Deployability of machine-learning prediction models in elite and high-level team sports: a task-specific systematic...

Machine-learning models are increasingly used for match forecasting and athlete monitoring in elite and high-level team sports, but high reported performance does not necessarily indicate readiness for real-world use. This systematic review evaluated whether models were developed and tested under conditions compatible with their intended decision time. Eight databases were searched, with searches completed by 12 July 2026. Eighty reports representing 77 studies contributed 87 analysis units: 55 Core, 23 post-event or high-leakage Comparators, and nine Boundary units. All 87 units were assessed for deployable prediction validity (DPV) across six domains, and 78 Core or Comparator units were also assessed with PROBAST. Amongst Core units, 26/55 (47.3%) were rated at least Moderate, including five Moderate-to-strong and 21 Moderate units. Pre-adjudication agreement for the overall DPV classification was 85/87 (97.7%; κ = 0.968; linearly weighted κ = 0.978), whilst domain-level exact agreement ranged from 75/87 to 87/87. In post hoc sensitivity analyses, 19/55 Core units met the primary DPV threshold and had Sufficient validation; 14/55 met the primary threshold and had at least partial evidence of predictive reliability or probability-quality; and 9/55 met the primary threshold and both additional conditions. PROBAST risk of bias was High for 56/78 units, Unclear for 21/78, and Low for one. The main limitations were evaluation designs that did not adequately reflect intended future use and limited assessment of predictive reliability. Overall, the evidence supports selective methodological readiness rather than general deployability, with different requirements for pre-game, in-game, and athlete-monitoring applications.

https://www.crd.york.ac.uk/PROSPERO/view/CRD420261425179 , identifier CRD420261425179.