Multi-model Transfer Learning and Genotypic Analysis for Seizure Type Classification

Abstract

The recent progress in phenotypic information and machine learning has led to a remarkable development in the accuracy of binary seizure detection. Yet the performance of classifying specific seizure types remains suboptimal due to the limited availability of annotated data with accurate seizure type labels. Transfer learning is promising to mitigate data scarcity to improve classification accuracy on smaller datasets. However, finding the best transferable model based on the specific training and testing dataset can be a complex and repetitive process, and a single-modelled approach may not fully capture the best feature representation of the input data. Moreover, genotypic data is often neglected in previous AI-based seizure detection studies, where analyses like Polygenic Risk Scores (PRS) could offer insights into genetic predispositions to seizures. To mitigate these challenges, we propose a seizure-type classification framework incorporating a multi-model weighting system designed to assign weights to different models, thus reducing computational complexity and processing time. In addition, we carry out a PRS analysis, aiming to bridge the gap between genotypic and phenotypic data, further enhancing the comprehensiveness and precision of seizure detection. Our model outperformed similar classifiers by more than 13–16% on the Temple University Hospital EEG Seizure Corpus dataset. This study represents a pioneering examination of the multi-source transfer learning framework in the field of type-specific seizure classification.