Electroencephalography and clinical Neurophysiology, 92 (1994) 89-92 89 © 1994 Elsevier Science Ireland Ltd. 0168-5597/94/$07.00 EEP 93531 Short communication Machine scoring of somatosensory evoked potentials H. Pratt *, N. Mittelman and A.B. Geva Evoked Potentials Laboratory, Behavioral Biology, Gutwirth Building, Technion - Israel Institute of Technology, Haifa 32000 (Israel) (Accepted for publication: 7 October 1993) Summary A machine-scoring algorithm was developed for automatic identification and measurement of the positive and negative peaks of short-latency somatosensory evoked potentials (SEPs). The algorithm enables objective and consistent identification and naming of specific components with minimal operator involvement, avoiding inaccuracies and variability resulting from differences in the criteria used by different operators, or by the same operator at different times. The algorithm is based on finite impulse response filtering of wave forms from 4 conventional recording channels at a bandpass of 90-240 Hz. The bandpass was based on the major lobe in power spectra of multiple records and was verified as effective by application to numerous wave forms. Peak identification is based on identifying the peak at its optimal channel and verifying its consistency with corresponding peaks in the other channels. The machine-scoring algorithm was validated on SEPs from 120 subjects. The machine-scored peak iatencies obtained with this procedure were significantly correlated with their manually measured counterparts. Key words: Somatosensory evoked potentials; Automatic peak measurement; Digital filtering; Machine scoring The most widely used data reduction of evoked potential wave forms is peak analysis, whereby the most positive or the most negative points, within a time interval, are defined and measured. The definition of peaks may be obscured by residual noise in the wave form or by partially overlapping peaks from different genera- tors. Peak definition is typically manual and thus prone to experi- menter bias, adding to test-retest variability of results and, thus, reducing sensitivity. In this report we introduce and validate a machine-based proce- dure for peak identification, naming and measurement for so- matosensory evoked potentials (SEPs). Validation was conducted on clinical records which were often contaminated by residual noise. It is with such noisy records that machine scoring is most needed. The procedure addresses noise biases by rigorous digital filtering. It addresses possible multiple generators by relating to homologous peaks in different recording channels. The machine-based nature of the procedure avoids experimenter bias, but does allow for user intervention in case of obvious misses. Methods One hundred and twenty adults (all males), ranging in age from 21 to 68 (average: 44+ 12) years, with no neurological complaints, participated in the study. Subjects included workers with no known exposure to hazardous substances (45 subjects), a group of 35 sub- jects exposed to lead with blood levels just under the maximum allowable level of 50 /zg/dl blood, and a group of 40 subjects exposed to low levels of mercury with blood levels within normal range. Subjects were chosen to represent a wide range of ages and * Corresponding author. possible systemic factors that might affect the potentials, without adverse effects on the definition of components. Stimuli were 200/~sec duration constant current pulses delivered to a pair of surface electrodes, 1" apart, over the median nerve at the wrist. Potentials were differentially recorded from silver disk elec- trodes (9 mm diameter) arranged in 4 montages: (1) an electrode over the primary sensory cortex (2 cm behind C3 or C4, contralateral to the stimulated side) referred to an electrode over Erb's point (C'c-Erb); (2) an electrode over the second cervical vertebra referred to an electrode on the middle of the forehead (CII-Fpz); (3) an electrode over the primary sensory cortex (2 cm behind C3 or C4, contralateral to the stimulated side) referred to the mid-forehead electrode (C'c-Fpz); and (4) an electrode over the primary sensory cortex contralateral to the stimulated side referred to an electrode over the primary sensory cortex ipsilateral to the stimulated side (C'c-C'i). A grounding electrode was placed on the arm, proximal to the stimulating electrodes. Electrode resistance was maintained be- low 5 kO. Potentials were differentially amplified (x 200,000) at a bandpass of 30-1500 Hz (-3 dB points, 6 dB/octave slopes), bringing the amplified potentials to just within the + 5 V range of the 12-bit analog-to-digital converter of the computer used for aver- aging. The amplified potentials were averaged using 128 addresses and a dwell time of 500 /~sec per address (i.e., analysis time of 64 msec, beginning with stimulus onset), using 1000 sweeps per average. Two repetitions of each average were performed to assess repro- ducibility. Potentials were displayed with positivity at the non-invert- ing electrode as an upward deflection. In the peak identification procedure, peaks and troughs in the filtered wave form were identified in the optimal channels for their detection, verifying identification by comparisons with additional channels. The narrow-band digital filtering resulted in elimination of flat peaks, high frequency noise, as well as stimulus artifacts. Cou- pled with verification across channels, the procedure negated identi- fication of noise- and artifact-related peaks or ambiguity with broad peaks. The procedure began with digital filtering of the averaged SSDI 0013-4694(93)E0261-4