Predicting normative walking biomechanics across the lifespan using seven simple features

Bernard X. W. Liew, Rachel Senden, David Rügamer, Emanuel Sommer, Kenneth Meijer, Qichang Mei, Richard J. Foster, Matthew TaylorJournal of NeuroEngineering and Rehabilitation 2025 · DOI

Abstract

The assessment of gait impairments requires a normative reference for comparison. For a fair assessment, comparisons must be made against a reference standing after controlling for sex, anthropometry, and walking characteristics. This study aimed to develop statistical models that predict the lower-limb kinematics and kinetics of walking across the lifespan of healthy participants, using seven simple covariates. Sixteen statistical models predicted 16 joint kinematics and kinetics during walking using the covariates of sex, age, height, mass, side (laterality), walking speed, and cadence, which were developed based on 301 participants between three to 91 years old. The root mean squared error (RMSE) ranged from 4.71° to 7.97° for joint angles, within 0.07 N/kg for ground reaction forces, 0.09 to 0.15 Nm/kg for joint moments, and 0.33 to 0.39 W/kg for joint powers. We provide both online and local apps which can be easily used by clinicians and scientists to generate normative walking data with uncertainty values, which can be used for movement impairment analysis ( https://github.com/EmanuelSommer/ShinyFOSR ).

Introduction

In clinical settings, identifying a gait impairment requires a baseline: what does a healthy person of a specific age, size, and walking speed look like? Traditionally, clinicians use gait indices to quantify deviations from the norm. However, these indices often fail to account for the natural changes in walking mechanics that occur as individuals age or change their walking speed. Most clinics lack the resources to collect extensive normative data from healthy volunteers across all demographics. Consequently, there is a significant need for a "plug-and-play" reference model that can generate a personalized healthy baseline for any patient based on simple clinical characteristics.

Methods and Materials

The study utilized a combined dataset of 301 healthy participants ranging from 3 to 91 years old. Data was aggregated from two previous studies, including 55 typically developing children and 246 healthy adults. 3D gait analysis was performed using a 12-camera Vicon system and an instrumented split-belt treadmill (CAREN system).

The researchers employed Function-on-Scalar Regression, specifically Functional Additive Models (FAM). Unlike standard linear regression, which predicts discrete values, these models predict entire continuous curves representing the gait cycle (0% to 100%). The models utilize seven simple covariates: sex, age, height, mass, side, walking speed, and cadence.

Distribution of body mass across age categories in the study population.

The FAM approach allows for non-linear "smooth" effects, capturing the complex relationships between age and walking mechanics that linear models might miss. For example, the maturation of gait in children and the decline in older adults are non-linear processes.

Results

The models predicted 16 different kinematic and kinetic trajectories with high accuracy. Sagittal plane movements (flexion/extension) were the most accurate, with Root Mean Squared Errors (RMSE) for joint angles ranging from 4.71° to 7.97°. Ground Reaction Forces (GRF) showed extremely high accuracy, with errors within 0.07 N/kg. Joint moments were predicted within 0.09 to 0.15 Nm/kg, and joint powers had an RMSE of 0.33 to 0.39 W/kg.

Variation in walking cadence across different age categories.

The study confirmed that gait changes non-linearly with age. For instance, hip extension and ankle plantarflexion show distinct peaks and troughs in behavior at different life stages. Notably, these models achieved error rates comparable to complex machine learning models that require high-tech sensor inputs, despite using only seven simple features. The models for sagittal plane angles fell near or within the "minimal detectable change" threshold of 5°, making them highly viable for clinical use.

Discussion and Conclusion

The developed models provide a robust, accessible tool for clinical gait analysis. By using a train-test split (80/20) for validation, the authors offer a realistic estimate of model performance on new patients. To facilitate clinical adoption, the researchers released an open-source Shiny App that generates normative curves with 95% confidence intervals based on patient statistics.

While the models are highly effective, some limitations remain. The data was collected on a treadmill, which can differ slightly from overground walking. Additionally, the models do not yet account for complex 3D rotations or internal bone geometry. Future work may integrate these statistical models with computer vision tools to enable full gait analysis using only a smartphone.

Summary updated

Abstract

The assessment of gait impairments requires a normative reference for comparison. For a fair assessment, comparisons must be made against a reference standing after controlling for sex, anthropometry, and walking characteristics. This study aimed to develop statistical models that predict the lower-limb kinematics and kinetics of walking across the lifespan of healthy participants, using seven simple covariates. Sixteen statistical models predicted 16 joint kinematics and kinetics during walking using the covariates of sex, age, height, mass, side (laterality), walking speed, and cadence, which were developed based on 301 participants between three to 91 years old. The root mean squared error (RMSE) ranged from 4.71° to 7.97° for joint angles, within 0.07 N/kg for ground reaction forces, 0.09 to 0.15 Nm/kg for joint moments, and 0.33 to 0.39 W/kg for joint powers. We provide both online and local apps which can be easily used by clinicians and scientists to generate normative walking data with uncertainty values, which can be used for movement impairment analysis ( https://github.com/EmanuelSommer/ShinyFOSR ) .
BibTeX
@article{liew2025predicting,
  author = {Bernard X. W. Liew and Rachel Senden and David Rügamer and Emanuel Sommer and Kenneth Meijer and Qichang Mei and Richard J. Foster and Matthew Taylor},
  title = {{Predicting normative walking biomechanics across the lifespan using seven simple features}},
  journal = {Journal of NeuroEngineering and Rehabilitation},
  year = {2025},
  doi = {10.1186/s12984-025-01768-9},
}
Figure