Wearable fall risk assessment by discriminating recessive weak foot individual

Sensor-based technologies have been widely used in fall risk assessment. To enhance the model's robustness and reliability, it is crucial to analyze and discuss the factors contributing to the misclassification of certain individuals, enabling purposeful and interpretable refinement.

This study identified an abnormal gait pattern termed "Recessive weak foot (RWF)," characterized by a discontinuous high-risk gait on the weak foot side, observed through weak foot feature space. This condition negatively affected the training and performance of fall risk assessment models. To address this, we proposed a trainable threshold method to discriminate individuals with this pattern, thereby enhancing the model's generalization performance. We conducted feasibility and ablation studies on two self-established datasets and tested the compatibility on two published gait-related Parkinson's disease (PD) datasets.

Guided by a customized index and the optimized adaptive thresholds, our method effectively screened out the RWF individuals. Specifically, after fine adaptation, the individual-specific models could achieve accuracies of 87.5% and 73.6% on an enhanced dataset. Compared to the baseline, the proposed two-stage model demonstrated improved performance, with an accuracy of 85.4% and sensitivity of 87.5%. In PD dataset, our method mitigated potential overfitting from low feature dimensions, increasing accuracy by 4.7%.

Our results indicate the proposed method enhanced model generalization by allowing the model to account for individual differences in gait patterns and served as an effective tool for quality control, helping to reduce misdiagnosis. The identification of the RWF gait pattern prompted connections to related studies and theories, suggesting avenues for further research. Future investigations are needed to further explore the implications of this gait pattern and verify the method's compatibility.