Quality-Aware Multimodal Seizure Detection via Dynamic Motor Fusion and Autonomic Gating

Abstract

The detection of epileptic seizures is essential for effective clinical surveillance, but current algorithms are impeded by intense motion artifacts and the insufficient learning of low-frequency physiological signals. This study uses multimodal physiological signals, including three-axis acceleration (ACC), three-axis gyroscope (GYR), Pitch, Roll, surface electromyography (sEMG), and electrodermal activity (EDA), collected with wearable devices from 29 patients with epilepsy, totaling 200.9 h of recordings and 54 seizure events. To achieve epilepsy detection based on these wearable signals, we propose the quality-aware motor-autonomic gated encoding (Q-MAGE) framework. In the signal preprocessing phase, we apply intersection-based signal quality (ISQ) preprocessing to ensure global data reliability by retaining only windows that simultaneously satisfy quality thresholds across all modalities. Within the backend motor-autonomic gated encoding network (MAGE-Net), a dynamic motor fusion (DMF) module adaptively weights complementary motor features based on global context. Furthermore, to resolve modality imbalance, an EDA-gated residual (EGR) mechanism uses low-frequency autonomic embeddings as gating parameters to modulate and reinforce high-frequency motor representations via residual connections. Experimental evaluations demonstrate that Q-MAGE achieves a sensitivity of 0.9420, an $F1$ -score of 0.9420, and a false alarm rate (FAR) of 0.47/h. Notably, ISQ preprocessing alone yields an 11.9 % improvement in $F1$ -score by eliminating artifact-induced errors, while the full framework outperforms comparable state-of-the-art methods by over 30 % in sensitivity. The proposed Q-MAGE provides a reliable solution for accurate seizure detection using wrist-worn devices in noise-prone environments such as hospitals.