Multisensory assessment and machine learning for athlete classification in talent identification

Talent identification in elite sport is challenged by maturation confounding and limited objective assessment tools. This preliminary study examined whether visual-vestibular-somatosensory and autonomic (VVS-A) measures distinguished podium-level from entry-level divers using machine learning.

(1) Identify VVS-A features distinguishing podium-level divers from a Come and Try group using traditional statistical comparisons; (2) evaluate machine-learning models' ability to classify podium-level athletes; and (3) examine the distribution of classification probabilities using lift-curve analysis.

Cross-sectional exploratory study with machine-learning classification.

Sixty participants from an Olympic diving talent identification programme underwent VVS-A assessment. Somatosensory function was evaluated via ankle proprioception using the AMEDA device. Visual, vestibular, and autonomic functions were assessed using the Prism-Neuro Eye system. Group differences were examined using independent-sample Student's t-tests. Supervised ML models were trained on selected VVS-A measures and evaluated using cross-validation and a held-out test set.

Podium-level athletes demonstrated superior ankle proprioception (Left: p < 0.001, d = 1.57; Right: p < 0.001, d = 1.83) and visual-vestibular smooth pursuit (p = 0.001, r = 0.51). No group differences were observed for voluntary saccades or autonomic metrics. A calibrated Ridge Logistic Regression model classified podium-level athletes with high accuracy within this sample (94.4%; AUC = 0.889).

Selected VVS-A measures were associated with differences in current performance level in Olympic diving. However, the cross-sectional design, age differences between groups, and limited sample size preclude conclusions regarding predictive validity, necessitating longitudinal sport-specific validation before informing applied practice within talent identification contexts.