Tibial Acceleration-Based Prediction of Maximal Vertical Loading Rate During Overground Running: A Machine Learning Approach

Rud Derie, Pieter Robberechts, Pieter Van den Berghe, Joeri Gerlo, Dirk De Clercq, Veerle Segers, Jesse DavisFrontiers in Bioengineering and Biotechnology 2020 · DOI

TL;DR

This paper introduces a machine learning approach to predict the maximal vertical instantaneous loading rate (VILR) during running using only wearable tibial accelerometers. By leveraging time-continuous 3D acceleration data and advanced ML models like XGBoost, they achieve significantly more accurate VILR predictions than traditional single-metric methods, paving the way for real-time, in-field biomechanical analysis.

What Problem Does This Paper Address?

Running biomechanics are crucial for understanding injury risk and athletic performance. A key metric is the maximal vertical instantaneous loading rate (VILR), which quantifies the impact force during foot strike. Traditionally, VILR is measured using expensive and lab-bound force platforms. While axial peak tibial acceleration (APTA) from wearable sensors offers a field-friendly alternative, it's a simplified, time-discrete measure that neglects much of the rich information available in continuous 3D acceleration data. This study aims to bridge this gap by exploring whether machine learning can extract more predictive power from this richer data to provide a more accurate, real-time VILR estimation outside the lab.

Methods

The researchers employed a machine learning pipeline that involved several key steps:

  • Data Acquisition: Bilateral tri-axial accelerometers were attached to the tibias of 93 recreational and competitive runners. Ground reaction forces were simultaneously recorded using force plates for ground truth VILR calculation.
  • Signal Preprocessing: Raw accelerometer data was filtered using a second-order band-pass Butterworth filter (0.8-45 Hz) to remove noise and gravity components. Ground reaction force data was filtered with a 60 Hz low-pass Butterworth filter.
  • Feature Engineering: This was a critical step. They extracted a comprehensive set of features from the 3D acceleration signals, including:
  • Auto-generated statistical features: Using the tsfresh Python package, they derived hundreds of statistical features (mean, max, peaks, wavelet coefficients, etc.) from the acceleration waveforms within each stride.
  • Trial-specific features: Running speed and ground contact time were included. Ground contact time was itself predicted using a deep structured recurrent neural network to handle the complex temporal relationships between gait events.
  • Subject-describing features: Body weight and shoe type (neutral, stabilization, or racing flats) were incorporated.
  • Feature Selection: A univariate feature selection method (FRESH procedure) was used to reduce the number of auto-generated features to 1662, controlling for false discovery rates.
  • Machine Learning Models: They evaluated three regression techniques: Linear Regression with Elastic Net (EN) and LASSO regularization, and Gradient Boosted Regression Trees (XGBoost).
  • Learning Settings: Two settings were explored:
  • Subject-dependent: A personalized model was trained for each individual runner using only their data.
  • Subject-independent: A single model was trained on data from all runners except one, and then evaluated on the held-out runner. This simulates real-world application where data from a new runner is unavailable during training.
  • Evaluation: Models were evaluated using Mean Absolute Error (MAE) and the coefficient of determination (R² score). Diagnostic ability was assessed using Receiver Operating Characteristic (ROC) curves.

Datasets

The study utilized data collected from 93 recreational and competitive runners. This dataset, comprising synchronized tibial acceleration and ground reaction force data, was used for both training and evaluating the machine learning models. The paper does not introduce a new benchmark dataset but rather uses existing data to develop and validate their ML approach.

Key Insights and Ideas

  • Richer Data, Better Predictions: The core insight is that by leveraging time-continuous, multi-dimensional (3D) tibial acceleration data and applying machine learning, significantly more accurate VILR predictions can be achieved compared to relying on a single, simplified metric like APTA. This highlights the power of ML in uncovering complex, non-linear relationships within biomechanical signals.
  • XGBoost Dominance: Gradient Boosted Regression Trees (XGBoost) consistently outperformed other regression techniques, suggesting its robustness in handling a large number of features and potentially non-linear relationships.
  • Subject-Specific vs. Generalizable Models: While subject-dependent models achieved the highest accuracy, the subject-independent models (especially when incorporating body weight and shoe type) demonstrated a strong balance between predictive power and practical applicability. This is crucial for real-world deployment where personalized training data might not be available.
  • Ground Contact Time Prediction is Key: The significant computational time for prediction was largely attributed to estimating ground contact time. This suggests that further optimization in this area could lead to even faster, real-time VILR estimation.
  • APTA is a Limited Proxy: The study confirms that APTA, while useful, is a crude approximation of VILR. The ML models showed a much stronger correlation with VILR than APTA alone, underscoring the limitations of the traditional approach.

Key Results

  • Superior Predictive Accuracy: The subject-dependent XGB model achieved the most accurate VILR estimates with a Mean Absolute Error (MAE) of 5.39 ± 2.04 BW·s⁻¹ and a mean absolute percentage error of 6.08%.
  • Strong Subject-Independent Performance: The subject-independent XGB model, incorporating subject-describing features, achieved an MAE of 12.41 ± 7.90 BW·s⁻¹ and a mean absolute percentage error of 11.09%. This is still a significant improvement over the APTA baseline (MAE: 21.07 ± 8.13 BW·s⁻¹).
  • Diagnostic Ability: The subject-independent XGB model demonstrated a stronger diagnostic ability (Area Under the Curve = 0.92) in identifying high-impact runners compared to the APTA model (AUC = 0.82).
  • Real-time Prediction: The prediction of VILR can be made within one foot contact (160-350 ms), making it suitable for real-time applications.
  • Feature Importance: While not explicitly detailed, the study implies that a combination of features, including those derived from 3D acceleration, is superior to single features like APTA.

Limitations and Future Work

  • Footstrike Pattern: The study primarily included rearfoot strikers. Future work should investigate how these models perform with other footstrike patterns (e.g., midfoot, forefoot) and whether models need to be adapted for them.
  • Lab Environment: Data was collected in a controlled laboratory setting. Future research should explore how performance is affected by different running surfaces and environments, as variations in surface can influence impact loading.
  • Additional Sensors: The authors suggest that incorporating trunk acceleration data could further improve predictive accuracy, as VILR relates to the deceleration of the entire body.
  • Leg Asymmetry: The study noted a significant asymmetry in VILR between left and right legs. Future work could explore building separate models for each leg to potentially improve accuracy.
  • Data Volume: While the study used 93 runners, more data could further enhance the robustness of subject-independent models.
  • Wearable Implementation: The authors envision implementing these algorithms on wearables, which would require further optimization for power consumption and computational efficiency on embedded systems.

Summary updated

Abstract

Ground reaction forces are often used by sport scientists and clinicians to analyze the mechanical risk-factors of running related injuries or athletic performance during a running analysis. An interesting ground reaction force-derived variable to track is the maximal vertical instantaneous loading rate (VILR). This impact characteristic is traditionally derived from a fixed force platform, but wearable inertial sensors nowadays might approximate its magnitude while running outside the lab. The time-discrete axial peak tibial acceleration (APTA) has been proposed as a good surrogate that can be measured using wearable accelerometers in the field. This paper explores the hypothesis that applying machine learning to time continuous data (generated from bilateral tri-axial shin mounted accelerometers) would result in a more accurate estimation of the VILR. Therefore, the purpose of this study was to evaluate the performance of accelerometer-based predictions of the VILR with various machine learning models trained on data of 93 rearfoot runners. A subject-dependent gradient boosted regression trees (XGB) model provided the most accurate estimates (mean absolute error: 5.39 ± 2.04 BW⋅s<sup>-1</sup>, mean absolute percentage error: 6.08%). A similar subject-independent model had a mean absolute error of 12.41 ± 7.90 BW⋅s<sup>-1</sup> (mean absolute percentage error: 11.09%). All of our models had a stronger correlation with the VILR than the APTA (<i>p</i> < 0.01), indicating that multiple 3D acceleration features in a learning setting showed the highest accuracy in predicting the lab-based impact loading compared to APTA.
BibTeX
@inproceedings{derie2020tibial,
  author = {Rud Derie and Pieter Robberechts and Pieter Van den Berghe and Joeri Gerlo and Dirk De Clercq and Veerle Segers and Jesse Davis},
  title = {{Tibial Acceleration-Based Prediction of Maximal Vertical Loading Rate During Overground Running: A Machine Learning Approach}},
  booktitle = {Frontiers in Bioengineering and Biotechnology},
  year = {2020},
  doi = {10.3389/fbioe.2020.00033},
}
Figure