Automated seizure detection is essential for effective epilepsy management, yet patient-specific models often generalize poorly across patients due to distribution shifts. Unlike methods focused solely on training-phase generalization, we propose a test-time adaptation strategy that dynamically adjusts model parameters using test samples. Built on ResNet-18, our approach incorporates a learnable consistency loss as an auxiliary objective during training, and introduces adaptive blocks for dynamic updating during testing. Our method has been extensively evaluated on two publicly available datasets, CHB-MIT and Siena, demonstrating strong performance across key metrics. On the CHB-MIT dataset, the proposed approach achieves an accuracy of 95.24%, a sensitivity of 94.69%, and a specificity of 95.85%. Similarly, on the Siena dataset, it attains an accuracy of 91.88%, a sensitivity of 92.32%, and a specificity of 91.59%. These results highlight the promise of test-time adaptation methods for cross-patient seizure detection and provide new insights for future research in this field.
Feng et al. (Wed,) studied this question.