Vol. 19, No. 2 Printed in U.S.A. PSrCHOPHYSIOLOGY Published 1982 byThe Society for Psythophysiological Research. Inc. Modification of the EEG Time Constant by Digital Filtering THEO GASSER, ALOIS KNEIP, AND ROLE VERLEGER Zentralinsiitut fur SeelLsche Gesundheil. Mannheim. Federal Republic of Germany ABSTRACT A method is proposed for modifying the time constant of an EEG amplifier after the recording. It is based on digital filtering and works for single sweeps, averaged potentials, and other neurophysiological data, e.g. spontaneous EEG. The validation is based on calibration signals and data from a study on eventrelated potentials. DESCRIPTORS: Time constant. Slow potentials, Digital filtering. We present here a computer algorithm for the modification of the time constant of an EEG amplifier to any desired vaiue. This possibility is of interest in particular for researeh on slow potentials and to a somewhat lesser degree for the spontaneous EEG: as has been pointed out by DuncanJohnson and Donchin (1979), there is a great variability in the time constant used which may lead to spurious differences between studies. Furthermore, EEG amplifiers in clinical laboratories often have time constants ranging from 0.1 to 1.0 see only, preventing those laboratories from doing research on slow potentials, where much higher time constants are needed. Especially with behaviorally disturbed subjects it may even be advantageous to record the EEG with a time constant as low as 0.1 sec; this can alleviate problems with the dynamic range of the equipment which oceur more frequently with higher time constants. Our method applies to single sweeps, averaged potentials, and also to the continuously recorded EEG. It is based on digital filtering using the Fast FourierTransform(FFT)(Cooley&Tukey, 1965). That it works will be shown by filtering calibration signals, a single sweep, and an averaged potential. A comparison is made with a different approach by Elbert and Roekstroh (1980). Method Digital Filtering In an EEG laboratory data is recorded via an amplifier with the time constant T set at some value. In a more general and more formal way we can say that the EEG signal is passed through an analog filter which is characterized by its transfer function H^(v) {v = frequency in radians). The function H (i-) tells us by what amount an incoming sine-wave of frequency v is attenuated or amplified ("gain"), and how much it is shifted in time ("phase"). The transfer function of the analog filter of the EEG amplifier is such that the amplitude of slow components is attenuated and shifted in time, and this to an increasing degree for slower frequencies v and smaller time constants T (compare e.g. Cooper, Osselton, & Shaw, 1974). Assume that these properties are undesirable and that some "virtual" amplifier with transfer function H^ (v) is wanted instead. This is accomplished on the computer by digital filtering in the frequency domain rather than the time domain both for speed of computation and ease of programming (compare e.g. Oppenheim & Schafer, 1975). Figure 1 illustrates this modification. original omplider signal A,IU with tronster output signoi o( H] function Hj I " yirluol" nrnplifjer This work has been performed as part of the research program of the Sonderforschungsbereich 123 (project BI) and the Sonderforschungsbereich 116 (project M2}. both at the University of Heidelberg, and was made possible by financial support from the Deutsche Forschungsgemeinschaft. Address requests for reprints to: Theo Gasser, Zcntralinstitut fur Scclischc Gesundheit, Postfach 5970, 68(.X) Mannheim 1, Federal Republic of Germany. with desired Iranslef function H2 1^1 output signal • ot H2 = '-, Itl Figure I. Diagram showing the relationship of the ohtainod signal lo the calculated signal with the desired time constant. 237 0048-5772/82/020237-04$0,4(}/0 1982 Tbe Society for Psychophysiological Research, Inc. 238 Gasser, Kneip, and Verleger In order to obtain x,(t), when in fact x(t) is recorded, the following steps have to be performed: (i) Forward Fourier-transformation of data recorded to obtain the frequency domain representation: x(t)->X(v) (ii) Application of a compensating filter H ^ to obtain an estimate for the input signal X|(t) in the frequency domain: X,(v) = X(v)/H,(v) (iii) Filtering forward with a transfer function H^ of your choice: (iv) Inverse Fourier-transformation to obtain the digitally filtered data in the time domain: Whenfilteringmass data (e.g. the ongoing EEG), the data has to be segtnented because of limited computer memory: in order to avoid boundary effects, we implemented an overlapping algorithm. The programs can be obtained from the authors on request. By using the Fast Fourier Transform (FFT) (Cooley & Tukey, 1965), we assume the number of data points to be a power of two. This requirement can always be met by taking more points than absolutely necessary. This is advisable also in order to improve the frequency resolution and in order to reduce end effects due to filtering. A word of caution is appropriate at this point: It is not possible to filter any signal in an arbitrary way without loss of accuracy. Whenever the original transfer function H|(y) has attenuated the signal by a high proportion in some frequency range, its restoration by digital filtering tends to become inaccurate. It may, therefore, become illusory to filter EHG data recorded at very low time constants to very high time constants with high accuracy. Characterization ofthe EEG Amplifier In a first step a digital method has to compensate for the amplitude and phase distortion introduced by the analog devices of the EEG amplifier. The step response ( = the calibration signal) SR^ is usually modeled by an exponential (or equivalently by its transfer function H, ^ which has to be substituted for H, in relation (ii) given above): Vol. 19. No. 2 that our amplifier (Schwarzer Encephysioscript 1630) has a cascade of two such elements—one fixed at T, = 4.2 sec—which is better modeled in the following way: SR(t)= J c - ( — ' — e " ^ ' ^^—e " (3) t<0 (4) Based on non-linear regression analysis T in relation (3) could be identified and thefitwas good as judged from the residuals. {For values of T = 0.1 and 1.0 sec given by the manufacturer we determined values of 0.1084 and 1.088 sec respectively.) A FORTRAN routine described inDeuflhardand Apostolescu(Note 1) was used, but any other stable non-linear least-squares program should do as well. It is essential to subtract the average of the baseline prior to the calibration signal. As to the transfer function H, in (iii) we are free to specify it according to our goals. We usually define H, by relation (4) for desired time constants below 4.2 sec and by relation (3) for those above 4.2 sec. In many instances it may be worthwhile to take the absolute value of H, ^. thereby eliminating any phase distortion; our limited number of runs clearly demonstrated a rather gross distortion of waveforms due to the running phase for time constants as low as 0.1 sec. Tests ofthe Method Calibration Signals Calibration signals recorded frotn the EEG amplifier, with time cotistants set at 0.1 and 1.0 sec, were filtered, and the time constant of the filtered signals was estimated as described above. Table 1 demonstrates that a time constant of 1.0 sec can be reliably reduced or enhanced in a range from 0.05 to 10.0 sec. For a time constant as low as 0.1 sec, it is possible to come as far as about 7 sec with filtering which should be sufficient for most purposes. Figure 2 shows graphically the effect of filtering a calibration signal. EEG Data SR^(t) c e" t<U 0 2'rrivT-l (1) (2) where: T = time constant of EEG amplifier X.v = time respectively frequency variable c - amplitude of step response i = imaginary unit of a complex number Based on (1), we tried to estimate the time constants by regression analysis, using calibration signals as data. Due to an undershoot, not compatible with (I), thefitwas rather poor. Further information from the manufacturer showed To check the method further, it was applied to event-related potentials recorded with a time constant of 1.0 at C^, while meaningful pictures were presented to the subject (see Verleger, Gasser, Bacher, & Weingartner, 1981, for a preliminary report). We first took a single sweep: Figure 3 shows the original curve together with filtered versions with (desired) time constants 0.1, 4.0 and 10.0 sec. To bring out the effects of filtering more clearly, we computed differences in the following way {compare Figure 4): the signal filtered to 10.0 sec was considered to be the true curve; then differences were successively formed to a filtered version with 239 Modification of Time Constant March, 1982 TABLE I Time constants determined from gauge signals and filtered gauge signals Time Constant of EEG Amplifier Time Constant Desired Constant Reached O.i084 0.U5 0.3 1.0 2.0 4.0 6.0 10.0 0.048 0.2% 1.026 2.127 4.242 5.888 7.439 1.088 0.05 0.1 0.3 2.0 4.0 6.0 10.0 (1.049 Q.im 0.303 1.819 4.112 6.357 10.37(1 TIME 15EC1 Figure 3. Single sweep of event-related potential, from bottom to top; 0.l-sec filtered, l.()88-sec original. 4.()-sec tiltered, 10.0-sec filtered. TIME 3. 4. S. TIME 15EC) Figure 2. Calibration signals, from top to bottom: 1 .DHHsec original, filtered to 0.1 sec. 0.1084-sec original (for graphic representation a time scale different from the EEG trace was chosen). (SECJ Figure 4. Differences of traces 1, 2, and 3 with respect to the uppermost trace 4 (10.0-sec time constant) in Eigure 3. when using different time constants. This holds in particular for research on slow potentials of the brain. We proposed a versatile method, based on digital filtering, to solve this problem, and proved it T = 0.! sec. The loss of structure is drastic for II.1 to work properly. Its implementation on a laboraand 1.0 sec. but already tnild for 4.0 sec. Similar tory computer is feasible both regarding core and speed of computation. To achieve high accuracy results were obtained for averaged potentials. results, a closer look at the characteristics of the Numerical A ccuracy EEG amplifier is appropriate; any deviations from The followitig test for the nutnerical accuracy of the situation we encountered can be treated in a the tnethod is adtnittedly an incomplete one but similar way, following the recotnmendations of this could nevertheless serve as a guideline. A sweep article. Judging from our work with mentally handrecorded with a time constant of 1 sec was filtered icapped children, it is advantageous to record with to 10.0, 4.0. and 0.1 sec and afterwards filtered low time constants even if a higher one is available, and filter later to higher time constants: as Table 1 back to 1 sec. The differences from the original shows, a time constant as low as 0.1 sec is feasible curve were in all cases of the order of 10 ' ^lV for recording. which is completely negligible. The approach by Elbert and Rockstroh (1980) is Discussion simpler but less versatile. It allows only for the corIn the past much controversy has arisen out of rection of the time constant to d.c. level (time conthe technical question of comparability of results stant equal to infinity). 240 Gasser, Kneip, and Verleger We have applied both methods to an averaged potential with a time constant of O.I sec. From our resultsweconcludethat both methods brought out the slow components; however, Elbert and Rockstroh's method introduced a falling time trend for which there was no evidence in the original data. (A comparison with data recorded with a time constant of 1.0 sec leads to less drastic differences.) It has to be conceded that the method suggested by Elbert and Rockstroh is designed for filter charac- Vol. 19. No. 2 teristics(l)and(2),andthatamodificationrelating it to (3) and (4) would probably improve the results. We consider it an advantage that two methods are now available to solve problems associated with recordings of different or inadequate time constants; the one introduced by Elbert and Rockstroh is highly specific and restricted, but simpie, the one suggested by us is more complex, but highlyfiexibleand easy to generalize to other needs and different equipment. REFERENCES Cooley, J.W., & Tukey, J.W. An algorithm for the machine calculation of complex Fourier series. Mathematics of Computation, 1965,29,297-301. Cooper, R., Osselton, J. W., & Shaw, J. C. EEG techno!ogy. London; Butterworths, 1974. Duncan-Johnson,C. C.,& Donchin, E.Thetimeconstantin P300 recording. Psychophysiology. 1979,16. 53-53. Elbert, T., & Rockstroh, B. Some remarks on the developmeni of a standardized time constant. Psychophysiology, 1980,/7,504-505. Oppenheim, A. V.,& Schafer, R. W. Digital signal process//j^. Englewood Cliffs, New Jersey: Prenlice Hiill, 1975. Verleger, R.. Gasser. T., Bacher P., & Weingartner. O. Analysing poor performance by recording event related potentials of the EEG. In P. Mittler (Ed.), Frontiers of knowledge in mental retardation: Biomeirical aspects. Baltimore, Maryland: University Park Press, 1981. Pp. 187-196. REFERENCE NOTE 1. Deufihard, P., & Apostolescu, V. A study of ihe GaussNewton merhod for the solution of nonlinear least sqtiares problems (Tech. Rep. Nr. 51, SFB 123). Heidelberg: Universitat Heidelberg, 1980. (Manuscript received February 19, I981;acceptedforpublication July 29, 1981) Announcements Research Associate Position In Clinical Psychophysiology and Behavioral Medicine Duke University has a position open as of January 1,1982, for a new Ph.D. with training in psychophysiology and/or experimental clinical psychology and a strong interest in autonomic nervous system physiology. This is an opportunity to work with a strong interdisciplinary team actively engaged in research in the area of behavioral medicine. Interested applicants should send curriculum vita to: Richard S, Surwit, Ph.D., Box 3926, Duke University Medical Center, Durham, NC 27710. Postdoctoral Fellowships At Duke University The Duke University Center for the Study of Aging and Human Development is offering postdoctoral fellowships in adult development and/or aging. Research training is available in many biomedical, behavioral, and social science disciplines. To request application materials, reply with vita and brief description of proposed research area. Completed applications should be received by April 26, 1982 for positions starting in June 1982 and positions for the 1982-1983 academic year. Contact: Dr. Ilene C. Siegler, Box 3003. Duke University Medical Center, Durham, NC 27710. An equal opportunity/affirmative action employer.