Exploratory development and internal validation of leakage-resistant, imbalance-aware machine-learning models for incident tendinopathy in athletes
Background Tendinopathy is common in physically active populations yet difficult to predict in screening settings where cohorts are small, and outcomes are rare. Before machine learning (ML)-assisted screening can be justified, models must be evaluated with leakage-resistant methods and mapped to clinically meaningful operating points. Objective To develop and benchmark a reproducible machine-learning framework for predicting tendinopathy risk in student-athletes under small-sample, class-imbalanced conditions. Methods A prospective cohort of student-athletes ( n = 56) was modelled using baseline predictors spanning demographics/anthropometrics, neuromuscular measures, and sleep/burnout constructs. An end-to-end pipeline integrated missing-data handling, imbalance mitigation (including synthetic oversampling), hyperparameter optimisation, probability calibration, and decision-threshold selection within repeated stratified nested cross-validation (20 outer test folds). Thresholds were selected using a guarded, fixed-precision policy targeting PPV ≥ 0.60 and sensitivity ≥ 0.40 (with predefined fallback rules when constraints were infeasible). Discrimination was contextualised using permutation-based null comparisons; reliability was assessed using the Brier score and fold-wise zero-recall analysis. Results Across 58 model–imbalance configurations, tree-based ensembles were the most stable internally. Random Forest with SMOTE achieved balanced accuracy 0.630 (95% CI, 0.509–0.751), ROC-AUC 0.669 (95% CI, 0.463–0.875), and PR-AUC 0.625 (95% CI, 0.430–0.819). Random Forest with ADASYN achieved ROC-AUC 0.679 (95% CI, 0.485–0.872). Leading Random Forest configurations showed Brier scores of approximately 0.20 and 0% zero-recall folds. The corrected repeated-CV intervals were wide, reflecting substantial uncertainty under the small-sample design. Conclusions Tree-based ensembles, particularly Random Forest configurations, showed the most favourable internal-validation profile among the evaluated models. However, discrimination was modest and uncertainty was substantial; these findings should be regarded as exploratory model-development evidence and require external validation before clinical screening use.
Authors
- Rawad Hammad (ORCID: https://orcid.org/0000-0002-7900-8640)
- Ahmed Saber Elgebaly (ORCID: https://orcid.org/0000-0002-8324-0960)
- Duncan Muir (ORCID: https://orcid.org/0000-0001-5635-2337)
- Hassan S. Abdalla
- Majed Saleh Alsulaiman
- Mohamed Imam
Institutions
- Imperial College Healthcare NHS Trust (GB)
- University of East London (GB)
Publication Details
- Journal
- Frontiers in Medicine
- Published
- 2026-11-19
- DOI
- https://doi.org/10.3389/fmed.2026.1918715
- Primary Topic
- Tendon Structure and Treatment
- Type
- article
- Field-Weighted Citation Impact
- 0.00