Introduction
Musculoskeletal injuries represent a significant challenge to military readiness, with approximately 55% of active-duty soldiers affected, primarily by overuse injuries. While wearable technology offers a potential solution for monitoring physical strain, traditional Human Activity Recognition (HAR) models often fail when applied to the "wild," continuous, and noisy data generated in military environments. This paper presents WearableMil, an end-to-end framework designed to process large-scale wearable data and provide actionable insights for injury prevention.
Methodology
The WearableMil framework addresses the complexities of field data through a multi-stage pipeline designed for data alignment, imputation, and classification.
- Data Alignment: Using a Linear Truncated Model (LTM), high-frequency heart rate (HR) data and lower-frequency activity logs are synchronized onto a unified one-minute grid.
- Physiologically-Informed Imputation: To handle data gaps, a rule-based system utilizes personalized HR thresholds ($minHR$ and $maxHR$) to recover missing sleep and awake states. This method reduced unknown sleep states from 40.38% to just 3.66%.
- Hierarchical Deep Learning: A two-layer Bidirectional LSTM (Bi-LSTM) architecture performs classification. Level 1 identifies broad states (Sleep, Awake, Activity), while Level 2 classifies specific military activities (e.g., Running, Obstacle Courses).
- Hierarchical Focal Loss: A custom loss function is employed to mitigate class imbalance, ensuring the model accurately identifies rare but high-intensity activities.
Experimental Setup
The framework was validated using a dataset from 135 soldiers wearing Garmin Forerunner 55 smartwatches over six months. This resulted in over 15 million minutes of data, including heart rate, steps, and sleep cycles.
Results and Discussion
The WearableMil framework demonstrated high performance in state and activity recognition. Level 1 state classification achieved a 93.8% accuracy in temporal splits and 83.8% in user splits.
A key finding was the "Window Size Paradox." As shown in the performance metrics, while longer windows (45–60 minutes) generally improved broad state classification by providing more context, they tended to obscure short-duration, high-intensity events. The researchers identified 45 minutes as the optimal window size for balancing these needs.
Beyond classification, the system includes a "SMART" visualization tool using radar charts. By comparing individual metrics against group medians, trainers can identify soldiers experiencing disproportionate strain, serving as an early warning system for potential injuries.
Conclusion
WearableMil provides a robust solution for monitoring military personnel in the field. While the model faces challenges with activities where sensors are removed (like contact combat), it offers a significant step forward in automated monitoring. Future work will explore adaptive window sizes to capture transient events more effectively and the use of Generative Adversarial Networks (GANs) for more sophisticated data imputation.
Summary updated