1013 - Predicting aerobic capacity after cardiac rehabilitation using machine learning models

Objectives: Predicting aerobic capacity after cardiac rehabilitation using machine learning models

Design: Medical records of a total of 162 patients who underwent cardiopulmonary exercise tests (CPET) for outpatient CR were retrospectively analyzed.To predict VO₂ peak, linear regression and 9 ML models were analyzed. Two tasks were designed for VO₂ peak prediction. Task 1 estimated VO₂ peak at the same visit point using clinical characteristics and functional assessments excluding CPET. Task 2 predicted VO₂ peak at the next visit based on the inter-visit exercise amount and previous visit data including CPET results. Data analyzed for Task 1 consisted of variables for demographics, medical history, physical status, 6-minute walk distance (6MWD), hand grip strength (HGS), and questionnaires with functional status. In Task 2, CPET results from previous visit and exercise variables (type, intensity, frequency, and total time) were analyzed along with the top 12 variables with high importance in Task 1.

Results: The average age of a total of 162 patients was 60.0 (±13.4) years and 123 (75.9%) patients were males. In Task 1, LightGBM showed the highest performance (MAE: 3.23 (±0.63), MAPE: 16.69 (±3.59), RMSE: 4.19 (±0.72)). In feature importance analysis, 6MWD, HGS, and KASI were ranked the highest. In Task 2, LightGBM also showed the highest performance (MAE: 3.13 (±1.47), MAPE: 13.65 (±6.53) RMSE: 4.76 (±7.10)), outperforming Task 1. In the feature importance analysis, the importance of CPET outcomes including VO₂ peak, test duration and VE/VCO₂ was high overall. HGS, 6MWD, total length of hospital visits, and EQ5D were also ranked high in importance. Among exercise variables, total time of resistance training was the most important.

Conclusions: Using the ML model, VO₂ peak was successfully estimated based on functional evaluation and clinical characteristics without CPET in patients underwent CR. Using the previous CPET and exercise amount, the VO₂ peak could be predicted more accurately.