Machine learning-based fall risk prediction in community-dwelling older adults using dual-task spatiotemporal gait parameters
4 affiliations
- 1School of Sport Medicine and Physical Therapy, Beijing Sport University, Beijing, 100084 China
- 2BDA New Town School of The High School Affiliated to Renmin University of China, Beijing, 100176 China
- 3China Institute of Sport and Health Science, Beijing Sport University, Beijing, 100084 China
- 4Medical examination center, Third Hospital, Peking University, Beijing, 100089 China
Abstract
Background: With global aging, falls have become a major threat to geriatric health. Traditional clinical assessments often suffer from a “ceiling effect,” failing to detect subtle motor impairments in high-functioning older adults. Moreover, steady-state gait analysis lacks ecological validity in capturing dynamic stability during complex activities. To address these gaps, this study utilizes a dual-task paradigm and machine learning to develop a multidimensional, high-fidelity fall-risk classification framework.
Methods: A retrospective case-control study was conducted involving 209 community-dwelling older adults (50 fallers and 159 non-fallers). A multi-dimensional feature set was established, comprising demographics, anthropometrics, physical activity, and a battery of clinical functional assessments. An optical motion capture system was utilized to extract spatiotemporal gait parameters during a dual-task paradigm involving surface perturbations and a modified Stroop task. Eight machine learning algorithms, including Naive Bayes (NB), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost), were developed and validated to identify fall risk.
Results: The machine learning models demonstrated robust and competitive performance, with the top models achieving high discriminatory accuracy (up to 0.879) and excellent area under the curve (AUC up to 0.927) on the test set. Ten key predictors were identified, with dual-task gait variability—specifically the CV of cycle time (SHAP value = 0.083)—contributing most significantly to the screening framework. While traditional functional scales showed a “ceiling effect” with limited discriminative power, dual-task gait parameters effectively unmasked latent motor instabilities. SHAP analysis further demonstrated that the NB model provided highly interpretable, continuous probability scores for individual fall risk screening.
Conclusions: Combining dual-task gait variability with clinical functional assessments via machine learning algorithms provides a robust and interpretable framework for faller identification. The integration of optical gait analysis and machine learning facilitates early risk screening and personalized retrospective evaluation, offering a promising tool to assist in detecting older adults at risk for fall-related injuries in the aging population.
Supplementary Information: The online version contains supplementary material available at https://doi.org/10.1186/s12877-026-07998-3.
Related
- Association of Serum GDF-15 with Bone and Muscle Parameters in Community-Dwelling Individuals
- Interpretable prediction model for severe community-acquired pneumonia in adults: machine learning development and validation using routine laboratory parameters
- The role of sarcopenia in all-cause mortality prediction among community-dwelling adults with eGFR 15–89 mL/min/1.73 m2
- Analysis of barriers to self-monitoring of blood glucose among community-dwelling elderly patients with diabetes mellitus: a cross-sectional study integrating artificial intelligence-based risk prediction.
- Analysis and Prediction of Relapse Risk Factors in Community-Dwelling Individuals With Schizophrenia: Retrospective Cohort Study.
BibTeX
@article{dong2026machine,
title = {{Machine learning-based fall risk prediction in community-dwelling older adults using dual-task spatiotemporal gait parameters}},
author = {Gengxin Dong and Haizhou Hu and Yuxin Guo and Qi Gao and Dapeng Bao},
journal = {BMC Geriatrics},
year = {2026},
month = jul,
doi = {10.1186/s12877-026-07998-3},
url = {https://doi.org/10.1186/s12877-026-07998-3},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - Machine learning-based fall risk prediction in community-dwelling older adults using dual-task spatiotemporal gait parameters
AU - Gengxin Dong
AU - Haizhou Hu
AU - Yuxin Guo
AU - Qi Gao
AU - Dapeng Bao
PY - 2026
DA - 2026/07/31
JO - BMC Geriatrics
DO - 10.1186/s12877-026-07998-3
UR - https://doi.org/10.1186/s12877-026-07998-3
AB - Background: With global aging, falls have become a major threat to geriatric health. Traditional clinical assessments often suffer from a “ceiling effect,” failing to detect subtle motor impairments in high-functioning older adults. Moreover, steady-state gait analysis lacks ecological validity in capturing dynamic stability during complex activities. To address these gaps, this study utilizes a dual-task paradigm and machine learning to develop a multidimensional, high-fidelity fall-risk classification framework. Methods: A retrospective case-control study was conducted involving 209 community-dwelling older adults (50 fallers and 159 non-fallers). A multi-dimensional feature set was established, comprising demographics, anthropometrics, physical activity, and a battery of clinical functional assessments. An optical motion capture system was utilized to extract spatiotemporal gait parameters during a dual-task paradigm involving surface perturbations and a modified Stroop task. Eight machine learning algorithms, including Naive Bayes (NB), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost), were developed and validated to identify fall risk. Results: The machine learning models demonstrated robust and competitive performance, with the top models achieving high discriminatory accuracy (up to 0.879) and excellent area under the curve (AUC up to 0.927) on the test set. Ten key predictors were identified, with dual-task gait variability—specifically the CV of cycle time (SHAP value = 0.083)—contributing most significantly to the screening framework. While traditional functional scales showed a “ceiling effect” with limited discriminative power, dual-task gait parameters effectively unmasked latent motor instabilities. SHAP analysis further demonstrated that the NB model provided highly interpretable, continuous probability scores for individual fall risk screening. Conclusions: Combining dual-task gait variability with clinical functional assessments via machine learning algorithms provides a robust and interpretable framework for faller identification. The integration of optical gait analysis and machine learning facilitates early risk screening and personalized retrospective evaluation, offering a promising tool to assist in detecting older adults at risk for fall-related injuries in the aging population. Supplementary Information: The online version contains supplementary material available at https://doi.org/10.1186/s12877-026-07998-3.
SN - 1471-2318
ER -