An Explainable Multimodal Data Fusion Approach for Heart Failure Detection

Heart Failure (HF) poses a significant healthcare challenge, marked by escalating prevalence and profound impacts on patient well-being. While current diagnostic protocols rely predominantly on clinical assessments, including electrocardiograms (ECG), echocardiograms, questionnaires, and blood tests, there exists a potential to enhance diagnostic precision through the integration of information from multiple modalities. This paper introduces an explainable multimodal data fusion strategy, which amalgamates electrocardiogram signals with carefully selected blood test results, aiming to refine heart failure detection. Utilizing a Convolutional Neural Network (CNN) model for electrocardiogram classification and an XGBoost algorithm for blood test analysis, the research utilizes the MIMIC-IV database, including MIMIC-IV-ECG, and employs robust pre-processing methods to ensure data quality. The study explores late fusion techniques to effectively integrate electrocardiogram and blood test data. By examining feature importance scores, the research identifies key blood tests for further investigation. Notably, the late fusion approach achieves an impressive accuracy of 97.46%, accompanied by robust sensitivity (97.16%) and specificity (97.67%). These findings underscore the potential of multimodal fusion techniques in augmenting heart failure detection accuracy, with profound implications for enhancing patient care and outcomes. The study contributes to the advancement of diagnostic methodologies and underscores the pivotal role of leveraging diverse data modalities in medical diagnosis.