A Dual-Task Gait Fusion Framework for Classifying Parkinson's Disease Severity from Wearable Sensor Data

Area of Science:

• Biomedical Engineering
• Neurology
• Data Science

Background:

• Accurate Parkinson's disease (PD) severity classification is crucial for effective management.
• Sensor-based gait analysis faces challenges due to overlapping disease stages and imbalanced data.
• Developing robust models for PD assessment is an ongoing need.

Purpose of the Study:

• To develop and evaluate a novel dual-task gait fusion framework for improved PD severity classification.
• To integrate gait signals from self-selected walking and walking with turning for enhanced feature representation.
• To address class imbalance and model robustness in sensor-based PD assessment.

Main Methods:

• A cohort of 87 participants with PD (mild to moderate stages) was analyzed.
• Task-specific temporal gait representations were learned and fused.
• Delayed class re-weighting was employed for optimization.
• Performance was validated using cross-validation and an independent PhysioNet cohort.

Main Results:

• The dual-task fusion framework achieved high accuracy (90.23%) and balanced accuracy (89.61%) on the primary cohort.
• Macro F1-score (89.10%) and AUC (94.43%) indicated strong classification performance.
• The model demonstrated generalizability, achieving 84.34% accuracy on an external dataset.

Conclusions:

• Integrating turning gait with self-selected walking significantly enhances wearable sensor-based PD severity classification.
• The proposed framework shows cross-dataset robustness.
• Further validation in larger, more diverse cohorts is recommended for complete severity stage coverage.