ADMP-GCN: Interpretable Adaptive Message Passing Graph Learning for Patient-specific EEG Seizure Prediction with Human-in-the-Loop Validation

Abstract

Automated seizure prediction based on multichannel Electroencephalography (EEG) using Artificial Intelligence (AI) is a challenging task that mandates accurate and interpretable results for improved clinical decisionmaking.However, brain connectivity modelling on the evolving patient-specific and seizure-dynamic EEG signals makes the problem highly complex.Recently, Graph Neural Network (GNN)-based approaches have shown promising results for improved seizure prediction by modelling brain functional connectivity and inter-channel interactions.However, they employ redundant channel representations and a fixed propagation mechanism, which potentially limit their practicality.To address these potential limitations, this paper introduces a unified framework that employs an explanation-constrained topology-aware graph learning model.In this context, a novel Adaptive Depth Message Passing Graph Convolutional Network (ADMP-GCN) is introduced that models topology-conditioned message propagation to enable heterogeneous receptive fields within EEG connectivity graphs while considering channel reduction for improved efficiency and robustness.Following model training, multilevel post hoc explainability is performed using SHAP, Integrated Gradients (IG), and GNNExplainer to interpret feature, layer, and graph-level contributions to seizure prediction.Finally, a human-aligned composite supervision is introduced that balances prediction accuracy and explanation quality.The proposed model is validated using the CHB-MIT dataset, and it is found that the proposed model results in improved accuracy of 95.10% and sensitivity of 96.33% compared to existing graph-based seizure prediction models.In conclusion, the results reveal a viable and clinically relevant seizure prediction that balances performance and applicability.