Applying machine learning to diagnose obstructive coronary artery disease using calcium scoring, coronary CT angiography, and clinical data

Introduction: Traditional diagnostic risk assessment in patients undergoing CT-based calcium scoring (CACS) and coronary CT angiography (CCTA) is based on a limited number of clinical and imaging findings. Machine learning has demonstrated its suitability to use a greater number of features to model complex non-linear relations leading to a more accurate diagnostic performance. Our aim was to develop and validate a machine learning model to diagnose obstructive coronary artery disease (oCAD) in patients without prior history of CAD, using CACS, CCTA and clinical data.

Methods: We retrospectively included 1254 patients without a prior history of CAD who underwent CACS and CCTA. Cardiac risk factors; smoking, hypertension, hypercholesterolemia, diabetes, family history of CAD; age; gender; body mass index; creatinine serum values; CCTA-outcomes, CACS and medication usage were registered at time of the scan. The entire dataset was split 4:1 into a training and test dataset, respectively. A XGBboost model was developed on the training set using fivefold stratified cross-validation and hyperparameter tuning. The test dataset was used to compare the diagnostic performance of the model to the performance of expert readers, looking at specificity, sensitivity, and ROC-analysis. The primary endpoint was oCAD on invasive coronary angiographies (ICA). Patients who were not referred for ICA were presumed not to have oCAD.

Results: A total of 85 (6.8%) of the 1254 patients (46% male, 60 ± 8 years of age) were diagnosed with oCAD. Age, total calcium score, body mass index, left anterior descending artery calcium score, and creatinine serum level were the highest-ranking features to predict oCAD. The performance of the XGBoost model did not differ from expert readers in predicting oCAD (p=0.84). The model and expert readers achieved a specificity of 0.97 vs 0.95, sensitivity of 0.71 vs 1.00, and an AUC of 0.96 vs 0.97, respectively.

Conclusion: The XGBoost model demonstrated a diagnostic performance comparable to that of expert readers. Features with negative predictive value in the diagnosis of oCAD were most valuable to the model such as high CACS. Currently, the model cannot replace expert readers, but it can contribute to the consistency of oCAD diagnoses and can be used to assist physician experts.