Customized GPT model largely increases surgery decision accuracy for pharmaco-resistant epilepsy

Abstract

Background: To develop an enhanced epilepsy diagnosis system by integrating an expert-informed ontology with a custom generative pre-trained transformer (GPT), validated by inferring possible seizure lateralization and localization using retrospective textual data from the pre-surgical assessments of patients with pharmaco-resistant epilepsy (PRE).

Methods: We developed an AI system for epilepsy diagnosis using Protégé with OWL/SWRL, integrating a knowledge base with seizure semiology, seizure types EEG descriptors, expert insights, and literature to pinpoint seizure locations. A customized GPT model was then tailored for specific diagnostic needs. Validated through 16 surgical cases, the system's accuracy in seizure localization and the JSON (JavaScript Object Notation) Epilepsy Matcher's term matching capabilities were confirmed against a Protégé-based knowledge base.

Results: A total of 117 patients with PRE underwent video-EEG monitoring at a single institution. However, only 16 of these patients received epilepsy surgery. The Protégé system achieved 75 % accuracy in diagnosing epilepsy from 16 cases using semiology, which increased to 87.5 % with EEG data. The Json Epilepsy Matcher further improved accuracy to 87.5 % with symptoms alone and 93.8 % when including EEG data, highlighting the benefits of applying GPT techniques.

Conclusions: This study highlights the efficacy of the JSON Epilepsy Matcher in improving seizure diagnosis accuracy. When combined with EEG data, it achieves a 93.8 % accuracy rate, suggesting a potential improvement in the practicality and generalizability of the original ontology expert system, boosting physicians' confidence in confirming surgery and potentially sparing many children from prolonged suffering. This innovative approach not only improves diagnostic accuracy but also sets a precedent for future applications of AI in neurology.

Keywords: Generative pre-trained transformer; Large-scale language model; Localization; Seizure descriptors; Semiology.