In a recent study published in Scientific Reports, researchers investigated the performance of a machine learning (ML)-based model in evaluating radiomic features to diagnose coronary artery disease (CAD) and its susceptibility using myocardial perfusion imaging (MPI) single-photon emission computed tomography (SPECT) images.
Study: Machine learning-based diagnosis and risk classification of coronary artery disease using myocardial perfusion imaging SPECT: A radiomics study. Image Credit: mi_viri/Shutterstock.com
Cardiovascular diseases (CVD) are a primary source of morbidity and death worldwide, with CAD being one of the most lethal. As a result, recognizing risk factors for this condition is crucial to take the appropriate precautions. MPI-SPECT imaging is a great asset for CAD diagnosis since it may offer a functional evaluation of the myocardium and heart arteries non-invasively.
However, the optical assessment of MPI SPECT is observer-dependent, error-prone, and time-consuming. As a result, automated, objective approaches for measuring cardiac MPI SPECT are in great demand.
About the study
In the present radiomics study, researchers investigated MP-SPECT image-based CAD diagnosis by ML. In particular, the team evaluated the performance of different ML models applied to delta, stress, and rest MPI SPECT radiomics for CAD diagnosis and risk classification.
The performance of classifiers built from three feature selections (FS) and nine ML-based algorithms was comparatively evaluated to identify the most accurate model for CAD status evaluation. The classifiers were gradient boosting (GB), extreme GB (XGB), K-nearest neighbor (KNN), decision tree (DT), multi-layer perceptron (MLP), random forest (RF), logistic regression (LR), support vector machine (SVM), and Naive Bayes (NB).
The three methods used for feature selection were Maximum Relevance Minimum Redundancy (mRMR), Recursive Feature Elimination using the Random Forest classifier (RF-RFE), and Boruta. The study included 395 individuals with suspected CAD who underwent a 48-hour rest-stress MPI SPECT. The enrolled population did not include individuals with myocardial infarction.
Among the participants, 78 were normal and 317 individuals were prone to CAD, among whom 135, 127, and 55 had low-, intermediate-, and high-risk, respectively. The left ventricular (LV) myocardium, eliminating the heart cavities, was delineated manually on rest-stress MPI-SPECT scans to determine the desired volume for investigation. Stress was induced by dobutamine, dipyridamole, and exercise.
In addition to clinical variables (family history, age, gender, smoking habits, ejection fraction, and diabetes mellitus status), 118 radiomic features were retrieved from the scans to delineate feature sets, such as the stress-, delta-, rest-, and combined feature sets. Feature extraction was based on the image biomarker initiative standardization (IBSI) and assessed using the Standardized Environment for Radiomics Analysis (SERA) protocol.
Of the data obtained, 80% was used to train and 20% to validate the model. Classifier performance was determined in two tasks, including (i) normal (CAD absence) and abnormal (CAD presence) classification and (ii) low and high-risk classification. Metrics such as area under the receiver operating characteristic curve (AUC), specificity (SPE), accuracy (ACC), and sensitivity (SEN) were determined to evaluate model performance.
Data were analyzed by two nuclear medicine physicians, and disagreement was resolved by consensus or consulting a senior physician. The physicians could access conventional SPECT scores, including the summed stress score (SSS), summed rest score (SRS), summed difference score (SDS), and wall thickening and motion data.
Results
The stress features model (in comparison to those based on other features) and those used for the CAS risk stratification task (in comparison to the first task models) showed better performances. The Stress-feature set using the mRMR-KNN classifier showed the best performance in the first task with SPE, SEN, ACC, and AUC values of 0.6, 0.64, 0.63, and 0.61, respectively.
The Boruta-gradient boosting model performed the best in the second task, with SPE, SEN, ACC, and AUC values of 0.76, 0.75, 0.76, and 0.79, respectively. Dependence counts normalized for non-uniformity, from the neighboring grey level dependence matrix (NGLDM) family and the status of diabetes mellitus from the clinical parameters were most frequently chosen from the stress set for classifying CAD risk.
Implications
Overall, the study findings highlighted the potential of machine learning models for classifying CAD risk using MPI-SPECT images. These models can significantly reduce the time-consuming MPI SPECT analysis for CAD diagnosis and risk evaluation. They also provide clinicians with insights into factors contributing to diagnosis, enhancing interpretability and trust in artificial intelligence-based automated models.
The model’s performance could be improved by developing it separately for each type of induced stress. Future studies should include patients with myocardial infarction and CAD-related clinical factors such as body mass index (BMI) and hyperlipidemia to enhance the generalizability of the study findings.
Wearable sensors and machine learning predict worker fatigue to improve safety and productivity