CARDIOPREN: An Explainable Autoencoder–RNN Ensemble Framework for Accurate Cardiovascular Disease Prediction

Cardiovascular diseases are a leading cause of death throughout the world. Thus, there is a pressing requirement for developing a model for early and effective prediction based on symptoms. Here, a novel system for predicting heart attack using deep learning is proposed which blends a combination of temporal and unsupervised learning approaches in order to enhance diagnostic accuracy. To ensure model validity, proposed methodology includes proper data preprocessing which includes handling missing values, cleaning contradictory records and making use of large patient health databases like UCI Cleveland Heart Disease Dataset and Kaggle Heart Failure Clinical Records Dataset. Feedforward Neural Networks, Recurrent Neural Networks, Autoencoders and a hybrid ensemble model referred as CARDIOPREN are some of the deep learning frameworks which are compared. Feature extraction and dimensional reduction are taken up by autoencoders, while patient histories' sequential patterns are identified by RNNs which are augmented with Long Short-Term Memory units. From the results it is observed that 84%, 85.25%, and 84% accuracy were achieved by FNN, RNN, and Autoencoder-classifier combinations respectively. However, proposed CARDIOPREN model is a combination of Autoencoder and RNN modules which is superior over any other model with accuracy 93.4%, F1-score 0.91 and AUC-ROC 0.95. It is revealed in this study how accurately cardiac attack is predicted when it is combined in ensemble strategies which take up temporal modeling along with feature compression. A robust, universally applicable solution for clinical decision support and early cardiovascular disease diagnosis is offered by CARDIOPREN architecture.