A Real-time Cross-patient Seizure Detection Algorithm Based on Adaptive Template Matching

Performance of EEG-based seizure detection across patients is heavily influenced by individual variability. In this work, we propose an adaptive template matching method based on an end-to-end convolutional neural network. The detection results can be obtained by matching the test data with the feature templates, which are fine-grained by K-means clustering and linear discriminant analysis (LDA). As the calibration data and feature templates are continuously updated during the testing process, the detection performance is gradually improved. Experiments on the CHB-MIT dataset show that our method achieves 84.75% sensitivity, 92.31% specificity and 91.92% accuracy. Experiments on an FPGA-based embedded system show that our method has good real-time performance.