Development and internal validation of survival models for predicting a composite clinical endpoint using clinical variables and quantitative SPECT/CT imaging features

Quantitative parameters derived from single-photon emission computed tomography myocardial perfusion imaging (SPECT-MPI) may provide prognostic information in patients with coronary artery disease (CAD). This study aimed to develop and internally validate survival models for a composite clinical endpoint (CCE) comprising cardiovascular readmission, coronary revascularization, non-fatal myocardial infarction, and death by integrating clinical variables with quantitative SPECT-MPI parameters. A total of 1,674 patients with suspected or confirmed CAD who underwent resting gated SPECT-MPI were retrospectively included and randomly divided into a training set ( n = 1,171) and an internal validation set ( n = 503). Quantitative perfusion, left ventricular function, and phase parameters were extracted from SPECT-MPI images. Seven survival models were evaluated: Cox proportional hazards (CoxPH), LASSO Cox, single decision tree (RPart), random survival forest (RSF), gradient boosting machine (GBM), support vector machine (SVM), and extreme gradient boosting (XGBoost). Clinical variables and quantitative imaging features were considered as candidate predictors. Feature selection and model development were performed using the training set, while the internal validation set was kept separate from model fitting and hyperparameter tuning. Model performance was evaluated using the concordance index (C-index), time-dependent area under the receiver operating characteristic curve (AUC), graphical calibration, and decision-curve analysis (DCA). For XGBoost-based risk stratification, a risk-score cut-off derived exclusively from the training set was applied unchanged to the internal validation set, and event-free survival was compared using Kaplan-Meier analysis. During a median follow-up of 36.6 months (interquartile range, 20.5-58.7 months), 662 patients (39.5%) experienced the CCE. The retained imaging features included phase entropy, total perfusion deficit (TPD), and left ventricular ejection fraction (LVEF). Adding phase entropy to TPD and LVEF resulted in minimal improvement in validation-set discrimination (ΔC-index = 0.0009; 95% CI, − 0.0072 to 0.0090). RPart showed the highest 60-month AUC in the training set (0.825) but declined to 0.600 in the internal validation set, indicating substantial overfitting. In contrast, CoxPH, LASSO Cox, RSF, GBM, and XGBoost showed broadly comparable discrimination in the internal validation set, with 60-month AUCs ranging from 0.702 to 0.717 and C-indices from 0.662 to 0.674. XGBoost was subsequently used for exploratory risk stratification, rather than selected on the basis of superior discrimination. SVM showed weak and unstable discrimination, with a validation C-index of 0.508 (95% CI, 0.454-0.538) and a 60-month validation AUC of 0.483 (95% CI, 0.417-0.548). Graphical calibration showed general agreement between predicted and observed risks, although deviations were present for some models and time points. DCA showed positive net benefit for several models over portions of the evaluated threshold-probability range, without consistent dominance of a single model. The training-derived XGBoost risk-score cut-off separated patients into groups with significantly different event-free survival when applied unchanged to the internal validation set (log-rank P < 0.0001). Survival models integrating clinical variables with quantitative SPECT-MPI parameters provided moderate discrimination for the CCE. ML models did not demonstrate a clear discrimination advantage over CoxPH in the internal validation set, and RPart showed substantial overfitting. XGBoost was used for exploratory risk stratification, and the training-derived threshold separated patients into groups with different event-free survival when applied unchanged to the internal validation set; however, its discrimination was broadly comparable with that of several other models. These findings do not establish the superiority of a particular modeling approach. External validation using well-defined and independently adjudicated cardiovascular outcomes is required before clinical application.