Sex-Equitable Machine Learning Prediction of Ankle Reinjury in Individuals with Chronic Ankle Instability: A 2-Year Prospective Cohort Study with Fairness Audit
- 1Department of Rehabilitation Sciences, The Hong Kong Polytechnic University
- 2Graduate School of Sports Medicine, CHA University
- 3Department of Exercise Sciences, Brigham Young University
Abstract
Background Patients with chronic ankle instability (CAI) face a 2-year reinjury rate approaching 50%, but individualized prediction models that quantify which baseline impairments predict subsequent reinjury within an established CAI cohort remain scarce, and existing prediction models for ankle injury have not assessed sex-related fairness. The aim was to develop and internally validate sex-equitable machine learning (ML) models predicting 2-year ankle reinjury in individuals with CAI, identify the most influential baseline prognostic markers, and audit sex-stratified model performance. Methods One hundred physically active individuals (50 male, 50 female) with CAI (mean age 21.8 ± 2.3 years) were recruited from a university population between February 2017 and February 2018 and prospectively followed for 2 years; 96 (49 male, 47 female) with at least one follow-up response were retained for analysis (41 reinjury events, 43% reinjury rate). Fourteen baseline predictors spanning demographic, anthropometric, injury-history, patient-reported, biomechanical, postural-control, and functional-performance domains were collected. Six algorithms were trained with nested cross-validation: an unpenalized logistic regression baseline and five ML algorithms (Elastic Net, support vector machine, random forest, XGBoost, LightGBM). Predictive contribution of each feature was quantified using Shapley additive explanations and partial-dependence plots. A pre-specified fairness audit compared single global and sex-specific Youden thresholds derived out-of-fold using the equalized-odds gap and disproportionate-impact ratio. Results All algorithms achieved area under the curve (AUC) values above 0.70; Elastic Net was selected as the primary model (AUC = 0.807, 95% CI 0.714–0.893; calibration slope = 1.015). Sex, arch height index, previous sprains, dorsiflexion range of motion, and Biodex Overall Stability Index consistently emerged as the most influential prognostic markers. Within-sex discrimination was comparable between males and females, but a single global threshold produced an equalized-odds gap of 0.384 that was reduced by 57% (to 0.164) through sex-specific thresholding without retraining the model. Conclusions ML models with a fairness-audited threshold strategy predicted 2-year ankle reinjury in individuals with CAI, reducing the equalized-odds gap by 57%, though the disproportionate-impact ratio (0.683) remained below the 0.80 benchmark. External validation is required before clinical deployment.
Related
- Large-scale plasma proteomic profiling enhances prediction of heart failure among hypertensive individuals: prospective insights from the UK Biobank.
- Diagnostic accuracy of deep learning for microsatellite instability prediction in gastric cancer: A systematic review and meta-analysis
- Feedforward Pre‐Activation of Lateral Ankle Muscles Is Insufficient Against Excessive Ankle Inversion During Perturbation in Individuals With Chronic Ankle Instability
- Mortality prediction in patients with acute myocardial infarction and prior ischemic stroke: a prospective cohort study with explainable machine learning and SHAP analysis
- Development and validation of a machine learning-based prediction model for emergence agitation in geriatric patients undergoing orthopedic surgery: a prospective cohort study.
BibTeX
@article{hwang2026sexequitable,
title = {{Sex-Equitable Machine Learning Prediction of Ankle Reinjury in Individuals with Chronic Ankle Instability: A 2-Year Prospective Cohort Study with Fairness Audit}},
author = {Ui-jae Hwang and S. Jun Son and Siu-ngor Fu and Arnold Y. L. Wong and Dustin Bruening and Brent Feland and Matthew Seeley and J. Ty Hopkins},
journal = {Sports Medicine - Open},
year = {2026},
month = nov,
doi = {10.1186/s40798-026-01137-3},
url = {https://doi.org/10.1186/s40798-026-01137-3},
note = {CC BY},
}
RIS
TY - JOUR
TI - Sex-Equitable Machine Learning Prediction of Ankle Reinjury in Individuals with Chronic Ankle Instability: A 2-Year Prospective Cohort Study with Fairness Audit
AU - Ui-jae Hwang
AU - S. Jun Son
AU - Siu-ngor Fu
AU - Arnold Y. L. Wong
AU - Dustin Bruening
AU - Brent Feland
AU - Matthew Seeley
AU - J. Ty Hopkins
PY - 2026
DA - 2026/11/08
JO - Sports Medicine - Open
DO - 10.1186/s40798-026-01137-3
UR - https://doi.org/10.1186/s40798-026-01137-3
AB - Background Patients with chronic ankle instability (CAI) face a 2-year reinjury rate approaching 50%, but individualized prediction models that quantify which baseline impairments predict subsequent reinjury within an established CAI cohort remain scarce, and existing prediction models for ankle injury have not assessed sex-related fairness. The aim was to develop and internally validate sex-equitable machine learning (ML) models predicting 2-year ankle reinjury in individuals with CAI, identify the most influential baseline prognostic markers, and audit sex-stratified model performance. Methods One hundred physically active individuals (50 male, 50 female) with CAI (mean age 21.8 ± 2.3 years) were recruited from a university population between February 2017 and February 2018 and prospectively followed for 2 years; 96 (49 male, 47 female) with at least one follow-up response were retained for analysis (41 reinjury events, 43% reinjury rate). Fourteen baseline predictors spanning demographic, anthropometric, injury-history, patient-reported, biomechanical, postural-control, and functional-performance domains were collected. Six algorithms were trained with nested cross-validation: an unpenalized logistic regression baseline and five ML algorithms (Elastic Net, support vector machine, random forest, XGBoost, LightGBM). Predictive contribution of each feature was quantified using Shapley additive explanations and partial-dependence plots. A pre-specified fairness audit compared single global and sex-specific Youden thresholds derived out-of-fold using the equalized-odds gap and disproportionate-impact ratio. Results All algorithms achieved area under the curve (AUC) values above 0.70; Elastic Net was selected as the primary model (AUC = 0.807, 95% CI 0.714–0.893; calibration slope = 1.015). Sex, arch height index, previous sprains, dorsiflexion range of motion, and Biodex Overall Stability Index consistently emerged as the most influential prognostic markers. Within-sex discrimination was comparable between males and females, but a single global threshold produced an equalized-odds gap of 0.384 that was reduced by 57% (to 0.164) through sex-specific thresholding without retraining the model. Conclusions ML models with a fairness-audited threshold strategy predicted 2-year ankle reinjury in individuals with CAI, reducing the equalized-odds gap by 57%, though the disproportionate-impact ratio (0.683) remained below the 0.80 benchmark. External validation is required before clinical deployment.
SN - 2198-9761
ER -