Multi-Class Heart Disease Detection using ECG Images via Deep CNN Feature Extraction and Ensemble Stacking

Cardiovascular diseases (CVDs) continue to be the number one cause of mortality across the globe, illustrating the need for trustworthy and automated diagnostic methods. Electrocardiogram (ECG) analysis is a traditional method to identify cardiac abnormalities but the existing methods based on single convolutional neural networks (CNNs) or traditional machine learning (ML) classifiers suffer from overfitting, generalizing across different datasets, and addressing class imbalance, which in turn presents a barrier to developing robust systems with clinical deployment intent. This research addresses these issues by using a hybrid ensemble framework for multi-class ECG image classification. Our hybrid ensemble framework follows the approach of using transfer learning from CNNs (VGG16, VGG19, ResNet50, and InceptionV3) for deep feature extraction, applying dimensionality reduction (via Principal Components Analysis) on the reduced features, and then classifying them using a stacking ensemble of Random Forest, XGBoost, LightGBM, Multilayer Perceptron (MLP), and Support Vector Machine (SVM), with Logistic Regression serving as the meta-learner. We augmented the classes by applying the Synthetic Minority Over-sampling Technique (SMOTE) to handle imbalanced datasets. Our trials on datasets from Pakistan, Mendeley, and Bangladesh verified the effectiveness of our model, as it scored 97.6% on accuracy, 97.59% on the F1 score, and 0.9992 on the macro-AUC score, continuously performing better than both traditional ML classifiers and individual CNNs. The findings indicate that CNN-derived features combined with different ML classifiers improve the robustness of the model, its scalability, and its ability to generalize across clinical datasets. They underscore the role of the proposed model in performing disease diagnosis in real-time from an ECG and act as part of the advanced clinical decision support.