Coronary artery disease detection based on iris images using machine learning

Coronary Artery Disease (CAD) occurs when the coronary arteries that supply the heart muscles become narrowed or blocked, limiting blood flow. It is the most common heart disease with the highest mortality rate. Early diagnosis of CAD can prevent the progression of the disease and facilitate the treatment process. In this thesis, an iris analysis and machine learning-based method is proposed for the non-invasive diagnosis of CAD using iris images. Iris analysis involves the blending of image processing methods and iridology. The study includes 155 volunteers in the patient group and 126 volunteers in the non-patient control group for CAD prediction. Eye images and biodemographic data of a total of 281 volunteers were obtained with the guidance of specialist physicians at the Cardiology outpatient clinic of Giresun University Training and Research Hospital. In this thesis, two different sets of attributes are utilized: iris and biodemographic attributes. The process of extracting iris features was carried out using image processing techniques. This process begins by employing the Integral-Differential operator method to determine the iris locations from the eye images. The iris, with its inner and outer boundaries determined through this method, was standardized by converting it into a rectangular format using the Rubber Sheet technique. The analysis region was obtained by cropping the heart region in the iris, referencing the Jensen iridology map. Image enhancement was applied to the analysis region using adaptive histogram equalization. For feature extraction, a 2-level wavelet transform was applied to the analysis region, yielding five first-order statistical, 22 Gray Level Co-occurrence Matrix (GLCM), and seven Gray Level Run Length Matrix (GLRLM) features for eight subcomponents. A total of 272 attributes were extracted, comprising 34 attributes for each subcomponent. The best features were determined using the Relieff feature selection method, and classification was performed using five different machine learning classifiers. The extracted features were transferred to the classification process based on the best features identified by the Relieff feature selection algorithm. The classification phase employed five different classifier families: Naive Bayes (NB), Decision Trees (DT), k-Nearest Neighbor (k-NN), Support Vector Machines (SVM), and Artificial Neural Networks (ANN). 10-fold cross-validation was utilized to evaluate the performance of the test data, considering accuracy, sensitivity, specificity, precision, F1 score, and AUC metrics. Performance measurements were conducted using iris parameters alone, as well as with the inclusion of physiological parameters of the volunteers as additional iris attributes. All analyses in this study were performed using the MATLAB programming language. In the proposed method, the highest value obtained, with 94.05%, belonged to the Linear SVM model as a result of iris analysis. SVM was followed by ANN with 92.86%, NN with 89.29%, DT with 88.1%, and finally, the k-NN classifier with 86.9%. SVM also performed the best in the metrics of Specificity, Precision, and F1 criterion with values of 92.31%, 93.48%, and 94.51% respectively. In the Sensitivity metric, k-NN was the best classifier with 100%, while the ANN model achieved the highest success with an AUC value of 97.01%. As a result of the analysis including iris and biodemographic data together, it was observed that the metric values increased. The accuracy metric increased to 95.24% in the SVM model. The SVM model also yielded the best results in Specificity, Precision, F1 criterion, and AUC metrics with values of 94.87%, 95.56%, 95.56%, and 98.12% respectively. In the Sensitivity metric, k-NN remained the best classifier with 100%. This study introduces a new model that is developed and evaluated for CAD detection, a common cardiovascular condition. The proposed model is compared to existing models previously used for CAD detection. The comparison results indicate that the proposed model outperforms the existing models in terms of the success rate in CAD detection. These findings demonstrate the effectiveness and superiority of the proposed model in accurately detecting CAD, highlighting its comprehensive and successful structure, unlike previous works in this field.