fastSeizureNet: Accurate and efficient knowledge-data fusion for semi-supervised seizure detection

Epilepsy is a neurological disorder affecting millions of patients worldwide. Electroencephalography (EEG) is widely used for epilepsy diagnosis due to its ability to capture both interictal and ictal patterns. For clinic usages, seizure detection should be achieved with both high accuracy and fast speed. Traditional approaches offer strong interpretability and robustness under limited labeled data, but their reliance on computationally expensive feature extraction leads to slow speed. In contrast, deep learning approaches enable fast end-to-end inference but are prone to overfitting when labeled data are limited. This paper proposes fastSeizureNet to improve both the detection accuracy and the inference speed. Three major challenges are addressed: 1) the accuracy-speed trade-off; 2) limited labeled data; and, 3) severe class imbalance. First, to mitigate the latency caused by manual feature extraction, we design lightweight neural networks to approximate the computationally expensive manually extracted features. Second, to combine the strengths of traditional and deep learning approaches, we explore multiple knowledge-data fusion strategies. Third, to alleviate class imbalance, we propose a balanced sampling strategy that encourages equitable learning across classes. Extensive experiments on two public EEG datasets were conducted under both within-patient and cross-patient scenarios with multiple backbone networks (EEGNet, TIE-EEGNet, and CE-stSENet). Further comparisons with state-of-the-art approaches consistently verified the superiority and robustness of fastSeizureNet.