Abstracts

The object is to develop the Video EEG Expert System (VES), a computer program to analyze patient data with the goal of finding anatomic seizure focus in the context of epilepsy pre-surgical evaluation. The system behavior is intended to model the decision making process of the epilepsy patient evaluation team which typically consists of neurologists, neurosurgeons, neuropsychologists, and nurses. Patient data included clinical information, MRI, PET, SPECT, and video EEG. The system is written in Lisp and a logic programming module, Prolisp, that supports hypothesis-driven rules. Benchmark files (controls) and corresponding anatomic localization hypothesis trees were created. A numeric instrument (called the confidence factor) for representing uncertainty was employed to encode data quality (e.g., a mildly suggestive MRI finding was assigned the value 0.6 instead of 1.0). An explanation facility was written to display the rules and confidence factors that were important to each diagnostic result. 23 experimental files obtained from charts were used to test correctness in localizing seizure focus. The principle developer was blind to these files. The system selected the correct localization for 100% of benchmark files (which is expected since these files were created with data clearly associated with a specific localization). For experimental files, VES selected the correct lateralization in 23/23 (100%), the correct localization in 18/23 (78%), and localization and lateralization in 18/23 (78%). First place ties were computed in 13%, and incorrect localizations were found in 9%. Analysis of incorrect results revealed some incomplete knowledge concerning temporal lobe localization. The explanation facility provided feedback on the decision analysis in a clear manner. VES performed well at identifying seizure focus. The tools developed for VES, especially Prolisp, should serve future neurological expert system development. Confidence factors provided an adequate mechanism for truth representation. Further work is planned to incorporate Bayesian statistical methods into the rule analysis thereby providing software users with probabilistic guides for seizure focus lateralization and localization.