A systematic review of machine learning algorithms for mortality risk, readmission and phenotype prediction in patients with heart failure: exploring key data sources, input variables and outcomes.

BACKGROUND: Heart failure is not only a prevalent disease with a high mortality rate, but also generates high costs for healthcare systems. By training artificial intelligence (AI) models on medical data, it is possible to predict changes in health status that may lead to hospital readmissions or death. Such predictions enable better patient care and a proactive response to deterioration. METHODS: We conducted a systematic literature review of relevant AI studies using multiple machine learning (ML) algorithms to predict readmission and mortality in heart failure in previously diagnosed patients as well as clustering phenotypes. We synthesized and categorized the studies by the outcome variables, i.e. mortality, readmission, and phenotyping. The Scopus database was searched in September 2024 for relevant studies published between 2014 and 2024. Studies were included if they focused on heart failure, used data from electronic health records or hospital records, adopted machine learning techniques, analyzed readmission, mortality or phenotypes, and included patients no younger than 18 years of age. RESULTS: A total of 109 relevant studies were identified. In the mortality group (68 studies), age, serum creatinine level, serum sodium level, systolic blood pressure and blood urea nitrogen were among the most frequently mentioned relevant variables for predicting mortality. Comorbidities, blood urea nitrogen and age were identified as the most relevant variables for readmission (32 studies). The remaining studies dealt with phenotyping or further outcomes. Within all groups, random forest was the most recommended ML algorithm for prediction, followed by support vector machines. Nine key implications were derived from this review to guide future research and practice in AI-based studies. These implications emphasize improving model generalizability, data quality, and explainability to enhance the robustness and effectiveness of AI applications. LIMITATIONS: We did not use formal tools to assess the risk of bias due to missing results but addressed potential reporting bias by documenting missing performance metrics as "not available" and prioritizing studies with comprehensive reporting to ensure transparency in data synthesis and interpretation. CONCLUSIONS: The review shows that it is useful to group the studies into outcomes of readmission, mortality and phenotyping. This made it possible to highlight the relevant variables in each group. In addition, the different predictive capabilities of each outcome were identified. This research was conducted within the scope of the project KardioInterakt, which is supported by the German Federal Ministry of Research, Technology and Space (BMFTR) (grant number 16SV8906). CLINICAL TRIAL NUMBER: Not applicable.