Abstracts

RATIONALE: To validate the FHS algorithm for real-time seizure detection, under conditions closely resembling those in which a portable/implantable device for automated seizure warning/therapy would operate.
METHODS: ECoG recordings (mean duration=112.4 hr.) from 14 subjects undergoing invasive monitoring for epilepsy surgery evaluation were analyzed using generic and adapted modes of the algorithm. Performance measures included: false positive (FP) and false negative (FN) rates. Speed of automated detection in reference to electrographic and clinical onset times from independent expert reviewers is currently under analysis and will be reported.
RESULTS: Ten of the 14 subjects had clinical seizures (CS; n=36) while the generic algorithm was running. The sensitivity for CS detection was 96%. FNs were limited to one subject in whom 2/5 seizures went undetected. Seven patients had no FPs and 3 others had less than 1 FP/day. In the remaining 4, there was disagreement between experts regarding the nature of certain detections: hippocampal spindles (HS) vs. very brief seizures (VBS). If HS, the mean FP rate was 9.3/day (over all subjects) vs. 0.9/day if VBS. Adaptations were only attempted on the 4 subjects with FP rates [gt] 1/day and on the subject with FNs. To test the adaptive power of the algorithm, initial adaptation was performed treating all detections in question as FPs.
To maintain a certain level of blindedness, algorithm adaptations (for FPs and FNs) were performed using the first true positive (visually verified CS) and up to the first three FPs of each type as identified by reviewer. The adaptations were successful in detecting all FNs, and in reducing FP rates ([dsquote]worst case[dsquote] mean [lt] 2.3/day over all subjects). A detailed comparison between generic and adapted algorithm performance will be provided.
CONCLUSIONS: The high sensitivity and specificity of the generic FHS algorithm (w/o training) was demonstrated using long ECoG recordings. Adaptation provided nearly perfect sensitivity and specificity. Given the inevitable differences among scorers regarding what is desirable to detect or ignore, this algorithm demonstrated the ability to conform to either choice. Additionally, the algorithm[scquote]s capability to quantify intensity and duration of activity provides further individualized control over what becomes detectable.
Support: NIH SBIR 5R44NS34630-03; Medtronic, Inc.
Disclosure: Grant - Medtronic to KUMC; Consulting - FHS from Medtronic; Ownership - FHS; Materials - Medtronic to KUMC and FHS; Other - FHS algorithm licensed by FHS from KUMC and sublicenced by FHS to Medtronic