Study exposes flawed AI testing that inflates accuracy claims in health predictions
Researchers found that standard validation methods overstate how well AI models perform on repeated patient measurements, a common scenario in digital health and sports medicine. The findings matter for hospitals and health tech firms relying on AI tools—using the wrong testing approach can lead to deploying unreliable systems into clinical practice.
Originaltitel: Participant-aware model validation for repeated-measures data: comparative cross-validation study
<p>Background: Repeated-measures datasets are common in biomechanics and digital health, where each participant contributes multiple correlated trials. If cross-validation (CV) ignores this structure, information can leak from training to test folds, inflating performance and undermining clinical credibility.</p><p>Objective: This study evaluates the impact of participant-aware validation strategies on model reliability in repeated-measures classification tasks, using fear of reinjury prediction following anterior cruciate ligament reconstruction (ACLR) as a case study.</p><p>Methods: We analyzed 623 hop trials from 72 individuals after ACLR to classify fear of reinjury based on biomechanical features. Four CV strategies were compared: stratified 10-fold CV, leave-one-participant-out cross-validation (LOPOCV), group 3-fold CV, and a nested framework combining LOPOCV (outer loop) with group 3-fold CV (inner loop). Ten supervised classifiers were benchmarked across classification accuracy, train-test generalization gap, model ranking consistency, and computational efficiency.</p><p>Results: Stratified 10-fold CV systematically overestimated model performance (eg, extra trees accuracy of 0.91 vs 0.66 under LOPOCV) due to participant-level data leakage. Group and nested CV strategies yielded more conservative and stable estimates. The nested LOPOCV + group CV framework achieved a good balance between generalization and participant-aware separation, with reduced bias and overfitting compared with nonnested alternatives.</p><p>Conclusions: Participant-aware validation strategies are essential for trustworthy machine learning (ML) evaluation in repeated-measures settings. Nested CV designs improve reproducibility, reduce selection bias, and align with regulatory expectations for clinical ML tools. These findings support best practices in model validation for biomechanics and digital health applications.</p>