300 Section - Trajectory-Based Modeling of Functional Recovery in Cardiac Rehabilitation

Understanding functional recovery dynamics is essential for personalized care in cardiac rehabilitation (CR). This study used a trajectory-based modeling to identify distinct recovery patterns and developed an early prediction model using initial outpatient data to classify patients into clinically meaningful subgroups.

Design: We analyzed 105 patients who underwent CPET during outpatient cardiac rehabilitation (CR) from March 2020 to May 2025. Patients with ≥3 visits (up to 7) were included. Each visit (spaced 2–3 months, last interval 6 months) collected data on demographics, 6MWD, HGS, BMI, body fat, SMI, KASI, EQ-5D, and CPET metrics (VO₂ max, MET max, Fat max). Dynamic time warping (DTW) measured patient similarity across recovery variables. A network graph (patients as nodes, DTW-based similarity as edges) was constructed, and cohesive cliques (20–40 patients) were extracted. A CatBoost classifier using early-visit data (2–3 visits) predicted clique membership. Model performance was evaluated via AUROC, focusing on identifying patients at risk of poor recovery.

Results: Mean age was 59.0 ± 11.9 years, and 78.1% were male. Four cliques were identified (n=32, 32, 25, 16), each showing distinct recovery patterns. Clique 1 showed consistent improvement, while Clique 4 had minimal or plateaued gains, especially in MET max, KASI, and SMI. Cliques 2 and 3 showed intermediate trends. KASI closely paralleled MET max, and 6MWD tracked VO₂ max, supporting their use as proxies. The model using two visits yielded AUROCs of 0.77 (Clique 1) and 0.70 (Clique 4), with overall AUROC of 0.65. Adding a third visit improved performance (overall AUROC 0.74), with Clique 1 and 4 reaching 0.86 and 0.81, respectively.

Conclusion:

Trajectory-based modeling effectively identified distinct functional recovery patterns in CR. Using early outpatient data, the model predicted subgroup membership with promising accuracy. These findings support the use of data-driven stratification to inform personalized, proactive interventions in CR settings.