Systematic Evaluation of Deep Learning Architectures for Epilepsy Detection with Patient-Based Validation

Abstract

EEG signal processing is crucial for diagnosing neurological diseases, such as epilepsy. This
study systematically compares the performance of time-domain and time-frequency domain
representations in detecting epilepsy seizures. The data were divided into four-second windows at a
sampling frequency of 256 Hz, and data leakage was prevented by applying a patient-based
discrimination protocol. The training, validation, and test sets were separated by patient IDs, and
windows from the same patient were not included in more than one partition. Six deep learning models,
including BiLSTM, 1D-CNN, STFT-CNN, STFT-CNN with attention mechanism, ResSTFT-CNN, and
ResSTFT-CNN with attention mechanism, were evaluated under the same training conditions. AdamW
optimization, label smoothing, and early stopping algorithms were applied to all models. The models
were compared in terms of accuracy, F1-Macro, AUROC, AUPRC, calibration metrics, and inference
rate. Results showed that the 1D-CNN model exhibited significant superiority. The model offered the
most balanced performance with an accuracy of 0.957, an AUROC of 0.993, and an inference time per
sample of 0.09 milliseconds. BiLSTM came in second with 0.936 accuracy. For STFT-based models, the
addition of an attention mechanism significantly improved performance, with STFT-CNN accuracy
increasing from 0.78 to 0.91. Model complexity analyses revealed that 1D-CNN strikes a balance
between high performance and a low parameter count. Working in the time-domain representation, 1D
CNN offers an optimal solution for epilepsy detection, striking a balance between accuracy, speed, and
clinical applicability. The patient-based discrimination strategy allows for the realistic generalization
capacity of the models to be evaluated.