Machine learning approach to identify clusters of patients with distinct cardiac chronotropic response during incremental exercise

Background The chronotropic cardiac response is generated by the interplay between sinoatrial and atrioventricular nodal functions, it is continuously regulated by the autonomic nervous system and depends on number of factors notably physical fitness, metabolism or treatments. The chronotropic cardiac response is responsible for the complex heart rate (HR) changes or trajectory that can be observed during an incremental exercise. However, no standardized computational framework exists to stratify patients by chronotropic phenotype at scale. Purpose To develop a machine-learning framework to identify subtle findings present in HR trajectories recorded during incremental physical exercise identifies Methods A large cohort of 3250 cardiopulmonary exercise tests (CPETs) performed in a routine cardiology hospital during 2017 and 2025 were analyzed Four parameters were algorithmically derived from each HR signal: (i) tHR - time of occurrence of a slope change in the HR trajectory, detected via segmented regression with Davies test validation and anchored to the first ventilatory threshold (VT1); (ii) αHR - angle of slope change at tHR, quantifying the direction and magnitude of HR inflection; (iii) εHR - a dimensionless elasticity-based metric of chronotropic responsiveness; and (iv) VO₂max - peak oxygen uptake normalized to body weight. A sinusoidal transformation was applied to αHR to limit the leverage of extreme angles. Agglomerative hierarchical clustering (Euclidean distance, complete linkage) was applied to identify exercise phenotypes. A Random Forest classifier was trained on the clustering solution to enable stable phenotype assignment across datasets. Results Four reproducible exercise phenotypes were identified. The largest cluster (~44%) represented the best exercise condition, characterized by the latest tHR, the flattest HR trajectory (αHR ~175°, near-linear), the highest εHR, and the longest exercise duration - predominantly deconditioned patients. A second cluster (~37%) showed earlier tHR, moderately deflected HR slope, and shorter effort duration, representing intermediate exercise capacity. A third cluster (~16%) presented the earliest tHR, the most negative αHR, the lowest VO₂max, and the shortest exercise duration. A fourth small but distinct cluster (~4%) was defined by a negative αHR combined with a negative εHR. The clustering was reproduced in a cohort of 158 patients with a deep cardiac phenotype. Patients with heart failure were in cluster 3 and 4 with those in cluster 4 being the most severe characterized by negative αHR and εHR (cluster 4). Conclusion A machine-learning pipeline applied to algorithmically derived HR trajectory parameters identifies four clinically meaningful chronotropic phenotypes ranging from high exercise performance to severe incompetence. The direction and timing of the HR slope change at VT1, captured by αHR and tHR, emerge as key discriminating axes