Adaptive ensemble sizing with reinforcement learning for real-time ankle injury detection in wearable sensor systems

TL;DR

This paper presents a real-time ankle injury detection system using wearable sensors and machine learning, achieving high accuracy and efficiency.

Contribution

The novel use of reinforcement learning to adaptively control an ensemble model for real-time injury detection in wearable systems.

Findings

The system achieved 87.4% overall detection accuracy with a 12.1% false alarm rate.

It predicted 76.3% of injury events at least 150 ms in advance with 17.2 ms latency.

The system reduced energy consumption by 35.4% and memory usage by 27.7% compared to baselines.

Abstract

Ankle injuries represent a leading cause of long-term impairment for athletes. Wearable inertial sensors have emerged for continuous joint monitoring yet implementing accurate real-time injury detection remains a challenge due to the latency, energy, and computational limitations. Effective solutions must therefore support fast, adaptive, and energy-efficient inference without compromising clinical relevance.

We implemented an adaptive ankle injury detection framework using the Ankle Motion Kinematics Dataset (AMKD), which synchronized inertial sensor and video-labeled data from 87 athletes across 12 sports. The system integrates a quantized 1D convolutional neural network (1D-CNN) and a pruned long short-term memory (LSTM) model into a lightweight ensemble. A reinforcement learning (RL) agent dynamically adjusts model parameters based on motion context, informed by a Gaussian process predictor that anticipates future kinematic shifts.

The core ensemble model achieved 94.3% classification accuracy on the test set. The full adaptive system, operating under real-time constraints, achieved 87.4% overall detection accuracy and a 12.1% false alarm rate (p < 0.01). It predicted 76.3% of injury events at least 150 ms in advance and maintained a low latency of 17.2 ms 34.8% faster than the best-performing baseline while reducing energy consumption by 35.4% and memory usage by 27.7%. The adaptive controller proactively detected 82.6% of high-risk transitions, and real-world deployment yielded 98.7% uptime across 8-hour sessions, confirming practical viability.

These results validate the framework as a viable, low-latency solution for real-time ankle injury detection in sports medicine and rehabilitation settings. Its modularity and efficiency enable seamless integration into existing wearable pipelines while maintaining responsiveness in dynamic conditions.

Figures