list of diseases.

The techniques of natural language processing provide a means to bridge the gap between textual and structured data, allowing humans to interact using familiar natural language, while enabling computer applications to process data effectively.

Linguistic Principles

Natural language processing (or computational linguistics) is a branch of computer science concerned with the relationship between information as expressed in natural language (sound or text) and repre­sented in formalisms that facilitate computer processing. Natural language analysis studies how to convert natural language input into a structured form, while natural language generation investigates how to produce natural language output from structured representations. Natural language processing investi­gates and applies scientific principles of linguistics through development of computer systems. One of the most important principles is that natural language is built up from several layers of structure, each layer defined as a set of restrictions on the previous layer [Harris 1991]:

TABLE 188.1 Layers of Linguistic Structure

Linguistic Layer

Description

Phonological

Mapping of sounds to phonemes (letters)

Morphological

Grouping of phonemes into morphemes (roots and inflections)

Lexical

Combining of morphemes into words

Syntactic

Combining of words into sentence structure

Semantic

Mapping of sentence structure into literal meanings

Pragmatic

Combining of sentence meanings into discourse meaning

These layers correspond roughly to the subfields of linguistics, and to the areas of study in natural language processing, which seeks to develop them into computer models. Knowledge about the rules of language structure at any or all of its levels is called competence. Most natural language processing applications in biomedicine do not even begin to approach the language competence of humans. For example, a transcription system may have knowledge about the sounds of a language but know nothing about the syntax of sentences. Similarly, a program that indexes scientific articles may know which terms to look for in the text, but have no ability to express how these terms are related to one another. However, it is important to emphasize that such applications may perform extremely useful tasks, even with very limited competence.

Many natural language processing applications exploit the fact that biomedical fields are restricted semantic domains, which means that natural language information associated with that field is focused on a narrow range of topics. For example, an article about cell biology is limited to discussions of cells and tissues, and is unlikely to mention political or literary issues. The natural language of a restricted semantic domain is called a sublanguage. Sublanguages tend to have specialized vocabularies and spe­cialized ways of structuring sentences (e. g., the “telegraphic style” of notes written about patients in hospital charts) and ways of organizing larger units of discourse (e. g., the format of a technical report) [Grishman and Kittredge 1986, Kittredge and Lehrberger 1982].

These properties of sublanguages allow the use of methods of analysis and processing that would not be possible when processing the language of newspaper articles or novels. For example, a program that indexes medical articles can select index terms from a list of terminology known to be of interest to researchers; a speech recognition system can exploit the fact that only certain words can be uttered by a user in response to a given prompt; a system that analyzes clinical reports can look for predictable semantic patterns that are characteristic of the given domain.

Applications in Biomedicine

There are four forms of natural language exchange possible between human and the computer: (1) the human can supply information to the computer using natural language; (2) the human can retrieve information from the computer through natural language; (3) the computer can supply information to the human in the form of natural language output; and (4) the computer can request information from the human using natural language questions. Applications of natural language processing may support one or several of these modes of exchange. For example, an automated questionnaire for patients may generate multiple choice questions but be able to receive only numbered answers, while a database interface may be able to accept natural language questions submitted by a researcher and return natural language replies. Applications can be classified according to the levels of language competence embodied in their design: (1) speech systems, (2) lexical systems, (3) syntactic and semantic systems, and (4) discourse systems.

Speech Systems

Speech recognition systems [Grasso 1995] process voice data into phonemic representations (such as text), while speech synthesis systems generate spoken output from such representations. Speech process­ing applications may work only at this level of language competence, e. g., software that transcribes spoken input into text, or software that generates speech when visual information cannot be employed, such as on the telephone. Applications in this category include systems that capture structured endoscopic reports [Johannes and Carr-Locke 1992], and emergency room notes [Linn et al., 1992]. Applications may also employ speech technology as part of a larger system. Examples include interfaces to diagnostic systems [Landau et al., 1989], [Shiftman et al., 1992], and systems for taking patient histories [Johnson et al., 1992].

Lexical Systems

Lexical systems work with language at the level of words and terms (short sequences of words). A lexicon is a special type of database that provides information about words and terms which may include pronunciation, morphology (roots and affixes), and syntactic function (noun, verb, etc.). The Specialist Lexicon provides information about medical terms and general English words [National Library of Medicine 1993]. A thesaurus groups synonymous terms into semantic classes, which are frequently organized into a hierarchical classification scheme. The Systematized Nomenclature of Medicine (SNOMED) classifies medical terms in several hierarchies, such as (1) topography, (2) morphology,

Etiology, (4) function, (5) disease, (6) occupation, and (7) procedure [Rothwell et al., 1993]. Medical Subject Headings (MeSH) classifies terms used in medical literature [National Library of Medicine 1990]. The Metathesaurus of the Unified Medical Language System (UMLS) combines SNOMED, MeSH, and other thesauri [Lindberg 1993].

Semantic classes are used to index natural language documents to facilitate their retrieval from a database. For example, MeSH is used to index the medical literature in the MEDLINE database [Bachrach 1978]. Semantic classes are also used to define the set of data elements (controlled vocabulary) that can be processed by a computer application. Controlled vocabularies are used by diagnostic systems and clinical information systems—by the programs that collect data, store it in databases, and display it [Linarrson and Wigertz 1989].

Syntactic and Semantic Systems

Lexical techniques can only approximate the meaning of an article or some text entered by the user of a computer application. One approach to obtaining a deeper understanding is to produce a representation of the syntactic structure of a sentence (e. g., determining the subject, the verb, and the object), and then map this structure into a representation of its meaning (such as predicate calculus). A computer program that analyzes the structure of sentences is called a parser, and uses a lexicon and a grammar (a formal representation of the syntactic rules of a language).

Representations of semantic content (or knowledge representation) vary, but most representations are variations of a structure called a frame, in which predefined slots are filled with information from natural language sentences. A number of systems process clinical reports into their own specialized frame representations, for example the Linguistic String Project [Sager et al., 1986], MedLEE [Friedman 1994], and SymText [Haug 1997]. The representation of “conceptual graphs” [Sowa 1984] has emerged as a potential standard for semantics and is used for medical language processing by a number of systems which include METEXA [Schroder 1992], MENELAS [Bouaud et al., 1997], and RECIT [Baud et al., 1992].

Information from the sentences of a discourse (e. g., an article or clinical report) combine in complex ways; for example, cross-references are made (e. g., using pronouns), and events are related temporally or causally. Understanding a discourse fully may require information about the context in which the natural language exchange occurs, or the background knowledge of a technical field. Systems based on syntax and semantics usually carry out an analysis of the discourse and context information as a subse­quent stage. Other systems attempt to approximate the semantic analysis of sentences, using medical knowledge to guide the overall processing of a text by defining the possible sequences of topics and subtopics characteristic of the domain. Domains in which these methods have been attempted include: (1) the history and physical sections of a patient chart [Archbold and Evans 1989]; (2) echocardiography reports [Canfield et al., 1989]; (3) discharge summaries [Gabrieli and Speth 1987]; and (4) radiology reports [Ranum 1988].

Natural language is also a convenient mode for the output of applications, for example in producing clinical reports [Kuzmak and Miller 1983]. Some expert systems have used natural language to express their recommendations [Rennels et al., 1989], and to provide explanations of their reasoning process [Lewin 1991].

Challenges

The design of controlled vocabularies is an important area of research [Cimino et al., 1989]. As yet, no comprehensive vocabulary for clinical medicine exists, and the effort to create one is beyond the means of any single institution. Indexing of articles and reports is currently a labor-intensive process, and quality control is a significant problem. Index terms are a poor approximation of the meaning of a text. Similarly, systems that capture clinical information through the entry of individual findings fall short of capturing the patient history. Natural language processing systems with greater understanding of syntax and seman­tics may help by providing richer information representations.

No natural language processing system currently covers the complete range of language interaction with significant competence at all levels—from speech to discourse. In the near future, the systems with the greatest practical success will specialize in performing selected language processing tasks in well — defined domains.

Defining Terms

Controlled vocabulary: A set of well-defined data elements intended for use in a computer application.

Frame representation: A structure for representing complex information, consisting of fixed, named

“slots” which may contain data values or frames.

Grammar: A formal representation of language structure, usually syntactic structure.

Language competence: Knowledge about the rules of a language.

Lexicon: A formal compilation of the words of a language, with information about phonetics, mor­

Phology, syntax, and semantics (depending on requirements of computer applications). Morphology: Field studying how sounds combine into morphemes, and how morphemes combine to

Form words.

Parser: A computer program that uses a grammar and lexicon to analyze sentences of a language,

Producing a syntactic or semantic representation.

Phonology: Field studying the rules that govern the sounds used in languages.

Pragmatics: Field studying how sentences are combined to form larger units of discourse, and how to

Represent meaning in the context of a situation.

Semantics: Field studying the relation between natural language sentences and formal representations

Of meaning.

Structured data: Data that has a explicit, unambiguous, and regular structure, making it amenable to

Computer processing.

Sublanguage: The natural language of a restricted semantic domain that is focused on a specific set of

Tasks or purposes, e. g., interpreting X-ray images.

Syntax: Field studying how words combine to form sentences.

References

Allen, J. 1995. Natural Language Understanding, Second Edition. Redwood City, CA: The Benjamin Cummings Publishing Company, Inc.

Archbold, A. and Evans, D. 1989. On the Topical Structure of Medical Charts. In: Proceedings of the 13th Annual SCAMC, p. 543-547. IEEE Computer Society Press, Washington, D. C.

Bachrach, C. A. and Chaen, T. 1978. Selection of MEDLINE contents, the development of its thesaurus, and the indexing process, Medical Informatics. 3(3): 237-254.

Baud, R. H., Rassinoux, A. M., and Scherrer, J. R. 1992. Natural Language Processing and semantical representation of medical texts. Meth. Inform. Med. 31(2):117—125.

Bouaud, J., Zweigenbaum, P., et al., 1997. A Semantic Composition Method Driven by Domain Knowledge Models. Twenty-First Annual Symposium of the American Medical Informatics Association.

Canfield, K., Bray, B., Huff, S., and Warner, H. 1989. Database capture of natural language echocardio­graphy reports: A Unified Medical Language System approach. In: Proceedings of the 13th Annual SCAMC, p. 559-563. IEEE Computer Society Press, Washington, D. C.

Cimino, J. J., Hripcsak, G., Johnson, S. B., Friedman, C., Fink, D. J., and Clayton, P. D. 1989. Designing an Introspective, Multi-Purpose Controlled Medical Vocabulary. Proceedings of the 13th Annual SCAMC, p. 513—518. IEEE Computer Society Press, Washington, D. C.

Friedman, C., Alderson, P. O., Austin, H. M., Cimino, J. J., and Johnson, S. B. 1994. A general natural language text processor for clinical radiology. JAMIA 1(2): 161—174.

Gabrieli, E. and Speth, D. 1987. Computer Processing of Discharge Summaries. In: Proceedings of the 11th Annual SCAMC, p. 137—140. IEEE Computer Society Press, Washington, D. C.

Haug, P. J. et al., 1994. A Natural Language Understanding System Combining Syntactic and Semantic Techniques. Eighteenth Annual Symposium of the American Medical Informatics Association.

Grasso, M. A. Automated speech recognition in medical applications. MD Computing, 1995;12(1):16—23.

Grishman, R. and Kittredge, R., Eds. 1986. Analyzing Language in Restricted Domains: Sublanguage Description and Processing. Erlbaum Associates, Hillsdale, New Jersey.

Harris, Z. 1991. A Theory of Language and Information—A Mathematic Approach. Clarendon Press, Oxford.

Johannes, R. S. and Carr-Locke, D. L. 1992. The Role of Automated Speech Recognition in Endoscopic Data Collection. Endoscopy. 24(Suppl 2): 493—498.

Johnson, K., Poon, A., Shiffman, S., Lin, R., and Fagan, L. 1992. A history taking system that uses continuous speech recognition. In: Proceedings of the 16th Annual SCAMC, p. 757—761. IEEE Computer Society Press, Washington, D. C.

Kittredge, R. and Lehrberger, J., Eds. 1982. Sublanguage—Studies of Language in Restricted Semantic Domains. De Gruyter, New York.

Kuzmak, P. M. and Miller, R. A. 1983. Computer-aided generation of result text for clinical laboratory tests. In: Proceedings of the 7th Annual SCAMC, p. 275—278. IEEE Computer Society Press, Wash­ington, D. C.

Landau, J. A., Norwich, K. N., and Evans, S. J. 1989. Automatic Speech Recognition—Can it Improve the Man-Machine Interface in Medical Expert Systems? Int. J. Biomed. Comp.. 24(2): 111—117.

Lewin, H. C. 1991. HF-Explain: a natural language generation system for explaining a medical expert system. In: Proceedings of The 15th Annual SCAMC, p. 644—648. IEEE Computer Society Press, Washington, D. C.

Lindberg, D. A.B., Humphreys, B. L., and McCray, A. T. 1993. The Unified Medical Language System. In: Yearbook of Medical Informatics, van Bemmel, J. H., McCray, A. T., Eds., p. 41-51. International Medical Informatics Association, Amsterdam.

Linarrson, R. and Wigertz, O. 1989. The data dictionary—a controlled vocabulary for integrating clinical databases and medical knowledge bases. Meth Inform. Med. 28(2): 78-85.

Linn, N. A., Rubenstein, R. M., Bowler A. E., and Dixon, J. L. 1992. Improving the Quality of Emergency Room Documentation Using the Voice-Activated Word Processor: Interim Results. In: Proceedings of the 16th annual SCAMC, p. 772-776. McGraw Hill, New York.

National Library of Medicine. 1990. Medical Subject Headings (NTIS NLM-MED-90-01). National Library of Medicine. Bethesda, Maryland.

National Library of Medicine. 1993. The Specialist Lexicon. Natural language systems group, National Library of Medicine, Bethesda, MD.

Ranum, D. 1988. Knowledge Based Understanding of Radiology Text. In: Proceedings of the 12th Annual SCAMC, p. 141-145. IEEE Computer Society Press, Washington, D. C.

Rennels, G., Shortliffe, E., Stockdale, F., and Miller, P. 1989. A computational model of reasoning from the clinical literature. AI Magazine. 10(1): 49-57.

Rothwell, D. J., Palotay, J. L., Beckett, R. S., and Brochu, L., Eds. 1993. The Systematized Nomenclature of Medicine. SNOMED International. College of American Pathologists, Northfield, Illinois.

Sager, N., Friedman, C., and Lyman, M. 1987. Medical Language Processing—Computer Management of Narrative Data. Addison-Wesley, Reading, Mass.

Scherrer, J. R., Cote, R. A., and Mandil, S. H. 1989. Computerized natural language medical processing for knowledge representation. North Holland, Amsterdam.

Schroder, M. 1992. Knowledge-Based Processing of Medical Language: A Language Engineering Approach. In: Advances in Artificial Intelligence, 16th German conference on AI, p. 221-234. Springer Verlag, Berlin.

Shiffman, S., Lane, C. D., Johnson, K. B., and Fagan, L. M. 1992. The integration of a continuous speech recognition system with the QMR diagnostic program. In: Proceedings of the 16th Annual SCAMC, p. 767-771. IEEE Computer Society Press, Washington, D. C.

Spyns, P. 1996. Natural Language Processing in Medicine. Meth. Inform. Med. 35:285-301.

Sowa, J. F. 1984. Conceptual Graphs: Information Processing in Mind and Machine. Addison-Wesley, Read­ing MA.

Van Bemmel, J. H. (Ed.). Meth. Inform. Med., 1998:4(5).

Further Information

For general information about the field of natural language processing, see [Allen 1987], and [Covington

1994]. Surveys of research in sublanguage can be found in [Kittredge and Lehrberger 1982] and [Grish —

Man and Kittredge 1986]. A variety of papers on medical language processing are collected in [Scherrer

Et al., 1989 and Van Bemmel 1998].

Geddes, L. A. “Historical Perspectives 5 — Electroencephalography ” The Biomedical Engineering Handbook: Second Edition.

Ed. Joseph D. Bronzino

Boca Raton: CRC Press LLC, 2000

Historical Perspectives 5

Electroencephalography

Leslie A. Geddes Historical Background

Purdue University Commercial Production of EEG Machines

Historical Background

Hans Berger (1929) was the first to record electroencephalograms from human subjects. However, before then, it was well known that the brain produced electrical signals. In fact, in Berger’s first paper there is a short history of prior studies in animals. Interestingly, the first person to demonstrate the electrical activity of the brain did not make recordings. In 1875, Richard Caton in the United Kingdom used the Thomson (Kelvin) sensitive and rapidly responding reflecting telegraphic galvanometer to display the electrical activity of exposed rabbit and monkey brains. His report [Caton, 1875], which appeared in the British Medical Association Journal, occupied only 21 lines of a half-page column. In part, the report stated:

In every brain hitherto examined, the galvanometer has indicated the existence of electric currents. The external surface of the grey matter is usually positive in relation to the surface of a section through it. Feeble currents of varying direction pass through the multiplier [galvanometer] when the electrodes are placed on two points of the external surface, or one electrode on the grey matter, and one on the surface of the skull. The electric currents of the grey matter appear to have a relation to its function. When any part of the grey matter is in a state of functional activity, its electric current usually exhibits negative variation. For example, on the areas shown by Dr. Ferrier to be related to rotation of the head and to mastication, negative variation of the current was observed to occur whenever those two acts respectively were performed. Impressions through the senses were found to influence the currents of certain areas; e. g., the currents of that part of the rabbit’s brain which Dr. Ferrier has shown to be related to movements of the eyelids, were found to be markedly influenced by stimulation of the opposite retina by light.

No recordings of the movement of the spot of light on the scale of the Kelvin galvanometer have been found, perhaps because, at that time, telegraphic operators used to read the dots and dashes of the Morse code by watching the movements of the spot of light on the galvanometer scale. Nonetheless, Caton’s description clearly shows than he witnessed the fluctuating potentials that we now know exist. Also important is the fact that Caton was the first to report visual-evoked potentials.

Berger, a psychiatrist in Jena, Germany, was aware of the several prior electroencephalographic animal studies and had conducted experiments using dogs. The only recording devices available to him were the string galvanometer, developed by Einthoven [1903] for electrocardiography and the capillary elec­trometer developed by Marey [1876]. Although there were a few mirror-type oscillographs available for recording waveforms from alternating current (50 to 60 Hz) generators and transformers, the sensitivity of such devices was very low, and they could not be used for bioelectric recording without a vacuum — tube amplifier.

The voltage appearing on the scalp produced by the brain is only about one-tenth that of the ECG detected with limb leads. To enable recording brain activity with the string galvanometer, the tension in the string was reduced, which increased the sensitivity but reduced the speed of response. This was the method used by Berger when he found that the capillary electrometer was unsatisfactory.

Natural Language Processing in Biomedicine

• I I I i ;

« * ‘W—* r^,^. ,

I > I I i i

FIGURE HP5.1 (a) The Edelmann double-string galvanometer of the type used by Berger. On the left it is a funnel

For detecting the arterial or venous pulse and on the right is a device for detecting the heart sounds, all of which can be recorded along with the two channels of ECG, and a time signal on 12-cm-wide photographic paper. (b) The phonocardiagram, venous-pulse record, and two recordings of the ECG. (From Zusatz apparate zum Elektrokar- diographen, Siemens-Reiniger-Veifa Berlin ca. 1926.)

Apart from problems of the lack of sensitivity and speed of response in the recording apparatus, Berger faced severe electrode problems. Because of the low amplitude of the cortical signals, the electrodes had to be very stable, producing no voltages that could be seen on the electrocortical recordings. In other words, the electrode noise and stability had to be in the low-microvolt range. Zinc electrodes were popular at that time, and Berger used them in his first dog studies just after the turn of this century. For human use, he used zinc-plated needles (insulated down to the tip) and inserted them through existing trephine holes of skull defects so that the tip was epidural. The electrodes were sterilized with 10% formalin solution. Berger first used the single-string Edelmann galvanometer designed for electrocardiography. Later he used a two-string unit of the type shown in Fig. HP5.1 sO that he could record the ECG along with the EEG. Such units also were equipped with devices to record the venous or arterial pulses and heart sounds. The photographic recording paper was 12 cm wide and 50 m in length.

The string galvanometer is a current-drawing device; therefore, a low electrode-subject resistance was necessary for adequate sensitivity. Berger stated that it was difficult to obtain a low resistance and reported a value of 1600 Q for his needle electrodes when placed 5 to 6 cm apart with the tips in the epidural space. A high-resistance electrode pair would reduce the amplitude of the recorded activity. Later Berger used chlorided silver needle electrodes to help solve this problem.

Difficulties with the zinc-plated needle electrodes led Berger to develop very thin lead-foil electrodes, wrapped in flannel, and soaked in 20% NaCl solution. The combination of saline-soaked flannel and the use of a rubber bandage to hold the electrodes on the scalp permitted recording the EEG for hours without the saline evaporating. The resistance measured between a pair of such electrodes ranged from 500 to 7600 Q depending on the size of the electrodes.

Berger’s final improvement to his recording equipment consisted of placing a capacitor in series with the electrodes and string galvanometer to block the steady potential difference due to slight electrochem­ical differences in the two lead electrodes. Recall that the amplitudes being recorded were in the range of tens of microvolts, and any steady difference in electrode potential would cause a steady deflection of the baseline of the recording. With the capacitor, this steady offset potential did not deflect the galva­nometer baseline.

Describing his first studies with the zinc-plated needle electrodes, Berger stated [translation by Gloor, 1969]:

As a general result of these recordings with epidural needle electrodes I would consequently like to state that it is possible to record continuous current oscillations, among which two kinds of waves can be distinguished, one with an average duration of 90 ct, the other with one of 35 ct. The longer waves of 90 ct are the ones of larger amplitude, the shorter, 35 ct waves are of smaller amplitude. According to my observations there are 10 to 11 of the larger waves in one second, of the smaller ones, 20 to 30. The magnitude of the deflections of the larger 90 ct waves can be calculated to be about 0.0007 to

00015 V, that of the smaller 35 ct waves, 0.00002 to 0.00003 V. [The symbol ct was used for millisec­onds.]

In his first paper, Berger called the dominant low-frequency first-order and the higher-frequency waves second order. In his second paper he stated:

For the sake of brevity I shall subsequently designate the waves of first order as alpha waves = a — w, the waves of second order as beta waves = P — w, just as I shall use “E. E. G.” as the abbreviation for the electroencephalogram and “E. C. G.” for the electrocardiogram.

The average duration of the waves reproduced in [his] Figure 1 is for the a — w = 120 ct and for the P — w = 30 to 40 ct.

Berger objected to the term electrocerebrogram to designate a record of the electrical activity of the brain. He stated:

Because for linguistic reasons, I hold the word “electrocerebrogram” to be a barbarism, compounded as it is of Greek and Latin components, I would like to propose, in analogy to the name “electrocar­diogram” the name “electroencephalogram” for which here for the first time was demonstrated by me in man.

Berger’s papers contain many EEGs from patients, but those from his son Klaus are discussed fre­quently. In fact, Klaus was used as a subject for electrode testing. For example, Berger wrote:

Klaus’ records were taken with every other possible type of electrodes: silver, platinum, lead electrodes, etc; also, different arrangements of these on the skin surface of the head were used. However, time and again it was found that the best arrangement was that with electrodes placed on the forehead and occiput. Of Klaus’ many records, I only want to show another small segment of a curve obtained in this manner [Fig. HP5.2]. In this instance head-band electrodes were applied to the forehead and

Y-(^1^//>/V//V//VNAA/VVVA^VVVVVVAAA/VVV^/vVVVVVVAAA/VVVSA/VVVVVVV^

FIGURE HP5.2 Klaus at the age of 15. Double-coil galvanometer. Condenser inserted. Recording from forehead and occiput with head-band electrodes. (Top) The record obtained from the scalp; (bottom) time in 1/10 second.

Occiput and were fixed with rubber bandages. From these head-band electrodes, records were taken with galvanometer 1 of the double-coil galvanometer; galvanometer 2 was set at its maximum sensi­tivity and was used as a control to make sure that no outside currents were entering the galvanometer circuit to disturb the examination. At that time I was still very distrustful of the findings I obtained and time and again I applied such precautionary measures. The record of galvanometer 2 ran as completely straight line, without my oscillation.

From the lengthy discussions in Berger’s papers, it is easy to see that few believed that his recordings originated in the brain. To dispel some of the uncertainty, Berger usually recorded the ECG along with the EEG. Occasionally, he recorded heart sounds and the arterial pulse. In addition in the electrical activity of the heart, various critics proposed that Berger’s recorded activity was due to friction of blood in the cerebral arteries, pulsations of the brain and/or scalp, respiration, contraction of piloerector or skeletal muscle, and glandular activity. Berger dealt with all the potential artifacts, pointing out that their time course and frequency were different from those of the EEG and showed that the EEG continued with transient slowing of the heart rate. Finally, he stated (perhaps in exasperation):

I therefore believe I have discussed all the principal arguments against the cerebral origin of the curves reported here which in all their details have time and again preoccupied me, and in doing so I have laid to rest my own numerous misgivings.

Among the first to publish an English-language verification of Berger’s observations were Jasper and Carmichael [1935], then at Brown University in the United States. Silver electrodes (1 to 2 cm in diameter), covered with flannel and soaked in saline, were connected to an amplifier/mirror-oscillograph system. They were able to confirm Berger’s findings and extend them, showing that with a two-channel system used to record the EEG of a girl with a convulsive disorder, the alpha wave frequency was 10 per second on the left side of the head and 6 to 8 per second on the right side, one of the early indications that the EEG was altered by brain pathology. Figure HP5.3a sHows one of the records obtained by Jasper and Carmichael.

Jasper later came to the Montreal Neurological Institute (McGill University) and created the Electro — pysiological Laboratory, which was formally opened with a celebration meeting held on February 24-26, 1939. In attendance were the world leaders in electrophysiology.

Clinical EEG at the Montreal Neurological Institute was inaugurated using a machine built by Andrew Cipriani. The inkwriter was of unique design and featured a strong magnetic field produced by an electromagnet. In this field was a circular coil coupled to an inkwriting pen. It is interesting to observe that this principle had been used by d’Arsonval [1891] with pneumatic coupling to a tambour that caused a pen to write on a rotating drum, as shown in Fig. HP5.4a. LAter in the 1930s, this coil design was coupled to a conical diaphragm and became the first dynamic loudspeaker. The pen motor devised by Cipriani is sketched in Fig. HP5.4b. A small vane, affixed to the writing stylus, dipped into an oil chamber (dashpot) to provided damping. This four-channel instrument was in routine use when the author first came in contact with Jasper in the early 1940s; it was replaced in 1946 by a six channel model III Grass EEG.

Meanwhile, at Harvard University (Boston, Mass), Gibbs, Davis, and Lennox [1935], were pursuing their interest in epilepsy. Recognizing the potential of the EEG for the diagnosis of epilepsy, they initiated a series of studies that would occupy the next many decades. Their recorder consisted of an ink-writing telegraphic recorder called the Undulator. In December of 1935, they published their first paper on EEG which carried a footnote that read: “This paper is no, XVII in a series entitled Studies in Epilepsy.” Citing the Berger papers and that by Jasper and Carmichael, Gibbs et al. stressed the importance of direct-inking pens for immediate viewing of the EEG so that the effect of environmental factors could be identified immediately. They reported:

The method is exceedingly simple. Electrical contact is made to two points on the subject’s head. Except for the study of grand mal epileptic seizures we regularly employ as electrodes two hypodermic

Natural Language Processing in Biomedicine

FIGURE HP5.3 The first U. S. records of the EEG to confirm Berger’s report. (a) The records obtained by Jasper and Carmichael (1935). The first channel shows the alpha waves, and the second shows the electrical activity detected by electrodes on the leg above the knee. The second record shows alpha inhibition by illumination of the retina, and the third record shows return of the alpha waves when the light was extinguished. (b) The first records published by Gibbs et al. (1935) showing alterations in normal subjects by various types of sensory stimulation eyes open and closed, problem solving, noise (rattle), and smelling ether. (Both by permission.)

Needles inserted one into the scalp at the vertex of the skull and the other into the lobe of the left ear.

Enough procaine hydrochloride is injected previously to insure the continued comfort of the subject.

Figure 3b iS a reproduction of the first EEG obtained by Gibbs et al. Soon Gibbs produced his well — known Atlas of Electroencephalography, which first appeared in 1941 and became the “bible” for training electroencephalographers.

The first recording equipment used by Gibbs et al. [1935] was built by Lovett Garceau. It consisted of a four-stage, singled-sided, resistance-capacity-coupled amplifier made with high-grain, screen-grid tubes driving a distinct-linking telegraphic recorder called the Undulator (U in Fig. 5) obtained from the Western Union Telegraph Company, Fig. 5 shows the circuit diagram. the overall high-frequency response extended to 25 Hz [Grass, 1984].

Slightly earlier, in Germany, Toennies [1932] had developed a direct-linking recorder that he called the Neurograph. Fig. HP5.6 shows a picture of the instrument and a record of the human electrocardio­gram and the response of canine eyes to light. Note that the chart is running in the opposite direction to conventional recordings. The time marks are 1/5 second.

Natural Language Processing in Biomedicine

FIGURE HP5.4 Electropenumatic inkwriter, described by d’Arsonval (1891) (a) and sketch of moving-coil inkwriter devised by Cipriani in the late 1930s for EEG at the Montreal Neurological Institute (b).

Natural Language Processing in Biomedicine

FIGURE HP5.5 The single-sided, five-stage, resistance-capacitor-coupled amplifier developed by Gareau to drive the Western Union inkwriting telegraphic recorder (Undulator) used by Gibbs et al. to record EEGs (From Garceau et al. 1935. Arch Neurol Psychiatry. With permission.)

Natural Language Processing in Biomedicine

FIGURE HP5.6 Toennies Neurograph and recordings showing a 1-mV calibration, the human ECG, and electrical activity from the canine eye, time marks 1/5 s. (From Toennies [1932].)

Commercial Production of EEG Machines

The excellent collaboration between Gibbs and Grass in 1935 resulted in replacement of the Undulator telegraphic inkwriter with a robust, d’Arsonval-type inkwriter in early 1936. Meanwhile, the Grass Instrument Company had been founded (1935) to produce electrophysiologic equipment at the well — known address, 101 Old Colony Avenue, Quincy, Mass; the address in the same today.

In 1937, Grass adopted folding chart paper and by 1939 was providing three-, four-, and six-channel EEGs. The Model III shown in Fig. HP5.7 iS the machine recognized by all and founded many EEG laboratories. It featured a knee-hole console with the pens at the right and ample viewing space for the record as it evolved. The chart speed was 30 mm/s, which ultimately became the standard.

At about the same time Grass was building EEG machines in Quincy, Mass., Franklin Offner, a research associate of Ralph Gerard at the University of Chicago, started his own company at 5320 North Kedzie Avenue, Chicago, to produce an EEG using the Crystograph, a high-speed, piezoelectric inkwriter that was described by Offner and Gerard [1936]. It consisted of two slabs of rochelle salt crystals, three corners of each were clamped, and the fourth was free to move when a voltage was applied to electrodes on the crystals. The moving corners were mechanically linked by a slender brass belt that caused motion of the rod that carried the inkwriting stylus. The sinusoidal frequency response extended uniformly to 100 Hz; the chart speed 25 mm/s, the same as for ECG. However, a gear shift provided chart speeds above and below 25 mm/s.

The Crystograph was used with the first Offner EEG machines, and it was ideally suited for high efficiency energy transfer from vacuum-tube amplifiers because of its high impedance. The differential amplifiers were housed in two 19-in relay tracks, as shown in Fig. HP5.8a; The Crystograph rested on an adjacent table; a six-channel Crystograph is illustrated. Damping was adjusted by a series variable resistor to achieve an excellent response to a step function. The record shown in Fig. HP5.8a wAs produced by a 3-p. V (peak-peak) square wave, showing the excellent transient response and the remarkably low internal noise level of the amplifiers.

Natural Language Processing in Biomedicine

FIGURE HP5.7 The Grass Model III EEG, circa late 1940s. (Courtesy Grass Instrument Co., Quincy, Mass.)

Although the Cyrstograph EEG produced elegant records, temperature and humidity played havoc with the rochelle salt crystals, and Offner replaced it with the Dynagraph, shown in Fig. HP5.8b, The recorder being a low-impedance, robust d’Arsonval-type moving-coil inkwriter.

Efficient coupling between the output amplifier stage and the moving-coil inkwriter was always a problem that was solved by Offner in his Dynagraph. A synchronous vibrator sampled the signal from the preamplifier, amplified the pulses, and recombined them after passing through a step-down trans­former, thereby providing an efficient impedance match between the amplifier output stage and penwriter. Fig. HP5.8b iS an illustration of the Offner Dynagraph that featured a sit-down console for viewing the record. The antiblocking feature of this design is shown at the bottom of Fig. HP5.8b By a sine wave record in which a transient, 100 times larger than the recording, was presented and the recording was restored a fraction of a second after the transient. Offner later replaced the vacuum tubes with transistors, bringing out the first transistorized EEG.

In the late 1940s, Warren Gilson, who was a pioneer in devising physiologic instruments, started production of EEGs in Madison, Wisc.; Fig. HP5.9 iS a photograph of one of his later instruments (circa 1959).

It is noteworthy that following the Grass Model III EEG, all subsequent instruments were of the console type with the penwriter at the right end of the desktop of the console. Not easily identified in any of the foregoing figures are the two rotary switches for each channel. Each switch could connect each side of the differential amplifier input to any of the 21 electrodes of the 10-20 system. Switching also was provided to apply a step-function calibrating signal to all channels simultaneously, the step function typically being produced by depressing a pushbutton or rotating a knob. In addition, a low-current ohmmeter was provided to permit measurement of the resistance of any pair of electrodes. Some EEGs included a high — frequency filter for each channel to exclude muscle artifacts often seen when patients clenched their jaws.

Electroencephalographs have changed little since their development in the post-World War II days. Transistors have replaced the vacuum tubes, but the chart speed and frequency response are the same as those established by the first manufacturers of EEG machines, although many new recording techniques have been introduced.

Natural Language Processing in Biomedicine

FIGURE HP5.8 The Offner EEG which used the Crystograph recorder (a) and the Offner Dynagraph EEG (b), circa late 1940s (Courtesy of Franklin Offner.)

Natural Language Processing in Biomedicine

FIGURE HP5.9 The Gilson EEG, circa 1950 (Courtesy of Warren Gilson).

References

Caton R. 1875. The electric currents of the brain. Br Med J 12:278.

D’Arsonval A. 1891. Galvanograph et machines producant des courants sinusoidaux. Mem Soc Biol 3 (new series):530.

Einthoven W. 1903. Ein neues Galvanometer. Ann Phys 12 (suppl 4):1059.

Gibbs FA, Davis H, Lennox WG. 1935. The electroencephalogram in epilepsy as in conditions of impaired consciousness. Arch Neurol Psychiatry 34(6):1135.

Gloor P. 1969. Hans Berger on the electroencephalogram. EEG Clin Neurophysiol (suppl 28):350.

Grass AM. 1984. The Electroencephalographic Heritage. Quincy, Mass, Grass Instrument Co.

Jasper HH, Carmichael L. 1935. Special article: Electrical potentials from the intact human brain. Science 81:51.

Marey EJ. 1876. Des variations electriques des muscles du coeur en particulier etudiee au moyen de l’ectometre de M. Lippman, C R Acad Sci 82:975.

Offner F, Gerard RW. 1936. A high-speed crystal inkwriter. Science 84:209.

Toennies JF. 1932. Der Neurograph. Die Naturwiss 27(5):381.

Saha, S., Bronzino, J. D. “Ethical Issues Associated with the Use of Medical Technology.” The Biomedical Engineering Handbook: Second Edition.

Ed. Joseph D. Bronzino

Boca Raton: CRC Press LLC, 2000

Добавить комментарий

Ваш e-mail не будет опубликован. Обязательные поля помечены *