A Comparison among Different Machine Learning Pretest Approaches to Predict Stress-Induced Ischemia at PET/CT Myocardial Perfusion Imaging

Abstract

Traditional approach for predicting coronary artery disease (CAD) is based on demographic data, symptoms such as chest pain and dyspnea, and comorbidity related to cardiovascular diseases. Usually, these variables are analyzed by logistic regression to quantifying their relationship with the outcome; nevertheless, their predictive value is limited. In the present study, we aimed to investigate the value of different machine learning (ML) techniques for the evaluation of suspected CAD; having as gold standard, the presence of stress-induced ischemia by ⁸²Rb positron emission tomography/computed tomography (PET/CT) myocardial perfusion imaging (MPI) ML was chosen on their clinical use and on the fact that they are representative of different classes of algorithms, such as deterministic (Support vector machine and Naïve Bayes), adaptive (ADA and AdaBoost), and decision tree (Random Forest, rpart, and XGBoost). The study population included 2503 consecutive patients, who underwent MPI for suspected CAD. To testing ML performances, data were split randomly into two parts: training/test (80%) and validation (20%). For training/test, we applied a 5-fold cross-validation, repeated 2 times. With this subset, we performed the tuning of free parameters for each algorithm. For all metrics, the best performance in training/test was observed for AdaBoost. The Naïve Bayes ML resulted to be more efficient in validation approach. The logistic and rpart algorithms showed similar metric values for the training/test and validation approaches. These results are encouraging and indicate that the ML algorithms can improve the evaluation of pretest probability of stress-induced myocardial ischemia.

Figure 1

Correlation matrix of the features used. The matrix elements are displayed in hierarchical clustering order. The numbers indicate the Spearman ρ coefficient between two features.

Figure 2

Importance of the features for each ML algorithm. ADA, AdaBoost, and Naïve Bayesian features importance were grouped into a single bar plot as the values for the two adaptive algorithms turned out to be equals, and Naïve Bayesian values differed with them by less than 5%.

Figure 3

Comparison among the ROC curves of the eight ML techniques considered. The ML performances are reported separately for the training/test approach (a) and validation approach (b). Parenthesis are reported the AUROC values.

Figure 4

Decision tree obtained by rpart algorithm. Each node or leaf is reported the prevalence concerning MPI outcome (nor: normal; isch: ischemic), the ratio between the number of prevalent and total patients, and the relative percentage.