Hierarchical Multi-Task Learning for Comprehensive Gait Assessment Using Wearable Inertial Sensors

Quantitative gait analysis with wearable inertial sensors holds substantial promise for scalable screening, diagnosis, and monitoring of neurological and musculoskeletal disorders. However, existing approaches predominantly adopt a single-task learning (STL) paradigm, training separate models for each clinical objective, which limits multi-objective coverage, parameter efficiency, prediction coherence, and cross-domain generalisability. Here we present H-MTL, to our knowledge the first hierarchical multi-task learning framework to unify as many as ten heterogeneous gait analysis tasks (pathological screening, multi-granularity disease classification, severity regression, and demographic estimation) within a single 0.61-million-parameter model, enforcing prediction coherence through a hierarchy consistency loss aligned with the clinical diagnostic tree. Under subject-wise 10-fold cross-validation on 260 subjects with seven gait pathologies, H-MTL achieves \((85.0 \pm 4.1%)\) screening accuracy with approximately 14-fold fewer parameters than ten separate models. Cross-dataset validation on four external cohorts (456 subjects, 34 clinical endpoints) spanning diverse sensor configurations, signal modalities, and disease populations reveals selective transfer, whereby pretrained initialisation outperforms training from scratch on 18 of 34 endpoints (53%), with the strongest gains on continuous mobility variables (up to 28.8% mean absolute error reduction) and neurodegenerative disease classification (30.7% relative \((F_1)\) improvement). These results demonstrate that the pretrained H-MTL model provides a lightweight, ready-to-deploy foundation for comprehensive, coherent, and transferable clinical gait assessment from body-worn inertial sensors.