Abstracts

Rationale: Automated seizure detection would have many uses in epilepsy research, including anti-epileptic therapy discovery. This study aimed to assess the accuracy of SeizureScan 1.00 (Clever Systems Inc, Reston VA) seizure detection software to measure behavioral-seizure frequency and severity. Methods: The behavior of kainate- and pilocarpine-treated rats was recorded for 2-4 h periods, and was manually observed, analyzed, and compared to results using SeizureScan. For experiments 1 and 2, behavior was recorded with a single camera, but two cameras were used for experiments 3 and 4. The positive predictive value (PPV, or true positives divided by the sum of true positives and false positives) and sensitivity (true positives divided by the sum of true positives and false negatives) were assessed. Results: In experiment 1, the seizure detection software identified 18 of the 23 (78%) manually detected seizures. Five of the 18 seizures were correctly classified according to the Racine motor-seizure scale, while the remaining 13 were identified but classified incorrectly. There were 5 false negatives (22%), 7 false positives and 15 seizures that the program divided into two seizures because of a pause in activity. The program had a sensitivity of 82.1% and PPV of 76.6%. In experiment 2, SeizureScan identified 60 of 82 (73%) manually detected seizures. Thirty seven of the 82 seizures were correctly classified according to the Racine scale, while the remaining 23 were identified but classified incorrectly. There were 22 false negatives (27%), 212 false positives and 1 divided seizure. The program yielded a sensitivity of 78.8% and PPV of 27.8%. In experiment 3, which utilized input from two cameras, the software detected 46 of 59 (77.96%) manually identified seizures in kainate-induced epileptic rats. There were 13 false negatives (22.04%) and 222 false positives recorded by the software. The sensitivity and PPV were 81.9% and 20.9%, respectively. In experiment 4, with pilocarpine-treated rats, improvements in background selection (contrast dark black background and recording cage clip without rat) were made, and SeizureScan identified 46 of 47 (97.8%) manually detected seizures, with only 1 false negative (2.2%) but 87 false positives recorded by SeizureScan. Twenty six of 46 seizures were correctly classified, while the remaining 20 were incorrectly classified. In this experiment, the sensitivity was increased to 97.91% and PPV was 35.07%. Conclusions: SeizureScan 1.0 appears to be efficient for automated seizure detection. Adjustments during video recording and in parameter settings, led to a significant improvement in seizure detection. However, large file-size and slow analysis-speed need to be addressed before this software can be used consistently and efficiently for chronic experiments that require long-term continuous video monitoring. (Supported by a subcontract from an NIH SBIR to Clever Sys. Inc.)