PEAK VERTICAL GROUND REACTION FORCE PREDICTION FROM KINEMATICS IN MALE RUNNERS - USING MACHINE LEARNING ALGORITHMS

Aug 3, 2026

This study evaluates whether peak vertical ground reaction forces during treadmill running can be predicted from lower-limb kinematics using machine learning. Eighteen healthy male runners provided sagittal hip, knee, and ankle angles plus subject metadata. Random forest, support vector regression, and multi-layer perceptron regressors were trained on steps sampled across three treadmill speeds and normalized across stance. Random forest achieved the highest accuracy (correlation 0.950, RMSE 0.456), outperforming other models, demonstrating that kinematic inputs can estimate peak VGRFs.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages4

Reading Time10 min

AccessFree to download

Introduction

Methods

Dataset and preprocessing

Model training and parameter tuning

Results

Conclusion

PEAK VERTICAL GROUND REACTION FORCE PREDICTION FROM KINEMATICS IN MALE RUNNERS USING MACHINE LEARNING ALGORITHMS Cody Dziuk and Janelle A. Cross Department of Orthopaedic Surgery, Medical College of Wisconsin, Milwaukee, WI, USA The purpose of this study was to examine if peak vertical ground reaction forces during treadmill running can be predicted from kinematic input using machine learning models. Eighteen healthy male runners' hip, knee, and ankle sagittal angles, with subject metadata, were input into random forest, support vector, and multi-layer perceptron regressors. Thirty strides per side at three speeds were pulled for the dataset. Random forest performed the best with a correlation coefficient of 0.950 and a root mean squared error of 0.456, while multi-layer perceptron was the worst with values of 0.948 and 0.462 respectively. The study showed machine learning models can predict peak vertical ground reaction forces. KEYWORDS: running, gait analysis, support vector, artificial neural network, random forest INTRODUCTION: Vertical ground reaction force (VGRF) is an important metric of running analyses that provides insight into running mechanics. VGRF characteristics have been shown to change in response to a variety of different factors including sex, fatigue, and footwear (BazueloRuiz et al. , 2018, Logan et al. , 2010) . Also, excessive ground reaction forces and knee joint loads have been identified as potential risk factors to the occurrence of running-related injuries (Messier et al. , 2008) . In the laboratory, ground reaction forces are measured via force plates or force transducers. This isn’t always practical depending on the situation. Kinematic variables are easier to measure than kinetics, with current technology providing measurements of sagittal kinematics using a phone or tablet (Mousavi et al. , 2020) . If such an app could be validated against optical three-dimensional (3D) motion capture to provide accurate kinematic data, a method of VGRF prediction based on kinematic input could be established for runners. A prediction model using a validated gold-standard data collection method needs to be established first. Prior studies have attempted to predict VGRFs by means of neural network models trained on accelerometer data (Ngoh et al. , 2018), 3D kinematic data with stacked machine learning models (Ong et al. , 2020), neural network models trained on inertial sensor data (Wouda et al. , 2018), and various other implementations. These methods require the use of complex electronic equipment. The goal of this study was to determine if peak VGRFs during stance phase of running gait can be predicted based on discrete lower body sagittal kinematics combined with subject metadata. METHODS: A public dataset of running C3D files from the Laboratory of Biomechanics and Motor Control was analyzed for this study (Fukuchi et al. , 2017) . Eighteen healthy male runners (age: 34 ± 6 years, mass: 70.0 ± 7.4 kg, height: 174.8 ± 6.8 cm) were selected. Trials consisted of treadmill running recorded at three running speeds (2 .5, 3.5, 4.5 m/s) . Kinematics were collected at 150 Hz using a 3D optical motion capture system. Kinetics were collected via a dual-belt forceinstrumented treadmill at 300 Hz. FS and FO events were marked using Visual 3D software (CMotion, Germantown, MD), then kinematic and kinetic data for each step normalized to 101 points of stance in order to compare left versus right sides. Ground reaction forces were normalized to bodyweight (N/kg) . Lower body kinematic input included hip, knee, and ankle sagittal plane angles at foot strike (FS) and foot off (FO) . Subject metadata consisted of height, weight, and running Published by NMU Commons, 2023 1 speed. Support vector regression (SVR), an artificial neural network (ANN), and random forest (RF) were the chosen prediction algorithms. Statistical parametric mapping (SPM) was performed on the stance phase sagitta

FAQs

What does the study predict in running biomechanics?

It predicts peak vertical ground reaction forces (VGRFs) during the stance phase of treadmill running.

What input variables are used for the machine learning models?

The models use sagittal hip, knee, and ankle angles at foot strike and foot off, combined with subject metadata such as height, weight, and running speed.

Which machine learning algorithm performed best for peak VGRF prediction?

Random forest performed best, with a reported correlation coefficient of 0.950 and a root mean squared error of 0.456.

Aug 3, 2026

Upload by Lucas Martin