Machine Learning Application to Predict Bicycle Ergometer Test Results: a Prospective Cohort Study

The aim of the study was to develop an optimal technique to predict the results of a bicycle ergometry test (BET) based on the parameters recorded during a six-minute walk test (6MWT) using machine learning methods.

The study involved 56 patients who had experienced acute myocardial infarction and were undergoing the second stage of cardiac rehabilitation. The patients underwent a complete examination, including history taking, physical examination, anthropometric assessment, as well as a symptom-limited BET and 6MWT. During the 6MWT, we recorded the following: the distance covered, the heart rate, the blood pressure, the oxygen saturation, Borg rating of perceived exertion, the number of steps taken, and the electrocardiographic data. The algorithms for random forest, gradient boosting, k-nearest neighbors, and multiple linear regression were used to construct the machine learning models. The performance of the models was evaluated based on a determination coefficient, a mean absolute error, a mean square error, and a root mean square error. SHAP analysis was applied to interpret the findings.

The gradient boosting model provided the best prediction quality with a high determination coefficient (R² being around 0.99) and low error values for both target metrics: the distance walked in the 6MWT and the metabolic equivalent achieved during the BET. The significance analysis of features revealed the heart rate, age, and the body mass index to have the greatest impact on predicting the 6MWT distance, while for predicting the metabolic equivalent, the distance covered, the number of steps, and the body mass index were the most significant.

The developed gradient boosting-based machine learning model demonstrated its high efficiency in predicting the results of the 6MWT-based BET. The suggested method can serve as a valuable auxiliary tool to plan cardiac rehabilitation programs, particularly in cases when BET is difficult or impossible to perform. The use of SHAP analysis helped to understand the contribution of each feature to the prediction, increasing the confidence in the model results.