ARISE-Net enables interpretable next-day wellness-risk prediction from longitudinal athlete monitoring data

Longitudinal athlete-monitoring data provide an opportunity to anticipate short-term wellness deterioration and support proactive training-load management, but many existing sports-injury and wellness prediction studies rely on small static tabular datasets, weak leakage control, limited calibration analysis, and insufficiently specified interpretability procedures. We propose ARISE-Net (Adaptive Risk-aware Interpretable Structure-Enhanced Network), a temporal risk-representation framework for next-day athlete wellness-risk prediction. Given a 14-day trailing window of daily training-load and self-reported wellness channels, ARISE-Net combines input feature attention, a bidirectional recurrent temporal encoder, prototype-based feature grouping, dual-path local–structural fusion, and a variational risk head trained with a supervised focal objective and KL regularization. The learned temporal risk embedding is then integrated with a gradient-boosted backbone through validation-only stacking. The primary evaluation was conducted on the longitudinal SoccerMon corpus, comprising 50 elite female football players monitored over two seasons. The task was to predict next-day elevated perceived stress from 15 daily monitoring channels under a temporally separated within-cohort holdout, supplemented by a subject-grouped evaluation for unseen-athlete generalization; the positive-class prevalence was 0.258. ARISE-Net achieved the highest point estimate of PR-AUC (0.880; subject-cluster bootstrap 95% CI [0.849, 0.906]) among fifteen methods, with an F1 score of 0.832 and an MCC of 0.768. No statistically significant differences in PR-AUC or AUROC were detected between ARISE-Net and the strongest modern baselines, including FT-Transformer and CatBoost. In contrast, ARISE-Net significantly outperformed the binary persistence baseline and several conventional baselines, including logistic regression, KNN, naïve Bayes, and decision tree, under paired subject-clustered testing. A complementary athlete-grouped evaluation preserved the relative model ranking with moderately lower absolute performance (fold-mean PR-AUC 0.812 ± 0.034), indicating that the findings were not driven solely by athlete identity. Interpretability was assessed using a bounded, multi-method procedure: attention was treated as an internal localization signal and was cross-checked with SHAP values, permutation importance, temporal-attention profiles, channel-by-day attention maps, and layer-wise representation separability. These analyses localized the model’s focus to recent fatigue and stress dynamics but were not interpreted as causal evidence. Auxiliary experiments on SIRP, UFIP, and AIP were retained only as static tabular sanity checks. Because the primary evidence comes from a single elite female football cohort and a subjective wellness endpoint, external validation across sports, sexes, monitoring systems, and clinically adjudicated injury outcomes is required before broad deployment.