Machine learning methods, applications and economic analysis to predict heart failure hospitalisation risk: a scoping review.

BACKGROUND

Machine Learning (ML) has been transformative in healthcare, enabling more precise diagnostics, personalised treatment regimens and enhanced patient care. In cardiology, ML plays a crucial role in risk prediction and patient stratification, particularly for heart failure (HF), a condition affecting over 64 million people globally and imposing an economic burden of approximately $108 billion annually. ML applications in HF include predictive analytics for risk assessment, identifying patient subgroups with varying prognoses and optimising treatment pathways. By accurately predicting the likelihood of hospitalisation and rehospitalisation, ML tools help tailor interventions, reduce hospital visits, improve patient outcomes and lower healthcare costs.

OBJECTIVE

To conduct a comprehensive review of existing ML models designed to predict hospitalisation risk in individuals with HF.

METHODS

A database search including PubMed, SCOPUS and Web of Science was conducted on 31 March 2024. Studies were selected based on inclusion criteria focusing on ML models predicting hospitalisation risks in adults with HF. The data from 27 studies meeting the criteria were extracted and analysed, with a focus on the predictive performance of the ML models and the presence of economic analysis.

RESULTS

Most studies focused on predicting readmission rather than first-time hospitalisation. All included studies employed supervised ML algorithms, with ensemble-based methods generally yielding the highest predictive performance. For 30-day hospitalisation or readmission risk, Extreme Gradient Boosting (XGBoost) achieved the highest mean area under the curve (AUC) (0.69), followed by Naïve Bayes (0.68) and Deep Unified Networks (0.66). For 90-day risk, the best-performing models were Least Absolute Shrinkage and Selection Operator and Gradient Boosting, both with a mean AUC of 0.75, followed by Random Forest (0.67). When the prediction timeframe was unspecified, Categorical Boosting achieved the highest performance with a mean AUC of 0.88, followed by Generalised Linear Model Net and XGBoost (both 0.79).Electronic health records were the primary data source across studies; however, few models included patient-reported outcomes or socioeconomic variables.None of the studies conducted an economic evaluation to assess the cost-effectiveness of these models.

CONCLUSIONS

ML holds substantial potential for improving HF care. However, further efforts are needed to enhance the generalisation of models, integrate diverse data sources and evaluate the cost-effectiveness of these technologies.