Data-Efficient Continuous Walking Fatigue Prediction using Kinect Stride Features

Technical Summary: Data-Efficient Continuous Walking Fatigue Prediction using Kinect Stride Features

Problem Statement
Continuous monitoring of walking fatigue is critical for clinical rehabilitation, sports science, and occupational safety. However, existing assessment methods face significant limitations. Traditional approaches rely on subjective self-reports (e.g., Borg Rating of Perceived Exertion) or invasive physiological measurements (e.g., sEMG), which are discontinuous, labor-intensive, and unsuitable for real-world, unobtrusive deployment. Conversely, while deep learning (DL) offers powerful predictive capabilities, its application in biomechanics is hindered by the "data bottleneck": the scarcity of large, diverse, labeled datasets required to train robust models. Collecting sufficient long-duration walking trials from large cohorts is time-consuming and physically demanding, leading to models that often overfit and lack generalizability to unseen individuals. This study addresses the need for a data-efficient, objective, and continuous method to predict fatigue trajectories in new individuals without requiring full trial data collection or extensive model retraining.

Methodology
The study proposes a novel predictive pipeline that leverages a database of previously characterized fatigue profiles to predict the continuous fatigue trajectory of a new (unseen) individual. The methodology integrates the following components:

1. Data Acquisition and Feature Extraction:

   • Data was pooled from 42 healthy young adults walking on a treadmill to fatigue.
   • A Microsoft Kinect v2 sensor captured 3D joint positions without body-worn markers.
   • Eight optimum stride-level features were extracted (hip sagittal angle, knee angles, angular velocities, step height, spine angle, and toe-out angle).
   • These features were anchored to physiologically validated references: Borg RPE and sEMG, utilizing coefficients previously optimized via the Grey Wolf Optimizer (GWO).

2. Similarity-Based Identification:

   • A 200-second initial gait signature from a new participant is used to identify similar individuals within the database.
   • Features are normalized to a common 0–100% Fatigue Progression Time (FPT) grid using Dynamic Time Warping (DTW).
   • An alignment-based similarity search computes the best temporal shift between the query and repository records, calculating a weighted Mean Squared Error (MSE) using GWO-optimized coefficients as feature weights.

3. Coefficient Transfer Strategies:
   Once similar individuals are identified, the system estimates the query participant's fatigue trajectory by transferring coefficient magnitudes from the database using three strategies:

   • Nearest Person (K=1): Directly assigns coefficients from the single most similar repository record.
   • K-Nearest Neighbors (KNN): Computes a similarity-weighted average of coefficients from the top-K most similar records.
   • Fuzzy Fusion (Proposed): A novel method that maps similarity indices to fuzzy membership values (Low, Medium, High) using Gaussian functions. It generates a smoothly blended coefficient vector by weighting the entire database based on continuous similarity, rather than relying on a discrete neighborhood.

4. Evaluation Protocol:

   • Leave-One-Subject-Out (LOSO): Rigorous cross-validation where each of the 38 participants in the main cohort was held out as a query while the remaining 37 formed the repository.
   • External Validation: The model was tested on three additional participants completely unseen during the training/repository phase.
   • Metrics: Performance was measured using Mean Absolute Error (MAE) and a derived "Fidelity Score" (100 - MAE), comparing predicted trajectories against DTW-normalized Borg RPE ground truth.

Key Contributions

• Data-Efficient Framework: The study demonstrates that a brief 200-second observation is sufficient to uniquely identify an individual's gait-fatigue signature (100% identification accuracy), enabling the system to bypass the need for full trial data collection for new users.
• Fuzzy Fusion Strategy: The introduction of a Fuzzy Fusion transfer learning method that outperforms traditional Nearest Neighbor and KNN approaches by creating a continuous, weighted blend of the entire database, thereby better handling inter-individual variability.
• Physiological Grounding: Unlike purely statistical DL models, the approach anchors predictions to GWO-optimized coefficients derived from physiologically validated features (Borg RPE and sEMG), ensuring the predicted fatigue reflects actual physiological states.
• Robust Generalization: The framework successfully predicts fatigue trajectories for unseen individuals using a modest repository size, effectively addressing the data scarcity issue in biomechanics.

Results

• Identification: The 200-second gait signature achieved 100% accuracy in identifying individuals within the database, with a statistically significant distinction between self-similarity and cross-similarity (Cohen's d = 3.22, p < 0.0001).
• LOSO Cross-Validation: The Fuzzy Fusion method significantly outperformed other strategies, achieving a mean Fidelity Score of 90.88% (MAE = 9.12%). This represented a substantial improvement over the Nearest Person method (86.05% Fidelity) and the best KNN variant (K=7, 87.06% Fidelity). Fuzzy Fusion also demonstrated greater consistency, with a lower standard deviation (2.70%) and a higher percentage of predictions falling within a 10% error margin (71.05%).
• External Validation: On three completely unseen participants, Fuzzy Fusion maintained a high Fidelity Score of 89.39% (MAE = 10.61%). This reduced the mean error by 49.15% compared to the Nearest Person approach and by 45.11% compared to KNN (K=7), confirming the model's robustness and generalizability to new populations.

Significance
The paper claims that this framework provides a practical, low-cost, and generalizable solution for continuous fatigue monitoring. By utilizing a brief initial observation to unlock personalized predictive models from a repository of similar profiles, the approach circumvents the data bottleneck inherent in deep learning. The authors assert that this methodology offers a viable route for real-time, objective physiological assessment in ecological settings, such as rehabilitation and occupational safety, without the need for invasive sensors or massive datasets. The success of the Fuzzy Fusion strategy highlights the potential of similarity-based transfer learning combined with physiologically anchored optimization to overcome the limitations of current fatigue monitoring systems.