AI Predicted Athlete Performance From Movement, Heart Data and Surveys

Athlete dashboards often separate movement, heart data and mental readiness into different screens. A new study tested whether one model could combine all three, reporting 91.2% cross-validated accuracy when classifying performance data from 150 athletes in team and individual sports.

KEY NUMBER
The combined model reached 91.2% accuracy under tenfold cross-validation using data from 150 athletes.

Three incomplete views became one prediction

The video pipeline used MediaPipe Pose to locate body landmarks in each frame. A Transformer then analyzed how those points changed over time, producing a representation of the athlete's movement rather than relying on a single still image.

The second input covered heart rate and heart-rate variability, measurements associated with cardiovascular load and recovery. The third drew on questionnaire scores for motivation, confidence and mental toughness.

The model fused those features before producing its prediction. It also reported an F1 score of 0.908 and an R-squared value of 0.895, outperforming the long short-term memory network, conventional deep neural network and random forest baselines chosen by the authors.

Removing any input made the model weaker

In ablation tests, performance fell when the researchers removed video, physiological or psychological features. That supports the narrower conclusion that the three sources contained complementary information within this dataset.

It does not show that every team should collect all three. Questionnaires can vary with context and reporting behavior, while video and heart sensors introduce calibration, privacy and data-quality requirements. Coaches would also need to know whether an apparently accurate score changes a real decision.

NewTqnia has previously examined wearable electrodes designed to keep recording through sweat and movement and the sensor inside the 2026 World Cup ball. Those systems measure observable signals directly. This model adds an interpretation layer whose reliability depends on how the target performance label was defined.

The next test must happen outside the original dataset

A useful follow-up would train the system on one group and evaluate it prospectively on athletes from different teams and sports. Researchers should publish how often the model changes a coach's recommendation, whether those changes improve a predefined outcome and which errors could put an athlete at risk.

The 91.2% figure shows that multimodal data can fit the study's performance categories. It does not yet show that the model can predict next week's match, prevent an injury or replace the judgment of a coach who knows the athlete.