Analysis of Machine Learning Models for Epileptic Seizure Detection Using Wearable Sensor Data

Epilepsy is a chronic neurological disorder characterized by recurrent and unpredictable seizures, which can result in significant impairment in quality of life and may cause sudden death, particularly in patients with drug-resistant epilepsy. Consequently, the early and reliable detection of epileptic seizures is critically important for ensuring patient safety and enabling timely clinical intervention. In this study, machine learning-based methods for epileptic seizure detection using data acquired from wearable devices are comparatively evaluated. The Open Seizure Database (OSDB) was utilized to extract time-domain and frequency-domain features from raw accelerometer signals. Recursive Feature Elimination (RFE), Principal Component Analysis (PCA), and Mutual Information (MI) techniques were employed for feature selection. For the classification task, Multilayer Perceptron (MLP) and Random Forest (RF) models were implemented. Model performance was assessed using accuracy, precision, recall, and F1-score metrics, with particular emphasis placed on recall for the seizure class due to its clinical relevance. The experimental results demonstrate that the Random Forest model achieves the highest generalization performance, attaining a recall of 85\% for the seizure class on the validation dataset. Moreover, SHAP-based explainability analysis demonstrates that high-frequency energy components and statistical measures reflecting signal variability are the most influential features in the decision-making process. Overall, the results suggest that explainable machine learning approaches based on wearable sensor data offer effective and clinically interpretable solutions for epileptic seizure detection.