TL;DR
This paper presents a novel deep learning approach using a single shoe-mounted accelerometer to accurately estimate lower extremity joint angles during running. This method significantly simplifies gait analysis by reducing sensor requirements and eliminating the need for participant-specific calibration, paving the way for more accessible and unobtrusive injury prevention tools.
What Problem Does This Paper Address?
Running injuries are a significant concern, often linked to abnormal running kinematics. Accurately measuring these kinematics is crucial for identifying potential issues and preventing injuries. Traditional methods, like optical motion capture, are accurate but confined to laboratory settings and require complex setups. Inertial Measurement Units (IMUs) offer a more portable solution, but existing IMU-based methods often demand multiple sensors, intricate calibration procedures, or participant-specific tuning, making them cumbersome for widespread use, especially for recreational runners. This paper aims to bridge this gap by developing a highly simplified, yet accurate, system for gait analysis using just one accelerometer.
Methods
The core of this research lies in a deep learning approach, specifically a one-dimensional Convolutional Neural Network (1D-CNN). The CNN was trained to map raw accelerometer data from a single shoe-mounted sensor to lower extremity joint angles (hip, knee, and ankle) in the sagittal plane.
- Sensor: A single accelerometer, integrated into a shoe-mounted IMU, was used. Crucially, only the accelerometer data was utilized, avoiding the complexities and potential interference associated with gyroscopes and magnetometers.
- Data Input: The CNN processed time-windowed accelerometer data, including past and future samples, along with a calculated root sum of squares of acceleration across the three axes. This latter feature was chosen for its robustness to IMU orientation changes.
- CNN Architecture: A custom 1D-CNN architecture was designed with multiple convolutional and pooling layers for feature extraction, followed by dense layers for regression. A dropout layer was incorporated for inter-participant scenarios to enhance generalization.
- Custom Loss Function: To address the higher estimation errors observed for hip and ankle angles compared to the knee, a customized loss function was implemented. This function weighted the errors for hip and ankle more heavily, aiming to improve their accuracy.
- Evaluation: The model was evaluated using two distinct approaches:
- Intra-participant: Models were trained and tested on data from the same participant, assessing the model's ability to learn individual gait patterns.
- Inter-participant: A leave-one-out cross-validation scheme was employed, where models were trained on data from nine participants and tested on the tenth, evaluating the model's generalizability across different individuals.
Datasets
The study utilized a dataset collected from ten healthy male participants running on a split-belt treadmill at five different speeds (8-12 km/h).
- Reference Data: Gold-standard joint angles were captured using a six-camera optical motion capture system (Vicon).
- Accelerometer Data: Raw accelerometer data was collected simultaneously from an Xsens IMU mounted on the shoe.
- Data Split: For intra-participant evaluation, data was split into 80% for training and 20% for testing. For inter-participant evaluation, a leave-one-out cross-validation approach was used.
Key Insights and Ideas
- Single Sensor Efficacy: The most striking insight is that a single accelerometer, strategically placed on the shoe, can provide surprisingly accurate estimations of lower extremity joint kinematics during running. This dramatically simplifies the sensor setup compared to previous multi-IMU approaches.
- Deep Learning for Generalization: The use of a 1D-CNN proved effective in learning complex patterns from accelerometer data and, importantly, demonstrated a degree of generalization across different participants, even without explicit participant-specific calibration.
- Shoe-Mounting Advantage: The choice of a shoe-mounted sensor is highlighted as a practical and convenient position for runners, avoiding the issues of improper fixation and soft tissue artifacts that can plague sensors placed on the shank or thigh.
- Gait Event Accuracy: Beyond continuous angle estimation, the model showed promising accuracy in identifying key gait events, which are critical for biomechanical analysis and injury risk assessment.
Key Results
The study reported impressive accuracy metrics for both intra- and inter-participant scenarios:
- Intra-participant:
- Root Mean Squared Error (RMSE) for hip, knee, and ankle angles was less than 3.4°.
- Normalized RMSE (NRMSE) was below 4.6%.
- R² values were consistently above 0.97, indicating a high degree of accuracy.
- Inter-participant:
- RMSE for hip, knee, and ankle angles was less than 6.5°.
- NRMSE was below 11.1%.
- R² values averaged 0.84 for the hip, 0.93 for the knee, and 0.78 for the ankle.
- Gait Events: Mean Absolute Error (MAE) for six key gait events was less than 2.5° in intra-participant and 6.5° in inter-participant scenarios. Notably, the ankle angle at initial contact was estimated with an MAE of less than 3° in the inter-participant method.
Limitations and Future Work
- Generalization: The authors acknowledge that data-driven methods, including their CNN, face limitations in generalization, especially with smaller datasets. While dropout layers were used, the inter-participant accuracy was lower than intra-participant accuracy.
- Sensor Placement Optimization: The study used an empirically determined shoe-mounting position. Future work could explore optimizing this placement (e.g., heel vs. insole) or investigating the potential of multiple tiny accelerometers on the shoe.
- Outdoor and Long-Term Use: The current study was conducted in a controlled laboratory setting. Future research should investigate the model's performance and repeatability in real-world outdoor running conditions and over extended periods to assess its robustness to factors like fatigue and varying terrain.
- Other Planes of Motion: The current model focuses on sagittal plane kinematics. Extending the approach to estimate frontal and transverse plane movements would provide a more comprehensive gait analysis.
Summary updated