Hierarchical multitask learning for upper extremity rehabilitation assessment using wearable sensors

Abstract

Upper extremity motor impairment following stroke critically limits patients’ functional independence, making continuous and objective rehabilitation assessment vital for guiding recovery. Current clinical evaluations remain largely subjective, confined to clinical settings, and restricted to isolated tasks, hindering frequent and fine-grained monitoring. To address these limitations, we present a hierarchical multitask learning framework driven by wearable sensors for simultaneous upper-extremity rehabilitation assessment. The wearable system comprises three low-cost wearable units, namely two inertial measurement units and an instrumented glove, enabling continuous kinematic analysis in unsupervised daily-life environments. The framework pretrains 15 subtask-specific expert networks to capture fine-grained movement representations; a dynamic gating network then fuses these representations to concurrently predict both the Brunnstrom Recovery Stage and Fugl-Meyer Assessment for the Upper Extremity from a single data acquisition. On the self-built RehabEx dataset, evaluated under a window-level split, the model achieves a BRS classification accuracy of 0.9695 and an FMA-UE R 2 of 0.9682 with a prediction error below the minimal clinically important difference. As this split is not participant-disjoint, these figures represent internal window-level performance, and subject-wise validation is required to establish generalization to unseen patients. Ablation studies further show that dynamic gating outperforms static fusion methods within this framework. The principal novelty is threefold: (i) the first explicit hierarchical multitask formulation of post-stroke upper-extremity assessment with subtask-to-scale decomposition, enabling synchronous BRS and FMA-UE evaluation from a single data acquisition; (ii) a Mixture-of-Experts–style dynamic gating mechanism applied to wearable IMU-based rehabilitation assessment, that adapts feature integration to each clinical objective rather than relying on a task-agnostic shared representation; (iii) empirical demonstration that a pretrain-then-freeze paradigm outperforms end-to-end joint training under realistic clinical sample sizes, providing a practical engineering paradigm for multitask learning in data-sensitive medical domains.