Artificial neural network predictive models for optimizing the training process in race walking: a longitudinal observational study

This study identified key physiological, biomechanical, and strength-related predictors of competitive performance in elite female race walkers and evaluated the effectiveness of classical and machine-learning models for individualized training optimization. Thirty nationally ranked female race walkers (25 ± 3 years) were assessed over four seasons (2021–2024). Laboratory and field tests included ergospirometry (VO 2 max, VCO2, VE (minute ventilation), respiratory exchange ratio (RER)), blood lactate (LA), heart rate (HR), gait kinematics (step length, speed), and lower-limb strength (1RM, maximal power). Temporal and seasonal dynamics were evaluated using Kruskal–Wallis and g-Fisher tests. Predictive models included multiple regression, polynomial regression, multilayer perceptron (MLP), and radial basis function (RBF) networks, developed with correlation-vector analysis (R0, R1), collinearity diagnostics, and interaction terms. The most influential predictors were HR, step length, VO 2 max, and 1RM (R0 > 0.70). RBF achieved the best predictive accuracy ( R 2 = 0.89; RMSE = 0.28), outperforming MLP ( R 2 = 0.87) and regression baselines ( R 2 = 0.61–0.81). Significant seasonal variation ( p < 0.001) underscored the value of time-dependent modeling. Conclusion: RBF neural networks offer superior performance for predicting race-walking outcomes; HR and step length are key real-time indicators, whereas VO 2 max and 1RM inform longer-term adaptation.