Abstract
INTRODUCTION: Heart failure (HF) is a global medical condition marked by substantial morbidity, mortality, and healthcare costs with complex pathophysiology and variation in definitions. Machine learning (ML) has emerged as a promising approach to improve HF classification and risk prediction by leveraging various data sources. This study aims to present the current state-of-the-art multimodal ML models for HF classification and prognosis prediction, focusing on their modalities, performance, and clinical utility. METHODS: Following PRISMA guidelines and registered with PROSPERO (CRD420250654631), this review searched across four electronic databases (November 2014 - November 2024) and identified 284 unique records, of which 15 were included in the final synthesis. The quality of the studies was evaluated using QUADAS-2 and QUAPAS. RESULTS: Our results showed that the two most common multimodal combinations were tabular-image and tabular-text. The algorithms of the models included convolutional neural networks for image data, transformer-based approaches for text, with well-known fused techniques (early, middle, late fusion). Overall, multimodal models demonstrated superior performance compared to unimodal approaches, achieving area under the receiver operating characteristic curve values frequently exceeding 80% and reaching as high as 98.2%. CONCLUSION: Despite promising results, challenges include inconsistent reporting of performance metrics and their 95% confidence intervals, limited external validation, a near absence of prospective studies, and a deficiency in integrating genetic or 'omics' information with conventional data. These challenges must be addressed to promote clinical adoption and future research. SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/PROSPERO/view/CRD420250654631, identifier CRD420250654631.