Machine learning approach to identify clusters of patients with distinct cardiac chronotropic response during incremental exercise
A Rammal, R Tchala Sare, C Kerdiles, R Isnard, M Zeitouni, M Ponnaiah, S N HatemAbstract
Background
The chronotropic cardiac response is the main mechanism by which cardiac output increased during exercise; its alteration is central to exercise intolerance during cardiopathies. Chronotropic cardiac response is generated by intrinsic heart rhythm mechanisms continuously regulated by the autonomous nervous system and resulting in complex heart rate trajectories during incremental exercise.
Purpose
To develop a machine learning framework that can analyze heart rate trajectories during an incremental exercise and categorize patients according to their chronotropic response patterns.
Methods
Exercise test data from 2,048 patients undergoing cardiopulmonary exercise testing on a ramp protocol were analyzed. Heart rate signals were processed using two statistical methods: (1) Davies’ test to assess linearity, and (2) a novel metric for HR elasticity to detect inflection points. The first ventilatory threshold (AT) was estimated from ventilatory equivalents (VE/VO2) and used to anchor trajectory analysis. Unsupervised inferential clustering was used to identify groups of patients with distinct heart trajectory during exercise.
Results
ML approaches identified 3 metrics that characterized heart rate trajectory i.e. - change in HR slope around AT defined by the time of occurrence (tHR) and the angle of slope, αHR , and - dynamic or elasticity εHR,of the trajectory that correspond to HR response. Using an AI-based hierarchical clustering on the cohort of 2,048 patients, 4 clusters of patients could be identified using the heart trajectory parameters generated by the algorithm. The clustering solution and derived marker were externally validated in an independent cohort of 638 patients, where all four patterns and marker distributions were consistently reproduced, confirming generalizability and robustness. Clusters of patients differed by exercise performance and clinical history. Notably, one cluster corresponded to a subpopulation of patients in severe heart failure (HF) showing an early tHR, a negative αHR and low εHR. We will present evidence that the auriculoventricular coupling is the major determinant of change in HR slope around AT.
Conclusion
We developed a machine learning algorithm to analyze heart rate trajectory during an incremental exercise. Using AI-based unsupervised hierarchical clustering, we identified 4 clusters of patients based on their heart trajectory and found that abnormal auriculoventricular coupling is a major mechanism of chronotropic incompetence during HF.