SHAP-Driven Feature Analysis Approach for Epileptic Seizure Prediction.

Abstract

Predicting epileptic seizures presents a substantial difficulty in healthcare, with considerable implications for enhancing patient outcomes and quality of life. This paper presents an explainable artificial intelligence (AI) that integrates a one-dimensional convolutional neural network (1D-CNN) with SHapley Additive exPlanations (SHAP). The approach facilitates precise and interpretable seizure prediction utilising electroencephalography (EEG) inputs. The suggested 1D-CNN model with SHAP attains superior performance, exhibiting an accuracy of 98.14% and an F1-score of 98.30% with feature-level explainability and high clinical insight using the CHB-MIT dataset. Through the computation and aggregation of SHAP values across time, we identified the most significant EEG channels, specifically "P7-O1" and "P3-O1", as essential for seizure detection. This transparency is crucial for building practitioners' trust and acceptance of the use of artificial intelligence-based solutions in the clinical domain. The technique can readily operate within portable EEG structures and hospital monitoring systems, triggering real-time alerts to patients. The outcome provides a timely intervention that could include anything from medication adjustments to responses in emergencies, preventing potential injury and improving safety. So, SHAP not only explains the model's predictions, but it also check and improve how much it relies on certain features, which makes it more reliable. Additionally, SHAP's interpretability aids physicians in understanding why the model arrived at its conclusions, increasing trust in the predictions and encouraging its extensive utilisation in diagnostic processes.