Machine Learning-Based Prediction of Aerobic Physiological Indicators in Athletes and Physically Active Individuals: A Systematic Review and Meta-Analysis

Background: Machine learning is increasingly used to predict aerobic physiological indicators in sport, but the reliability of reported performance is uncertain. Methods: We searched the PubMed, Embase, Web of Science Core Collection, Scopus, SPORTDiscus, and IEEE Xplore databases through 22 September 2026 and included 31 reports involving athletes or physically active individuals. Study information was extracted using CHARMS, and prediction-model evaluations were assessed with PROBAST+AI. Reported performance was summarized by outcome, metric, and validation design. An exploratory random-effects analysis was restricted to two lactate-threshold correlations from identifiable participant test sets. Results: Twelve reports assessed maximal or peak oxygen uptake, nine threshold-related outcomes, six continuous oxygen uptakes, two running economies, and one aerobic power. The two-study pooled correlation was 0.889 (95% confidence interval 0.090–0.992); the studies used different threshold definitions and target variables. Nine selected maximal oxygen uptake reports provided R2-related values of 0.45–0.969, but these were not pooled because definitions, evaluation designs, and sampling variances were not sufficiently comparable. No report was confirmed to provide locked-model validation in an independently recruited external cohort. The PROBAST+AI assessments classified 29 of 31 model evaluations as high risk of bias. No algorithm family was consistently superior. Conclusions: Strong within-study associations do not establish agreement or transportability. The small threshold synthesis is highly uncertain; participant-level validation, independent physiological reference measurements, and transparent reporting are priorities.