Monitoring kidney patients by syntactometric eeg analysis

MONITORING KIDNEY PATIENTS BY SYNTACTOMETRIC EEG ANALYSIS C. Hernbndez

Sande

and J.E. Arias

Rodriguez

ABSTRACT

refinement of the method also takes into account the second largest peak The distance of the resulting 32-character ‘sentences’ from a pattern is calculated using a string-tostring metric based on weighted insertions, deletions and substitutions (insertions and deletions are included to allow for artefacts detected tn the signal). Elementary weights have been assigned on em,Cical medical grounds taking into account the neurological significance of the vartous spectral bands and their correlation with the levels of creattnine and urea tn the bloodstream

The EEG of seriously ill ktdne! patients rejects changes in their condition which c-an be measured quantttatiuely b.y a syntacttc pattern recognition techntque. 128 s of an EEG were sampled at 100 Hz and segmented into 4 set blocks, each of which was labelled D (or d), T (or t), N (or n), S (or s), or B (or b) depending on the position of its main peak m the power spectrum: delta band (O-4 Hz), theta band (g-8 Hz), alpha band (a- 12 Hz), sigma band (12-11 Hz) or beta band (over 14 Hz). The use of capital or small Letters depends on the size of the peak A Keywords:

Patient

monitors,

elertroencephalo~raphs,

automauc

EEG processing,

METHODS

INTRODUCTION The EEG record is increasingly being used as a complementary aid in the diagnosis and monitoring of primary extra-cerebral disorders with secondary cerebral effects, since it allows noninvasive observation of the latter. In patients suffering kidney failure, slow delta and theta rhythms appear as their azotemic index rises, and fade on favourable response to treatment. Although the cerebral effects of kidney disease were first studied in the early 1940s’ systematic research was not carried out until 30 years late?. In 1976 the United States National Cooperative Dialysis Study initiated a nationwide programme to set up a kidney patient data bank employing uniform EEG reports, thereby permitting the development of a computer algorithm for their classifications. The analysis of EEG records by signal processing methods is more discriminating than visual inspection; it allows a refinement of the quantification of EEG abnormalities and a closer monitoring of patient progress. Given the fairly well-defined division of cerebral activity into different wavebands, syntactic pattern recognition techniques appear to be particularly appropriate for the description of EEG structur$, they offer the additional advantage of presenting the information in a form similar to that involved in the mental processes of the electroencephalographer carrying out a visual analysis. We describe a method of monitoring kidney patients’ progress in which the disparity between EEGs recorded at periodic intervals, is quantified by applying a string-to-string metric to EEG ‘words’ whose terminals are identified by a simple decision algorithm.

Department of Electronics, Faculty of Physics, Compost&, Spain. Reprints from Dr C. Hernindez Sande C 1985 Buttetworth 0141-5425/85/040334-03 334

J

Biomed

kidney disease

& Co (Publishers) $03.00

Eng. 1985, Vol. 7, October

University

Ltd

of Santiago

de

EEGs were recorded from a 50-year-old patient performing simple arithmetic operations with his eyes shut. A 16 channel electroencephalograph in a bipolar lo-20 arrangement was used, though only four channels were subsequently processed (P3:0 1, P3-C3, P4-02, F4-C4). ‘Analysis was carried out on an HP-9825-T system with 64 k words of memory. Preprocessing For each channel four 128s records exhibiting a minimum of irrelevant muscular activity were digitized at a frequency of 100 Hz, divided into four 32s blocks and after passing through a Hamming window, stored on magnetic tape. Each block was then subdivided into 4s segments whose spectra were obtained using a 1024 point fast Fourier transform. Segment

labelling

The two spectral peaks of greatest half-amplitude breadth were determined in each segment, their amplitudes and frequencies obtained, and the segments labelled D, T, N, S,B, A or X depending upon whether the dominant feature was a delta peak (O-4 Hz), a theta peak (4-8 Hz), an alpha peak (8-12 Hz), a sigma peak (12-14 Hz), a beta peak (14 Hz), or an artefactual peak due to irrelevant muscular activity or of such low amplitude that it could not be classified as belonging to any of the preceding categories. Following Bourne’, the importance of peaks was judged primarily on the basis of their width, the widest being deemed the dominant feature. If there were two peaks of unequal amplitude but equal breadth, the higher amplitude was preferred. Small letters were used for labels if the amplitude of the main peak was less than half the mean amplitude of the 32s block

Kidney patient monitoring: C. Hemhndez

Figure

1

Detection

Examples

of EEG segment labelling

of artefacts

The low frequency range of EEGs is prone to interference from two sources, artefact peaks due to muscular activity and slow rhythms due to drowsiness; both may be mistaken for slow pathological signals. The second problem may be avoided by recording EEGs from alert patients and asking them to carry out simple arithmetic operations during the recording session. Muscular and ocular artefacts are rather more difficult to exclude, and must be identified and neutralized from the traces in which they occur’. A peak initially classified as a delta was reclassified as an artefact if: a) its frequency

Sande and J. 6:. Arias Rodriguez

lay below 1.5 Hz

b) its low amplitude and frequency warranted a d label, because it might be due to baseline drift; or c) its width was greater than 5 Hz (or greater than 3 Hz if initially classified as a delta peak); or d) the second most relevant first.

peak overlapped

the

Segments dominated by artefact peaks were labelled according to the location of the second most relevant peak so long as the latter’s amplitude warranted a capital label, failing which they were labelled with an A. This is illustrated in Figure 1.

insertions and deletions) that one string must undergo in order to coincide with a second string. In the present case insertions and deletions need not be considered separately because the ‘gaps’ in the strings (i.e. those 4s segments lacking pronounced peaks or dominated by artefact peaks) were themselves assigned labels (X or A). The transformations were weighted in accordance with an estimate of the clinical significance of the corresponding spectral shifts (~&de I ), improvements being given negative weights and changes for the worse, positive weights. RESULTS

AND DISCUSSION

Figure 2 illustrates the calculation of a Levenstein distance between the four 32s blocks of one of the April 1983 (P4-02) records and the corresponding December 1982 blocks. The total distance of 8 between the two records suggests that the patient’s condition is worsening. T&e 2 shows, for each EEG channel, the total Levenstein distances of the four records taken in each 1983 session with reference to the December 1982 records; they faithfully reflect the steady worsening shown by the levels of creatinine and urea between December 1982 and August 1983, and the improvement observed in the final Table 1

The string-to-string

distance

EEGs were recorded in December 1982, April 1983, August 1983 and November 1983. On the last three occasions, once labelling had transformed each of the four 128 s records of each of the four channels processed, into four eight-letter strings, the disparity between each of these strings and the corresponding December 1982 string was quantified as a weighted Levenstein distance using an algorithm due to Wagner and Fishefl. Essentially, the Levenstein distance counts the number of transformations (substitutions,

Substitution

Weights

D

T

D T

0

-1

1

0

N

2 2

1 2

0 0

0 0

0 0

1

M

0

0

0

S,B .%x

N

S,B

.%x

-2

-2

-1

-1

-2

-1%

A substitution between a small letter label and capital label is given half the weight of the corresponding transformation between two small letter labels which is given a quarter of the weight of the corresponding capital transformation. Thus weight (n+ D) = 1 and weight (n + d) = 54.

J. Biomed Eng.

1985.

Vol. 7. October

335

&dney patient monitoring: C. Hernkdez Table

Sande and J. E Arias Rodriguez

2 Progress from

(mg/lOO

1982

to November

1983 Levenstein

Blood Ure;tNitrogen (mg/lOO ml)

Creatinine

Date

December

ml)

Distance

(Each value represents

4 I 28 s records) Channel

December April

1982

1983

August

1983

November

1983

l&c3

F4-C4

P4-02

P3-01 -

-

30

41

40

a.7

68

10.2

75

35

12.6

130

58

55

64

60

8.4

71

50

39

56

48

OAdtlnON fOTOOONO distance.2

O@ Dd

IO t I

RON nA0

0 0

distance-l

90 90 ?O

nnAAOOl0

OnOOIOnO distance+]

0 n 0 1 ND0 0 10 n ODD

T 0

60

distance+? 50 *0 30 20

10

DlC

JM

FEE

IU

9P9

191

Levenstein distance Figure 3 (sums for 4 128s records)

JUI

from

JUL

IU6

December

SIP

WI

801

DEC

1982 records

ACKNOWLEDGEMENT This work is financed by the Comision Asesora de Investigation Cientifica y Tecnica under Project No. 2806/8.3. REFERENCES N

Figure 2 Calculation of the Levenstein distance between the four 32s blocks of one of the April 1983 (P4-02) records and the corresponding December 1982 blocks. The total distance of 8 between the two records suggest that the patient’s condition is worsening

November 1983 session. The ability to quantify changes in EEG performance during observation of the progress of kidney failures may be illustrated by trend charts (Figure 3) and increases the utility of EEG monitoring in comparative studies of different therapies or in the evaluation of dialysis requirements. The system could include a heuristic algorithm for refinement of the string-to-string transformation weights on the basis of biochemical data, or a facility for the clinician to adjust weights himself in accordance with the peculiarities of each patient.

336

J. Biomed Eng. 1985, Vol. 7, October

Romano, J. and Engel, G.L Delirium. I. Electroencephalographic data Arch Neural Psychiat, 1944, 51,356377 Bourne, J. R, Miezin, F. M., Ward, J.W. and Teschan, P. E. Computer quantification of electroencephalographic data recorded from renal patients. Comput Biotned Rex, 1975, 8, 46 l-473 Chotas, H.G., Bourne, J.R and Teschan, P.E. Heuristic techniques in the quantification of the electroencephalogram in renal failure Conp. Biomed Rex, 1979, 12,299-312 Bowling, P.S. and Bourne, J.R. Discriminant analysis of electroencephalograms recorded from renal patients IEEE Trans. Biomed Eng., 1978, BME25, 12-17 Bourne, J.R, Hamel, B., Giese, D., Woyce, G.M., Lawrence, P.L, Ward, J.W. and Teschan, P. E. The EEG analysis system of the national cooperative dialysis study. IEEE Trans. Riomed Eng., 1980, BME21, 656-664 Boume, J.R, Jagannathan, V., Hamel, B., Jansen, B.H., Ward, J.W., Hughes, J.R and Erwin, C.W. Evaluation of a syntactic pattern recognition approach to quantitative electroencephalographic analysis. Elect Clin Neurophysial, 1981, 52, 57-64 Jansen, B.H., Bourne, J.R and Ward, J.W. Identification and labelling of EEG graphic elements using autoregressive spectral estimates. Compul. 3iaL Med, 1982, 12 97-106 Wagner, Rk and Fisher, M.J. The string-tostring correction problems. J Ass. Cornput. Mach, 1974, 21