Artificial intelligence and machine learning in heart failure: a systematic review from risk prediction to phenomapping
F Khan, N KhanAbstract
Background
Heart failure is a heterogeneous syndrome that defies precision with traditional models. AI and ML provide methods to synthesize complex, high-dimensional data that could transform our approach to HF management.
Aim
To systematically assess how AI/ML are being used across different data domains to improve risk prediction, diagnosis, and phenomapping in HF and how they perform.
Methods
A literature search of PubMed, IEEE Xplore, and Scopus was performed from January 2018 through December 2024. Studies developing or validating AI/ML models (supervised or unsupervised) for HF utilizing EHR, biomarkers, or imaging data were included. We extracted data on model type, data source, performance metrics, and clinical outcomes and synthesized the information narratively.
Results
Out of 1527 records, 89 passed the criteria for selection. Key Message: 1) EHR/Clinical Data: Machine learning models, particularly XGBoost and neural networks, outperformed conventional risk scores like MAGGIC for 1-year mortality risk, with median AUC improvements ranging from 0.07 to 0.11. NLP added prognostic information for clinical notes. 2) Biomarkers & Multimodal Integration: Integrations of proteomic/metabolomic information with clinical data, using machine learning, yielded novel prognostic models beyond BNP alone. 3) Imaging: Deep learning can automate ejection fraction and strain analysis for echocardiography and cardiac magnetic resonance, with expert performance, and predict outcomes directly from images. 4) Phenomapping: Using 22 studies, k-means clustering, and other unsupervised methods, consistently yielded 3 to 5 groups with varying outcomes, including a metabolic-inflammatory subgroup with increased risk for hospitalization.
Conclusion
AI and ML exhibit higher predictive performance in HF data domains than conventional approaches and can stratify HF into actionable phenogroups. The critical next step will be prospective validation and seamless integration of such models into clinical workflows to further precision medicine.For image description, please refer to the figure legend and surrounding text.For image description, please refer to the figure legend and surrounding text.