DualDyConvNet: Dual-Stream Dynamic Convolution Network via Parameter-Efficient Fine-Tuning for Predicting Motor Prognosis in Subacute Stroke

Abstract

Stroke is a significant impediment on a global scale, with the prognosis for motor ability contingent on initial rehabilitation and the severity of the injury. Consequently, the predictability of early recovery potential for personalized rehabilitation is crucial. However, studies predicting the prognosis of motor ability are still limited in performance. In this study, we propose a novel framework, called dual-stream dynamic convolution network (DualDyConvNet), to predict motor recovery for two months using resting-state electroencephalogram data in the subacute phase. Specifically, the channel-stream emphasizes distinct characteristics within each frequency band, while the spatial-stream integrates information across frequency bands to capture spatial patterns. We utilized the SMC and KIST datasets consisting of subacute stroke patients, and recovery potential was quantified using Fugl-Meyer Assessment of upper limb. As a result, we achieved average root mean squared error (RMSE) of 0.070 +/- 0.045 and 0.223 +/- 0.148 on the two datasets, respectively. This outperformed existing models, confirming the efficacy of our framework. Moreover, external (cross-dataset) validation was conducted under two conditions of with and without Euclidean-space alignment (EA) application, and DualDyConvNet outperformed comparative models, demonstrating strong generalization: pre-trained on SMC, it achieved mean RMSEs of 0.218 +/- 0.172 (w/o EA) and 0.215 +/- 0.201 (w/ EA); pre-trained on KIST, 0.160 +/- 0.087 (w/o EA) and 0.135 +/- 0.087 (w/ EA). The proposed framework holds significant potential in facilitating early rehabilitation planning by predicting motor function prognosis in stroke patients. Furthermore, it can contribute to enhancing the quality of life by providing patients with prognostication.