An Interpretable LES-Integrated Hybrid RNN-CNN Framework for Multimodal Cardiovascular Disease Prediction

Abstract

Coronary artery disease is one of the leading causes of death worldwide. There is a need for an efficient diagnostic system to detect CAD that can lead to effective clinical therapy. Existing deep learning models are mostly designed for either spatial imaging or temporal physiologic signal learning, but they have limited fusion ability and do not extract lesion-aware features. In response to these challenges, this study proposes an interpretable Local Energy Shape (LES)- integrated hybrid Recurrent Neural Network and Convolutional Neural Network (LES-RCNN) framework for cardiovascular disease prediction and classification using multimodal image and physiological datasets. The proposed framework implements LES-based feature extraction combined with recurrent learning and convolutional feature modeling to extract spatial, temporal, and lesion-based cardiovascular characteristics from CCTA, ECG image data, and structured CSV physiological records. Lesion localization is greatly enhanced by the structural representation provided by LES transformation. Also, the model's spatial and temporal feature learning is improved with a hybrid RNN-CNN architecture. Experimental assessment was carried out utilizing both CVD Atlas image datasets and UCI physiological datasets in binary and multi-label classification.  The proposed 17-layer LES-RCNN architecture obtained 98.5% accuracy, 98.2% precision, 98.7% recall, and 98.4% F1-score. This was better than existing state-of-the-art cardiovascular disease prediction models. Through comparisons and convergence studies, it was confirmed that the learning was more stable, lost less, and had better generalization capacity. The results of the study reaffirm that the LES-based multimodal framework can provide an accurate, interpretable & computationally efficient diagnosis of early CVD & clinical decision support for clinical decision-making.