Transfer Learning Based Hybrid VGG16-Machine Learning Approach for Heart Disease Detection with Explainable Artificial Intelligence

Eshetie Gizachew Addisu ^1,2* Tahayu Gizachew Yirga ³ Hailu Gizachew Yirga ⁴ Alemu Demeke Yehuala ⁵

• ¹ University of Gondar, Gondar, Ethiopia
• ² Department of Information System, College of Informatics, University of Gondar, Gondar, Amhara Region, Ethiopia
• ³ College of Natural and Computational Science, Mekdela Amba University, Tulu Awuliya, Ethiopia
• ⁴ department of computer science, college of informatics, university of gondar, gondar ethiopia, Gondar, Ethiopia
• ⁵ Department of Surgery, School of Medicine, College of Medicine and Health sciences, University of Gondar, Gondar, Amhara Region, Ethiopia

Heart disease is a leading cause of mortality worldwide, and accurate early detection is crucial for effective treatment and management. In this study, we propose a novel hybrid machine learning approach that leverages transfer learning using the VGG16 convolutional neural network (CNN) combined with multiple machine learning classifiers for heart disease detection. Conditional tabular generative adversarial network (CTGAN) was employed to generate synthetic data samples from the actual datasets; statistical metrics, correlation analysis and domain experts evaluation were incorporated to evaluate the quality of synthetic datasets. The dataset comprises tabular data with 13 features, which is reshaped into image like format and resized to 224x224x3 to be compatible with the input requirements of the VGG16 model. Feature extraction is performed using VGG16, and the extracted features are fused with the original tabular data. This combined feature set is then used to train various machine learning models, including Support Vector Machines (SVM), Gradient Boosting, Random Forest, Logistic Regression, K-Nearest Neighbours (KNN), and Decision Tree. Among these models, the VGG16-Random Forest hybrid achieved a promising result across all evaluation metrics, including 92% of accuracy, 91.3% precision, 92.2% recall, 91.82% specificity, 92.2% sensitivity, and 91.75% F1-score. The hybrid models were also evaluated by unseen datasets to determine generalizability capabilities of the proposed approaches and relatively promising results are achieved by VGG16-Random Forest. Additionally, explainability is integrated into the model using SHAP values, providing insights into the contribution of each feature on the model prediction. This hybrid VGG16-ML approach demonstrates the potential for highly accurate heart disease detection with interpretable outcomes, offering valuable assistance in clinical decision-making processes.