ExplainableAIinSportsManagement:FromtheLabtotheFederation

Every few months a new headline promises that artificial intelligence will revolutionize sport — predicting injuries before they happen, optimizing training loads, spotting the next elite athlete from a spreadsheet. Having spent the last years both researching explainable AI for human locomotion and leading the digital transformation of a national federation, I’ve learned that the gap between that promise and daily practice has very little to do with model accuracy. It has to do with trust.

The real bottleneck isn’t the model

In a research setting it’s tempting to treat a problem as solved once the metrics look good. In a federation, a sports medicine department, or a coaching staff, a model that nobody understands is a model that nobody uses. The physiotherapist who has treated an athlete for three seasons is not going to change a return-to-play decision because a black box printed a number.

This is the uncomfortable truth of applied sports AI: the limiting factor is almost never the algorithm. It’s whether the people responsible for an athlete’s health and performance can understand, question, and stand behind what the system suggests.

What sports management actually needs from AI

Strip away the hype and the real needs are refreshingly concrete:

• Injury prevention that explains itself. Not “risk: 0.72”, but which movement patterns, loads, or asymmetries are driving that risk — so a clinician can act on them.
• Monitoring that fits real workflows. Coaches, physios, nutritionists, and psychologists already work as a team. AI has to slot into that multidisciplinary conversation, not replace it with a dashboard nobody trusts.
• Decisions that survive scrutiny. In a federation, a call about an athlete can be reviewed by medical staff, technical directors, and sometimes the athlete’s own family. Every recommendation needs a why.

None of these are model-accuracy problems. They’re explainability and integration problems.

Explainability is the bridge

This is exactly where my doctoral research meets the field. Techniques like SHAP and counterfactual explanations turn a prediction into something a professional can reason about: these are the features that pushed the risk up; change this range of values and the prediction flips. Applied to clinical gait and foot-pressure data — baropodometry, optical motion capture, OptoGait, inertial sensors — that means a model can point to the specific biomechanical signature behind a decision, not just the decision itself.

When a physiotherapist can see that a flagged athlete’s risk is driven by an asymmetry the model learned to associate with a known pathology, the AI stops being an oracle and becomes a colleague. That shift — from trust me to here’s my reasoning — is what actually gets tools adopted.

From data to the federation

The other half of the story is organizational. Good models die in production when the surrounding structure isn’t ready for them. Digitizing a national judo federation through NextGenerationEU and CSD-funded programs taught me that the unglamorous work is what makes the intelligent work possible:

• Unifying performance, medical, nutrition, and psychology data that used to live in separate spreadsheets and separate heads.
• Building platforms multidisciplinary teams actually want to open.
• Establishing governance so that sensitive athlete data is handled responsibly from day one.

Only once that foundation exists does explainable AI have somewhere to land. You cannot bolt interpretability onto chaos.

Leakage-free by default

There’s a quieter lesson that research brings to sport: honesty about what a model really knows. A lot of impressive-looking sports AI collapses under rigorous, leakage-free validation — the same athlete appearing in training and test data, or a feature that quietly encodes the outcome. In a lab that’s a reproducibility problem. On the field it’s worse: a confident, wrong recommendation about a real person’s body.

Bringing the discipline of leakage-free pipelines and segmented notions of “normal” into applied settings is not academic pedantry. It’s the difference between a tool that helps and one that erodes the very trust it needs to be useful.

Where this goes next

The most interesting frontier in sports management isn’t a bigger model. It’s tighter loops between research and practice: methods that are interpretable by design, validation that is honest by default, and platforms built so that coaches, clinicians, and athletes are part of the reasoning rather than recipients of a verdict.

I sit at that intersection on purpose — as a researcher who insists on explainability and rigor, and as someone who has to make it work inside a real institution. Sport is a demanding place to deploy AI precisely because the stakes are human and the experts are skeptical. That’s not a bug. It’s exactly the pressure that produces AI worth trusting.