Stage-specific predictors of exercise adherence after percutaneous coronary intervention: a prospective multicenter...

The efficacy of exercise after Percutaneous Coronary Intervention (PCI) is compromised by poor long-term adherence. Although multiple influencing factors are recognized, their dynamic evolution and relative importance across different stages remain unclear, hindering the development of targeted interventions.

To characterize longitudinal patterns of exercise adherence and identify stage-specific predictors within 6 months post-PCI using machine learning (ML).

A prospective longitudinal study.

600 post-PCI patients were assessed at baseline (T0), 1 month (T1), 3 months (T2), and 6 months (T3) after discharge. We developed concurrent-model and longitudinal prediction models using Random Forest, LightGBM, and Lasso Regression, with interpretability enabled by feature importance analysis and Partial Dependence Plots. A tiered benchmarking strategy, including a linear model trained on the full feature set, was employed to isolate the added value of ML algorithms.

Exercise adherence showed significant longitudinal changes, remaining stable from T0 to T1, sharply declining at T2, and partially recovering at T3. All ML models performed well, with concurrent-model R² values ranging from 0.503 to 0.745 and longitudinal R² values ranging from 0.5 to 0.723, and substantially outperforming linear benchmarks. Feature importance analysis revealed a dynamic evolution of stage-specific predictors: action self-efficacy and kinesiophobia dominated the initiation phase (T0), while maintenance self-efficacy and coping planning became the paramount predictors thereafter. Critically, the level of maintenance self-efficacy at T1 served as a strong and consistent prospective predictor and was associated with long-term adherence at T2 and T3.

This study suggests that exercise adherence after PCI follows a dynamic trajectory, with the relative predictive relevance of psychological and behavioral factors varying across stages. The shift from initiation-related predictors to maintenance-focused predictors, together with the sustained predictive association of early maintenance self-efficacy, may inform the development of phased assessment and intervention strategies. However, these findings should be interpreted as predictive associations rather than causal mechanisms and require external validation before clinical implementation.

Not applicable as a clinical trial. This study was a prospective observational cohort study without investigator-assigned intervention allocation and was not prospectively registered.