5
ELSEVIER Electroencephalography and clinical Neurophysiology 103 (1997) 445-449 A practical analysis of computer based seizure detection during continuous video-EEG monitoring M.C. Salinsky* Oregon Health Sciences University and Portland Veterans Affairs Medical Center, Portland, OR, USA Accepted for publication: 14 March 1997 Abstract Computer based seizure detection (CSD) systems improve the efficiency of CCTV-EEG monitoring by capturing epileptic seizures which would have otherwise bee:a missed. We prospectively evaluated the yield of a commercial CSD system in 83 consecutive CCTV-EEG admissions. All seizures were coded as to the method of detection. The percentage of seizures detected only by CSD was calculated for each patient, and the impact on ',length of hospital stay was estimated. Overall, 33% of epileptic seizures were signaled by the patient, 45% were directly observed by family or medical personnel, and 22% were captured only by CSD. Forty admissions (48%) had at least one seizure captured only by CSD. The majority of these events were clinical and electrographic seizures (73%) and the remainder were purely electrographic. Five admissions concluded with all seizures captured only by CSD. We estimated an average saving of 1.3 hospital days per admission, based on the percentage of seizures captured only by CSD. © 1997 Elsevier Science Ireland Ltd. Keywords: Electroencephalography; Telemetry; Epilepsy; Seizure; Computer aided diagnosis 1. Introduction Continuous long-tern1 closed circuit TV-EEG (CCTV- EEG) monitoring, an accepted method for detailed analysis of seizures, is used for diagnostic evaluations and for sei- zure onset localization in patients being considered for epi- lepsy surgery (Gumnit, 1987; Kaplan and Lesser, 1990). The efficiency of CCTV-EEG monitoring is affected by the ability to detect seizures. Seizures are usually detected by patients pressing a bedside push-button alarm, or by direct observation of fanlily or medical personnel. However, patients may be unawale of seizures or unable to respond, and observers are not always present. In one series only 40% of 208 epileptic seizures (ES) were signaled by patients or observers (Pierelli et al., 1989). Additional seizures can be detected by reviewing the continuous EEG record. How- ever, a complete review of lengthy CCTV-EEG records is extremely time consurning and fatiguing, and impractical for routine clinical practice. Gotman (1982) described a * Oregon Health Sciences :University Epilepsy Center (CDW3), 3181 S.W. Sam Jackson Park Road, Portland, OR 97201, USA. Tel.: +1 503 4945682; fax: +1 503 4946558; e-mall: [email protected] method for automated computer based seizure detection (CSD) as an alternative to complete record review. The computer program is typically set to be overly sensitive in order to capture a high percentage of ES at the expense of many false positive detections. This results in an enriched EEG file, substantially shorter than the original continuous record but containing a relatively high concentration of ES. Improvements to the seizure detection algorithm have sub- sequently been described by Gotman (1990) and Qu and Gotman (1993), (1995). Alternative methods have been pro- posed (Principe and Smith, 1985; Panych et al., 1988; Murro et al., 1991). The practical value of CSD is determined by the number of ES detected only by this method, i.e. those which would have been missed by relying on direct observation and push- button alarms. Seizures detected only by CSD decrease the length of hospital stay by accumulating toward the total number of seizures needed to answer the specific clinical question. This decreases the total hospitalization cost. In a previous study, Gotman (1990) reported that 41% of 244 ES recorded from 49 patients were detected only by CSD. This prospective study evaluated the effect of CSD on the efficiency of inpatient CCTV-EEG monitoring. We deter- 0013-4694/97/$17.00 © 1997 Elsevier Science Ireland Ltd. All rights reserved PII S0013-4694(97)00025-4 EEG 96151

A practical analysis of computer based seizure detection during continuous video-EEG monitoring

Embed Size (px)

Citation preview

Page 1: A practical analysis of computer based seizure detection during continuous video-EEG monitoring

E L S E V I E R Electroencephalography and clinical Neurophysiology 103 (1997) 445-449

A practical analysis of computer based seizure detection during continuous video-EEG monitoring

M.C. Salinsky* Oregon Health Sciences University and Portland Veterans Affairs Medical Center, Portland, OR, USA

Accepted for publication: 14 March 1997

Abstract

Computer based seizure detection (CSD) systems improve the efficiency of CCTV-EEG monitoring by capturing epileptic seizures which would have otherwise bee:a missed. We prospectively evaluated the yield of a commercial CSD system in 83 consecutive CCTV-EEG admissions. All seizures were coded as to the method of detection. The percentage of seizures detected only by CSD was calculated for each patient, and the impact on ',length of hospital stay was estimated. Overall, 33% of epileptic seizures were signaled by the patient, 45% were directly observed by family or medical personnel, and 22% were captured only by CSD. Forty admissions (48%) had at least one seizure captured only by CSD. The majority of these events were clinical and electrographic seizures (73%) and the remainder were purely electrographic. Five admissions concluded with all seizures captured only by CSD. We estimated an average saving of 1.3 hospital days per admission, based on the percentage of seizures captured only by CSD. © 1997 Elsevier Science Ireland Ltd.

Keywords: Electroencephalography; Telemetry; Epilepsy; Seizure; Computer aided diagnosis

1. Introduct ion

Continuous long-tern1 closed circuit TV-EEG (CCTV- EEG) monitoring, an accepted method for detailed analysis of seizures, is used for diagnostic evaluations and for sei- zure onset localization in patients being considered for epi- lepsy surgery (Gumnit, 1987; Kaplan and Lesser, 1990). The efficiency of CCTV-EEG monitoring is affected by the ability to detect seizures. Seizures are usually detected by patients pressing a bedside push-button alarm, or by direct observation of fanlily or medical personnel. However, patients may be unawale of seizures or unable to respond, and observers are not always present. In one series only 40% of 208 epileptic seizures (ES) were signaled by patients or observers (Pierelli et al., 1989). Additional seizures can be detected by reviewing the continuous EEG record. How- ever, a complete review of lengthy CCTV-EEG records is extremely time consurning and fatiguing, and impractical for routine clinical practice. Gotman (1982) described a

* Oregon Health Sciences :University Epilepsy Center (CDW3), 3181 S.W. Sam Jackson Park Road, Portland, OR 97201, USA. Tel.: +1 503 4945682; fax: +1 503 4946558; e-mall: [email protected]

method for automated computer based seizure detection (CSD) as an alternative to complete record review. The computer program is typically set to be overly sensitive in order to capture a high percentage of ES at the expense of many false positive detections. This results in an enriched EEG file, substantially shorter than the original continuous record but containing a relatively high concentration of ES. Improvements to the seizure detection algorithm have sub- sequently been described by Gotman (1990) and Qu and Gotman (1993), (1995). Alternative methods have been pro- posed (Principe and Smith, 1985; Panych et al., 1988; Murro et al., 1991).

The practical value of CSD is determined by the number of ES detected only by this method, i.e. those which would have been missed by relying on direct observation and push- button alarms. Seizures detected only by CSD decrease the length of hospital stay by accumulating toward the total number of seizures needed to answer the specific clinical question. This decreases the total hospitalization cost. In a previous study, Gotman (1990) reported that 41% of 244 ES recorded from 49 patients were detected only by CSD.

This prospective study evaluated the effect of CSD on the efficiency of inpatient CCTV-EEG monitoring. We deter-

0013-4694/97/$17.00 © 1997 Elsevier Science Ireland Ltd. All rights reserved PII S0013-4694(97)00025-4 EEG 96151

Page 2: A practical analysis of computer based seizure detection during continuous video-EEG monitoring

446 M.C. Salinsky /Electroencephalography and clinical Neurophysiology 103 (1997) 445-449

mined the percentage of seizures captured only by CSD, and estimated the impact on length of hospital stay.

2. Methods

We studied all patients undergoing CCTV-EEG monitor- ing at the Oregon Health Sciences University/Portland Veterans Affairs Medical Center during 1992-1994. The hospitals shared a two bed CCTV-EEG unit. Patients were admitted for diagnostic telemetry or presurgical evaluation. Patients without definite electrographic seizures recorded during the admission were excluded from analysis. Eighty-three admissions met the criteria, including 72 patients and 489 monitoring days. The median length of stay was 5 days (range 2-13 days). The cable videoteleme- try systems utilized 24-64 EEG channels. Sixty-seven recordings (81%) used surface or surface/sphendidal elec- trodes, and 16 (19%) used intracranial electrodes (10 fora- men ovale, 4 subdural strips, 2 subdural grids).

CCTV-EEG was monitored 24 h per day, 7 days per week. EEG technologists were on site 10 h per weekday. Telemetry rooms had a guest bed permitting family mem- bers to stay overnight. Live video was displayed at the nur- sing station and EEG workroom. Telemetry beds were equipped with push-button alarm systems which produced audible alarms, and triggered a dedicated pager worn by the nurse.

CSD was performed on a personal computer using Stel- late systems MONITOR program. From 1992 through 1993, 16 EEG channels were used for CSD (selected based on seizure characteristics). Subsequently, detection was per- formed on up to 32 channels. Default CSD settings were used during the initial 24 h of monitoring. Settings were then modified based on performance. In addition to seizure detections, the program acquired 1 min of EEG every 30 rain between 2300 h and 0800 h and performed spike detec- tion with 5-25% of detected spikes saved as 2-s EEG seg- ments. EEG files written by the MONITOR program were reviewed by a clinical neurophysiologist daily. Segments suggestive of electrographic seizure activity were replayed from videotape for clinical correlation.

Each ES was logged at the time of review and classified by the method of detection, i.e. (P) push-button detection (the patient pressed the alarm button immediately before, during, or within 10 min after the ES), (O) direct observa- tion by family or medical personnel (including seizures detected by observation of post-ictal confusion), or (C) computer detection only. Many ES were detected by more than one method. For study purposes a single letter code was assigned to each ES based on the priority P > O > C. Computer detections were assigned lowest priority since the goal was to determine the percentage of seizures that would have been missed had CSD not been used. Computer detections were further classified as being clinical ES or purely electrographic seizures. For each patient, the length

of CCTV-EEG stay, total number of ES, and the number and percentage of P, O, and C detections were recorded.

To estimate the impact of CSD on the length of hospital stay (for each admission) the following steps were taken: (1) the number of observed or push-button detections (O + P) was divided by the actual length of stay (ALOS) to give the number of non-computer detected ES per hospital day (observed + push-button detection rate (OPD)); (2) the total number of recorded ES was divided by the OPD ((O + P + C)/OPD) to provide the projected length of stay (PLOS) required to reach the same total ES count in the absence of CSD; (3) the difference between the projected length of hospital stay and the actual length of stay ( P L O S - ALOS) was calculated for each CCTV-EEG admission; and (4) the cost associated with the additional hospital days was estimated by multiplying PLOS - ALOS by $1500/day (a conservative average of all charges per day in our geographic region). Calculations were based on the total number of CSDs, and separately for CSDs excluding purely electrographic seizures.

Two adjustments were made to provide a conservative estimate of additional hospital time and cost in the absence of CSD. Five CCTV-EEG admissions concluded with only C detections. This would have resulted in a division by zero (OPD = 0) using the above analysis. These patients were arbitrarily assigned a PLOS - ALOS difference of 1 hospi- tal day per ES detected. The second adjustment concemed the maximum value for PLOS - ALOS. In some cases the majority of epileptic seizures were detected only by CSD, and the value of PLOS - ALOS was large. From a practical standpoint there are limits on the length of stay in CCTV- EEG units, even if the total number of captured seizures has not reached the desired goal. Therefore, PLOS - ALOS was arbitrarily capped at a maximum value of 5 days per admis- sion.

3. Results

The mean age for the 72 patients was 32.4 years (range 11-56 years). A median of 5 (range 1-35) ES were detected per admission. Of 83 admissions, 27 (33%) yielded three ES or less, and 14 (17%) yielded more than ten ES. The number of seizures captured and length of hospital stay were not significantly correlated.

Fig. 1 illustrates the method of seizure detection for the 576 ES. The percent of seizures detected by each method was calculated for each admission and then averaged across all 83 admissions. An average of 78% of ES were captured by direct observation or by the patient reporting via the push-button alarm. The remainder (22%) were detected solely by review of CSDs. Seventy-two percent of directly observed seizures were also detected by CSD.

Forty admissions (48%) had at least one ES detected only by CSD. Thirty-five admissions (42%) had at least one clin- ical ES detected, and 13 had at least one purely electro-

Page 3: A practical analysis of computer based seizure detection during continuous video-EEG monitoring

M.C. Salinsky /Electroencephalography and clinical Neurophysiology 103 (1997) 445-449 447

graphic ES detected by CSD alone. Overall, 73% of CSDs were clinical and electrographic seizures, and 27% were purely electrographic. Fifty-five percent of these CSDs were sleep onset events, and 49% occurred during the night shift (2300-0700 h). The percentage of ES detected only by CSD was not significantly correlated with the total number of recorded seizures (Spearman correlation, P = 0.21). However, the 14 patients with more than ten ES had relatively few detected by CSD only (10.8%).

Five patients had all ES captured by CSD only. Their admission length ranged fi:om 4 to 13 days. Four of these patients had only one clinical ES, all complex partial. Three of these occurred between 2400 h and 0300 h, from sleep. One patient had four complex partial seizures all of which occurred from sleep betwe, en 0200 h and 0700 h. Patients were not aware that any of these seizures had occurred. Two of these patients had additional purely electrographic sei- zures.

The projected additional hospital stay for each patient, assuming that CSD was not available (PLOS - ALOS), is illustrated in Fig. 2. Forty-three admissions (52%) did not have any unique CSDs. Therefore, the length of stay would not have changed. The remaining CCTV-EEG admissions would have required additional hospital days to record the target number of ES. The number of additional hospital days (PLOS - ALOS) was capped at a maximum of 5 per admis- sion. With this adjustment~. 107 additional monitoring days would have been required (average of 1.3 additional days per admission). Based on the projected additional length of stay, and a total cost of $1500/day, a mean additional cost of $1933 per admission would have resulted. This cost is somewhat misleading in that more than half of the CCTV- EEG admissions did not benefit from CSD and the median additional cost for the 83 admissions was $0. However, considering the entire 3 year period, a total projected cost

of $160 458 would have resulted from extension of hospital stays ($124995 if purely electrographic seizures were excluded from consideration).

4. Discussion

The impact of CSD on the efficiency of CCTV-EEG monitoring depends on the circumstances under which the evaluations take place. For example, our data was obtained in a telemetry studio which does not have an EEG technol- ogist in attendance 24 h per day. In centers with 24 h tech- nologist surveillance the value of an automated seizure detection system may be reduced. Other centers may routi- nely use one on one nursing with similar results. On the other hand, we encouraged overnight family visitors, which potentially increased the number of directly observed seizures. Differences in alarm systems, physical design of the studio, recording systems and personnel all effect CCTV-EEG monitoring efficiency and the practical yield of CSD. Our data was collected using the commercial CSD program MONITOR. Other programs may have dif- ferent sensitivity and specificity.

CSD had a significant impact on CCTV-EEG monitoring efficiency at our unit with 22% of ES detected only by this method. Over 40% of admissions had at least one clinical ES detected by CSD only and in five patients all ES were found only by review of CSDs. These findings are in general agreement with Gotman (1990). The lower rate of ES detected only by CSD in this study may reflect a higher rate of directly observed seizures, and a different method for calculation of the overall percentages. CSD is particu- larly useful for the detection of nocturnal and sleep onset seizures. Hospital stays were projected to be an average of 1.3 days shorter due to the higher dally yield of ES. Con-

• • = i • • • • i ~'¸¸•

Fig. 1. Method of detection of 576 seizures during 83 CCTV-EEG admissions. Percentages were calculated for each admission and averaged across all admissions. Computer detected seizures were those which were not reported by the patient and were not directly observed.

Page 4: A practical analysis of computer based seizure detection during continuous video-EEG monitoring

448 M.C. Salinsky / Electroencephalography and clinical Neurophysiology 103 (1997) 445-449

5

"~ 4 J a

a

3 o

2 H

g I i

Individual CCTV-EEG admissions Fig. 2. Projected number of additional CCTV-EEG monitoring days for each of the 83 CCTV-EEG admissions, assuming that computer based seizure detection was not available.

sidering the entire group of 83 telemetry admissions over a period of 3 years, routine use of CSD led to an estimated 18% decrease in total hospital days. Estimated 3 year cost savings resulting from CSD were in the range of $125 000- $160 000. The cost of a two-bed CSD system similar to ours is <$40000. The system is clearly cost effective in our hospital. The impact of CSD would be greater in units where direct observation of ES is less likely. CSD detected 72% of our observed ES. In a worst case situation (no direct observation) only 13% of all detected seizures would have been missed. However, direct observation and ictaYpostictal examination are extremely important for the characteriza- tion of ES. CSD should be used as an adjunct to, not a replacement for, direct observation and examination.

CSD is fundamentally a labor saving device. In theory, the same or better results would have been achieved by complete manual review of the continuous EEG. Such review is an extremely time consuming process which we assume would not routinely be done (Pierelli et al., 1989). We also assumed that the total number of ES captured dur- ing each admission was the number necessary to answer the particular clinical question. However, in some cases an excess number of ES may have been recorded and the addi- tional information from CSDs may not have impacted the length of stay. Most epilepsy monitoring units attempt to minimize hospital stay by discharging patients as soon as the required information has been obtained. Our data was collected under such circumstances. The exceptions were a few patients who had flurries of seizures which prolonged hospitalization for treatment purposes. These patients were closely observed by medical personnel and the majority of ES were observed, as reflected in the relatively low percen- tage of unique CSDs (11% for admissions with >10 seizures versus 22% for the entire group).

This study did not assess the sensitivity or specificity of CSD. Pauri et al. (1992) have addressed some issues regard- ing the tradeoff between false positive detections and sen- sitivity. Our study was designed to provide practical information on the utility of CSD in clinical practice. Bal- ancing true positive and false positive detections required dally adjustment of seizure detection parameters based on the specific clinical questions. Even with the use of CSD some seizures may have been missed by all detection methods.

A c k n o w l e d g e m e n t s

Thanks to L. Struble R.EEGT, C. Caudle R.EEGT, and the EMU nursing staff, for their help in gathering this data; and to J. Cereghino MD, D. Smith MD and J. Buchhalter MD, PhD for comments and suggestions.

R e f e r e n c e s

Gotman, J. Automatic recognition of epileptic seizures in the EEG. Elec- troenceph, clin. Neurophysiol., 1982, 54: 530-540.

Gotman, J. Automatic seizure detection: improvements and evaluation. Electroenceph. clin. Neurophysiol., 1990, 76: 317-324.

Gumnit, R.J. (Ed.) Intensive Neurodiagnostic Monitoring. Advances in Neurology, Vol. 46. Raven Press, New York, 1987.

Kaplan, P.W. and Lesser, R.P. Prolonged extracranial and intracranial inpatient monitoring. In: J.A. Wada and R.J. Ellingson (Eds.), Clinical Neurophysiology of Epilepsy. Elsevier, Amsterdam, 1990, pp. 121- 154.

Murro, A.M., King, D.W., Smith, J.R., Gallagber, B.B., Flanagin, H.F. and Meador, K. Computerized seizure detection of complex partial seizures. Electroenceph. clin. Neurophysiol., 1991, 79: 330-333.

Panych, L.P., Wada, J.A., Janicijevie, R.D., Steinke, T.G. and Beddoes,

Page 5: A practical analysis of computer based seizure detection during continuous video-EEG monitoring

M.IZ Salinsky /Electroencephalography and clinical Neurophysiology 103 (1997) 445-449 449

M.P. Computer assisted operation of a facility for the long term mon- itoring of seizure patients. Arn. J. EEG Technol., 1988, 28: 211- 229.

Pauri, F., Pierelli, F., Chatrian, {3. and Erdly, W.W. Long-term EEG- video-audio monitoring: computer detection of focal seizure patterns. Electroenceph. clin. Neurophysiol., 1992, 82: 1-9.

Pierelli, F., Chatrian, G.-E. and Erdly, W.W. et al. Long-term EEG-video- audio monitoring: detection of partial epileptic seizure and psychogenic episodes by 24 hour EEG record review. Epilepsia, 1989, 30: 513- 523.

Principe, J.C. and Smith, J.R. Automatic recognition of spike and wave bursts. Electroenceph. clin. Neurophysiol., Suppl. 37, 1985.

Qu, H. and Gotman, J. Improvements in seizure detection performance by automatic adaptation to the EEG of each patient. Electroenceph. clin. Neurophysiol., 1993, 86: 79-87.

Qu, H. and Gotman, J. A seizure warning system for long-term epilepsy monitoring. Neurology, 1995, 45: 2250-2254.