Development and external validation of an explainable machine learning model for risk stratification of patients with chronotropic incompetence undergoing exercise stress SPECT-MPI: a dual-center study

Abstract Background Chronotropic incompetence is associated with increased cardiovascular risk in patients undergoing exercise stress single-photon emission computed tomography myocardial perfusion imaging (SPECT-MPI). However, prognostic tools integrating clinical, exercise, and perfusion imaging information in this population remain limited. We aimed to develop and externally validate an explainable machine-learning model for predicting major adverse cardiovascular events (MACE). Methods This retrospective dual-center study included 1,060 patients with chronotropic incompetence (heart rate reserve < 80%) undergoing exercise stress SPECT-MPI (derivation cohort, n = 765; external validation cohort, n = 295). The primary endpoint was composite MACE. Five survival algorithms were evaluated using strict five-fold nested cross-validation. All data-dependent preprocessing, LASSO-Cox stability selection, sample-weight selection, and hyperparameter optimization were confined to the corresponding outer-training folds, and internal performance was estimated from patient-level out-of-fold (OOF) predictions. The optimal algorithm was selected using a composite score integrating discrimination and prediction-error metrics. SHapley Additive exPlanations (SHAP) analysis was used for model interpretation. Results During a median follow-up of 63 months, 180 MACE events occurred. XGBoost achieved the highest composite score, with a nested OOF censoring-weighted C-index of 0.889 (95% CI, 0.861–0.917) and mean cumulative/dynamic AUC of 0.919 (95% CI, 0.892–0.943). Full-derivation stability selection retained five predictors: dyslipidaemia, diabetes, summed motion score, total perfusion deficit (TPD), and heart rate recovery at 3 min (HRR3). In the external cohort, the 5-year cumulative/dynamic AUC was 0.884 (95% CI, 0.818–0.933), compared with 0.834 for a Cox model using the same five predictors, 0.826 for TPD-plus-HRR3 XGBoost, and 0.811 for clinical-only XGBoost. Derivation-defined risk groups remained clearly separated for MACE in both cohorts (log-rank P < 0.001). Sensitivity analyses, including reintroduction of all six medication variables, supported the robustness of the final predictor set and model performance. Conclusions An explainable XGBoost survival model integrating clinical, exercise, and SPECT-MPI information provided robust risk stratification in patients with chronotropic incompetence and maintained discrimination in independent external validation. The model was implemented as a web-based research calculator. Its use should remain restricted to patients meeting the study eligibility criteria, excluding those with prior myocardial infarction or coronary revascularization, pending further prospective multicenter validation. Clinical trial number Not applicable.