Start with the label being learned
A supervised learning model is trained on examples paired with outcomes. In sport, those outcomes might describe medical-attention injuries, time-loss injuries or self-reported complaints. These categories overlap but are not equivalent. A model trained to identify one cannot be assumed to predict another. The prediction window also matters: a claim about the next session differs from a claim about the coming season.
Ask who recorded the outcome and how consistently it was recorded. An athlete who continues training with symptoms may be labelled differently across clubs. Access to medical care can influence which problems enter the dataset. Exposure matters too: an athlete with more playing time has more opportunity to sustain a recorded injury. Without a defensible outcome definition and follow-up process, sophisticated modelling cannot repair the basic ambiguity.
Features are selected, not discovered without context
Inputs can include movement features, previous injury, training exposure and contextual information. A simple regression model combines inputs under explicit mathematical assumptions. Tree-based methods and neural networks can learn more complex relationships, but complexity alone does not make the result more accurate, transferable or useful. Compare the model with a reasonable simple baseline rather than with no information at all.
When many features are tested against relatively few outcomes, an apparently promising pattern can reflect chance. Repeatedly tuning a model against the same evaluation set gradually turns that set into part of the development process. Feature selection, scaling and imputation must be fitted within the training procedure, not using information from the eventual test data.
Separate people, time and information carefully
Data leakage occurs when the training process receives information that would not be available at the intended prediction moment. A measurement recorded after symptoms began may quietly reveal an outcome that a supposedly preventive model is meant to anticipate. Administrative fields can also contain clues about treatment or absence that would not exist in ordinary prospective use.
Repeated observations create another trap. Randomly splitting individual sessions may place the same athlete in both development and evaluation groups. The model can then benefit from recognising an athlete rather than learning a transferable relationship. The split should reflect the intended use: new athletes, future sessions, different teams or a genuinely new setting. There is no universally correct split independent of that question.
Discrimination is not calibration
Discrimination describes how well a model separates people who later experience the defined outcome from those who do not. Calibration asks whether estimated probabilities agree with observed outcome frequencies in comparable groups. A model can rank athletes reasonably well while systematically overstating risk. An overall accuracy figure can look reassuring when the outcome is uncommon, even if the model rarely identifies useful cases.
Thresholds determine which estimates trigger a flag. Lowering a threshold may identify more eventual cases while increasing false alarms. Raising it may reduce unnecessary interventions while missing more cases. Sensitivity, specificity and predictive values help describe those trade-offs, but none selects the right action automatically. The costs of a missed concern and of an unnecessary restriction must be considered in context.
Independent validation is a new test, not a new slogan
Internal resampling can estimate how a development procedure behaves within the available data. External validation asks what happens in data collected separately, such as another club or season. Report population characteristics, missing data, uncertainty and performance across relevant groups. Equipment, coaching practice and participation patterns may shift over time, so a successful validation is not a permanent guarantee.
The TRIPOD+AI reporting guideline helps readers inspect studies of prediction models, while PROBAST+AI addresses risk of bias and applicability. The EQUATOR Network is a useful route into health-research reporting guidance. These tools improve scrutiny; a checklist completed by an author does not itself establish clinical benefit or make a product suitable for a particular team.
Association does not identify the best intervention
A feature may be associated with injury because it reflects fatigue, prior injury, exposure or another underlying factor. Changing the feature does not necessarily change injury occurrence. A model may also rely on a variable that cannot be modified responsibly. The step from “this predicts an outcome” to “changing this prevents that outcome” needs intervention evidence and an explicit account of potential harms.
The most useful model report therefore ends with boundaries: who the model was studied in, when it can be used, what information it requires and what decisions it has not been shown to improve. Prediction should support a transparent discussion, not replace one with a mysterious traffic light.
Limitation: an explanation chart is not a causal explanation
Feature-importance plots describe aspects of model behaviour. They do not establish that a feature causes injury, that an attribution is stable or that changing the feature will improve outcomes.