Hierarchical multi-task learning for comprehensive gait assessment using wearable inertial sensors

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Quantitative gait analysis using wearable inertial sensors offers scalable solutions for neurological and musculoskeletal disorders.
• Current single-task learning (STL) methods are inefficient, limiting multi-objective analysis, parameter efficiency, and generalizability.
• A unified approach is needed to overcome the limitations of isolated models for diverse gait analysis tasks.

Purpose of the Study:

• To introduce H-MTL, the first hierarchical multi-task learning framework for comprehensive gait analysis.
• To unify ten heterogeneous gait analysis tasks within a single, parameter-efficient model.
• To enforce prediction coherence aligned with the clinical diagnostic tree.

Main Methods:

• Developed H-MTL, a hierarchical multi-task learning framework integrating ten gait analysis tasks.
• Utilized a hierarchy consistency loss to ensure prediction coherence based on the clinical diagnostic tree.
• Validated H-MTL through subject-wise 10-fold cross-validation on 260 subjects and cross-dataset validation on four external cohorts (456 subjects).

Main Results:

• H-MTL achieved 85.0 ± 4.1% screening accuracy with significantly fewer parameters (0.61 million) compared to STL approaches.
• Cross-dataset validation demonstrated selective transfer learning, enhancing motor-severity and neurodegenerative classification.
• The framework provides a lightweight, coherence-aware solution for comprehensive gait assessment.

Conclusions:

• H-MTL represents a significant advancement in gait analysis using wearable inertial sensors.
• The hierarchical multi-task learning approach offers improved efficiency, coherence, and generalizability.
• This framework provides a foundation for advanced, unified clinical gait assessment.