Advancing Post-Traumatic Seizure Classification and Biomarker Identification: Information Decomposition Based Multimodal Fusion and Explainable Machine Learning with Missing Neuroimaging Data

Abstract

Late post-traumatic seizure (LPTS) is a complication of traumatic brain injury (TBI), which can lead to a potentially lifelong condition of post-traumatic epilepsy (PTE). Currently, the patho-mechanism that induces epileptogenesis in TBI subjects is unclear. As such, the epilepsy community strives to identify which TBI subjects will develop epilepsy and find potential biomarkers. To that end, this study collects longitudinal multimodal data from TBI subjects at multiple participating institutes. A supervised, binary classification task is formed with data from the LPTS versus no LPTS subjects. Missing modalities in certain subjects is handled in two ways. First, we extend a graphical model based Bayesian estimator to directly classify subjects with missing modality, and second, we investigate standard imputation techniques. The multimodal information is then combined, following several fusion and dimensionality reduction techniques found in literature, and eventually fitted to a kernelor a tree-based classifier. For this fusion, we propose two new algorithms: recursive elimination of correlated components (RECC) which filters information based on correlation, and information decomposition and selective fusion (IDSF) which meaningfully recombines information from decomposed multimodal features. Based on the cross-validated area under the curve (AUC) score, we find the proposed IDSF algorithm provides the best performance. Finally, following statistical analyses of the frequently selected features, we recommend alterations in inferior temporal gyrus as a potential biomarker.