Abstract
Cardiovascular Disease (CVD) epitomizes class of disorders that disturb the vessels of heart and blood, encircling circumstances such as heart failure, coronary artery disease, and strokes, and leftovers a foremost global cause of morbidity and mortality. The early diagnosis of CVD is decisive as it consents for opportune involvement and organization, plummeting the risk of complications, improving treatment outcomes, and preventing further progression of the disease, ultimately contributing to better patient outcomes and overall cardiovascular health. Furthermore, early detection and diagnosis of CVD benefit significantly from the utilization of electrocardiograms (ECGs) and phonocardiograms (PCGs). The application of DL algorithms for identifying CVDs using PCG and ECG data has gained substantial attention, although a predominant number of existing approaches hinge on data sourced from a single modality. Henceforth, the development of proficient multi-modal Machine Learning (ML) techniques is crucial for effective prediction and detection of CVD. In this paper, we have proposed multi-modality-based CVD diagnosis framework named as multi-class model. In order to classify cardiovascular diseases into several categories using structured clinical data, this study introduces MCC-CVD, a new multi-component deep learning model. A real-world dataset of 920 patient records was used to assess the model. This dataset contains 13 clinical parameters, such as age, cholesterol level, resting blood pressure, fasting blood sugar, and other risk markers. The model used a two-stage weight correction technique and a tri-pattern attention mechanism (TPAM) to achieve robust performance, which allowed for more subtle feature weighting and better interpretability. Here, we utilized both quality enhanced ECG and PCG data through performing multiple processes including noise reduction and normalization. Besides, to evade misclassification data enhancement in terms of false peak elimination is performed based on adaptive thresholding features. After that, we fed the processed data into a multi-class architecture made up of three modules following. For extracting appropriate features, we designed Spectral Spatial Temporal Pyramid Network (SST-PNet) module. Additionally, Weight Correction Module with Attention Mechanism (WCM-AM) employs for weight maximum approach with three-pattern attention mechanism. Finally, novel Multi-class EnDe-CNN classifier is introduced to classify various CVD in multiple classes. A stratified 10-fold cross-validation method was used to carry out extensive studies. Outperforming baseline classifiers like SVM, Random Forest, and Logistic Regression, the suggested MCC-CVD model attained an average accuracy of 92.4%, F1-score of 0.87, precision of 0.89, and recall of 0.85. With an area under the curve (AUC) of 0.94, the model clearly has good discriminative potential across various subtypes of CVD. Furthermore, sensitivity analysis showed consistent performance even when changing parameters or data, and statistical testing validated the model's superiority with p-values less than 0.05.