A Novel End-to-End Multisource Data Adaptive Framework for Epileptic Seizure Detection

Epilepsy is a chronic neurological disorder characterized by recurrent seizures that severely impact patients’ lives, making seizure detection an important research focus. Nonetheless, most existing methods are developed on single datasets with limited generalizability, leading to insufficient robustness in multi-center clinical settings. To this end, a multi-source data adaptive (MSDA) framework for multi-center epileptic seizure detection was proposed in this study, in which the Multi-Scale Residual and Multi-Dimensional Feature Enhancement Network (MSR-MDFENet) is integrated with a patient-adaptive transfer learning (PATL) strategy. Specifically, three datasets with different configurations (i.e., patients number, Electroencephalography channel number, etc.) were used for evaluation: the publicly available CHB-MIT dataset and two newly collected local datasets, JHCH and JHMCHH. Cross-patient seizure detection was first performed on each dataset by employing MSR-MDFENet together with a leave-one-out cross-validation strategy; superior performance was achieved over other recent benchmark models. After further integrating the PATL strategy-forming the MSDA framework-detection accuracy on the three datasets was significantly increased by 16.17%, 6.71%, and 11.28%, reaching 97.59 ± 3.56%, 96.74 ± 3.77%, and 96.52 ± 5.90%, respectively. Moreover, satisfactory detection accuracies were achieved on the JHCH (95.43 ± 4.52%) and JHMCHH (94.77 ± 5.97%) datasets when the MSDA framework was trained on the CHB-MIT dataset, thereby underscoring the generalizability of the MSDA framework at cross-patient/device/datasets levels. In summary, the proposed MSDA framework, with its advanced generalizability and performances, provides an effective and practical approach for multi-center seizure detection.