Interpretable Machine Learning-Based Algorithms for Cardiac Anomaly Detection

Interpretable Machine Learning-Based Algorithms for Cardiac Anomaly Detection Abstract: Cardiovascular Diseases (CDV) remain the leading cause of death globally. Accounting for an estimated 17.9 million deaths per year, according to the World Health Organization (WHO). In order to effectively treat and manage cardiac abnormalities, which lead to a number of mortality and also ultimately lower the morbidity linked to these disorders, the essential step is early detection of the disease. The main aim of this study is to look in more detail at the performance of supervised classification machine learning algorithms to detect cardiac anomalies from rest/stress myocardial perfusion imaging (MPI) in single-photon emission computed tomography (SPECT). A total of 268 patients who performed a 2-day stress-rest protocol MPI SPECT were suspected to have a cardiac abnormality. Altogether, 401 features were extracted from Rest-Stress MPI SPECT images. These features included different sets, consisting of Rest-Stress, Delta, and combined-radiomics (a combination of all sets of features). To have training and testing parts, the data was randomly divided into subsets of 75% and 25%. To evaluate the performance of machine learning, combinations of three scaling techniques, four feature selections, nine classifier algorithms, and two search strategies were used. To evaluate the model performance, different metrics consisting of Specificity (SPE), Sensitivity (SEN), Accuracy (ACC), and area under the ROC curve (AUC) were measured. Models built using a combination of Rest-Stress feature sets performed better than models of Rest and Stress alone. The metrics results were ACC = 0.83, AUC = 0.86, SPE=0.81, and SEN = 0.81 for Robust Scaler Scaling, the Logistic Regression (LR) classifier with selected features from the Model-based Feature Selection (FS), and Random Search for parameter optimization. For models with the highest performance, a designed interpretability model was implemented. Shapley values of features are calculated to find the features that have the highest effect on the output result. As an example, the “Zone distance variance” feature from the Gray-Level Zone Size Matrix (GLSZM) values, significantly impacted the final result. In this study, we developed a high-performance model for accurately detecting anomalies in cardiac SPECT images. By incorporating interpretability features, we enhanced the model's clinical utility, making it more suitable for implementation in clinical settings in the future. Our approach improves diagnostic accuracy and provides clear insights into the model's decision-making process, facilitating its adoption in routine clinical practice.