Abstract
665
Objectives: Optimal machine learning models should incorporate both clinical and imaging data to predict major adverse cardiac events (MACE) in patients undergoing myocardial perfusion imaging with SPECT (SPECT MPI). However, clinical variables can only be used if they are available in electronic health records or manually inputted during SPECT MPI - which is time-consuming and error prone. We aimed to evaluate MACE prediction using reduced clinical features.
Methods: This study included 20,414 patients from the multicenter REFINE SPECT registry with images, clinical data, and follow-up for MACE. The average follow-up interval was 4.7±1.5 years. During follow-up, 3541 patients experienced at least one MACE (3.7% annualized risk): 1617 deaths, 379 MI, 1895 revascularizations, and 300 admissions for unstable angina. Twenty-seven clinical variables and seventy-one imaging variables were available. We trained 3 Extreme gradient boosting (XGBoost) models to predict MACE: 1) Model trained using all variables (M-All), 2) Model trained with all imaging and clinical variables available from SPECT MPI image headers (M1), 3) Model trained with all imaging and key clinical variables, not present in image header (M2). Clinical variables used in M1 and M2 are in Table 1. Prediction performance for the 3 models was evaluated using 10-fold cross validation and area under the receiver-operating curve (AUC) and compared with the performance of 4-point scale visual diagnosis and stress total perfusion deficit (TPD). We also obtained a variable importance ranking - for both imaging and clinical variables - using XGBoost to evaluate the predictive information of the reduced clinical features for MACE.
Results: The AUCs for MACE prediction were M-All: 0.797, M1: 0.779, M2: 0.793. The AUC for all 3 models was better than MACE prediction by visual diagnosis and stress TPD (AUC=0.680 and 0.698, respectively; p < 0.001 for all). 11 of the top ranking 15 features could be obtained from images or image headers.
Conclusions: It is possible to significantly reduce the clinical variables used by machine learning for MACE prediction, using either clinical information stored in SPECT MPI image headers or adding as few as 4 clinical variables not available in image headers. This will allow for practical implementation of automated MACE prediction, even if electronic health records are not available for “real-time” integration with SPECT MPI.