Cardiac rehabilitation (CR) improves functional capacity and reduces cardiovascular morbidity, yet clinical response remains highly heterogeneous and difficult to stratify using conventional assessment. This study presents a machine-learning framework for the early stratification of CR patients into responders and non-responders based exclusively on pre-intervention baseline characteristics. A total of 122 patients undergoing an 8-week CR program were evaluated using 56 clinical, physiological and metabolic predictors. Multiple classification models were trained under a stratified 10-fold cross-validation scheme. Among them, an SVM-RBF classifier achieved the best performance and retained high discriminative capacity after dimensionality reduction. The final reduced model, based on the ten most informative features identified through convergence between Random Forest and SHAP analyses, preserved >95% of the full-feature performance. The predictors were physiologically coherent, reflecting muscular strength, ventilatory efficiency, chronotropic modulation and metabolic burden. SHAP-based explainability enabled patient-level attribution of improvement likelihood, identifying modifiable variables associated with favorable or limited training response. In parallel, we developed a web-based clinical decision-support prototype that estimates improvement probability and highlights the most influential determinants for each patient, illustrating translational applicability for precision rehabilitation planning. These findings support a transition toward personalized CR strategies guided by explainable AI and baseline phenotyping.

Cardiac rehabilitation (CR) combines supervised exercise with interventions targeting long-term health behaviours, yet outcomes are often evaluated primarily through changes in aerobic capacity. This study compared the standardised magnitude of improvements in Mediterranean-diet adherence and peak oxygen uptake (VO₂peak) and examined whether these changes tracked together within individual patients, and whether previous habitual physical activity was associated with subsequent aerobic response. A total of 112 patients completing an eight-week comprehensive CR programme were analysed. Pre–post changes in VO₂peak and PREDIMED score were quantified using paired tests and within-subject standardised effect sizes. Associations with previous physical activity were assessed using adjusted regression models, while direct correspondence between dietary and aerobic changes was examined using Pearson correlation and equivalence testing. A four-domain exploratory analysis jointly evaluated changes in VO₂peak, PREDIMED, handgrip strength, and waist circumference. VO₂peak increased by 1.80 mL · kg⁻¹ · min⁻¹ (95% CI 1.30–2.31; $d_z=0.67$), while PREDIMED increased by 1.71 points (95% CI 1.26–2.17; $d_z=0.71$). Their standardised effect-size estimates were similar, but individual changes were essentially unrelated ($r=0.003$; $p=0.976$), and equivalence testing supported that the association was contained within the pragmatic bounds of $r=\pm 0.20$. Previous physical activity was independently associated with greater VO₂peak improvement ($\beta =1.63$ mL · kg⁻¹ · min⁻¹; 95% CI 0.53–2.72; $p=0.004$). The multivariate change was significant (Hotelling’s $T^2=152.24$; $F_{4,108}=37.03$; $p<0.001$), with all four domains providing non-redundant information. These findings indicate that behavioural and physiological improvements may coexist at the group level without parallel individual-level change, supporting a multidimensional assessment of CR response.