Explainable Hybrid Deep Learning Framework for Cardiovascular Disease Prediction and Clinical Decision Support.
TL;DR
HCVDNet is a hybrid machine learning-deep learning model that can be used to improve the predictive accuracy of models as well as provide clinicians with clear explanations of the results of their predictions, and is capable of developing a high-performance, highly interpretable diagnostic framework for predicting cardiovascular risk.
Abstract
INTRODUCTION The main concern of the world's public health is cardiovascular disease. There are many different types of machine learning models available to help clinicians identify cardiovascular disease at an early stage and predict future risk. In this study, we propose HCVDNet, a hybrid machine learning-deep learning model that can be used to improve the predictive accuracy of models as well as provide clinicians with clear explanations of the results of their predictions.
Methods
The proposed architecture is a combination of three ensemble learning algorithms (XGBoost, CatBoost, and Random Forest) to learn features from data in the first step; then, the second step uses Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks to create a deep context model of the data. We have integrated two post-hoc explainable artificial intelligence techniques (SHAP and LIME) into our architecture to produce explanations at both the global and instance levels of data. Our experiments will use cardiovascular- related datasets found in the UCI and Kaggle databases.
Results
The performance of the proposed framework was significantly superior to that of other baseline machine learning and deep learning-based architectures in terms of accuracy, as it reached 96.94%, AUC reached 98.72%, and ECE was 0.021. As such, SHAP and LIME were used to evaluate the model's output to determine whether there were any clinically relevant and, therefore, reliable feature patterns within its outputs.
Discussion
The findings indicate that incorporating ensemble learning with deep temporal modeling enhances both the accuracy of predictive capabilities and the interpretability of these predictions. The interpretability modules further support physician confidence in the AI-assisted decision-making process by providing a clear explanation of how each feature contributes to the medically meaningful outputs generated from the model.
Conclusion
As a result, H-CVDNet is capable of developing a high-performance, highly interpretable diagnostic framework for predicting cardiovascular risk. In addition, because it is designed to be FHIR compliant, it can seamlessly integrate into existing electronic health record systems. Thus, there appears to be considerable opportunity to deploy this type of AI-based framework in real-time clinical settings and provide additional, dependable AI-based healthcare analytics.