Abstract

Background: Hospital readmissions exert significant clinical and financial strains on healthcare systems. This is especially the case where reimbursement is outcome-connected in value-based care models. The prediction of hospital readmission and healthcare cost burdens associated with readmissions has increasingly involved the application of Artificial Intelligence (AI) and machine learning (ML). However, the literature is deficient in providing an integrative overview of both clinical and economic dimensions.

Methods: A systematic review compliant with PRISMA 2020 guidelines was performed. For scientific studies published in the last 12 years (2014 to May 2026), a multitude of sources were searched, e.g., PubMed, Scopus, Web of Science, IEEE Xplore, and Google Scholar. The studies under consideration utilized AI or ML in predicting healthcare cost and/or hospital readmissions. After screening and assessing eligibility, 80 studies were considered. Relevant study characteristics and associated outcomes were recorded. The algorithms and cost-related variables were also documented, as well as the model performance. The risk of bias in the studies was evaluated based on the PROBAST guidelines, and due to heterogeneity, the findings were summarized in a narrative format.

Results: Across many clinical research contexts, ML models consistently outperformed traditional statistical models in predicting hospital readmissions. In the context of ensemble methods, random forest and gradient boosting methods, as well as XGBoost, usually achieved moderate to high AUC levels. In terms of large data set predictive modeling, deep learning techniques demonstrated strong predictive capabilities, although they suffered from interpretability issues and limited clinical integration. The majority of studies were conducted in heart failure, diabetes, and cardiovascular disease, including various surgical patient populations. About 50% of the studies included costs to the system, either in terms of costs of hospitalization, penalties for readmissions, or in terms of resource consumption, which shows increasing focus on value-based care. The main limitations of the reviewed studies included variability in the definition of costs and a lack of external validation.

Conclusion: The use of artificial intelligence to construct predictive models of healthcare system readmission stratification and cost optimization yields beneficial outcomes. However, challenges relating to their interpretability, standardization, and external validation remain obstructive to their application in a clinical setting. Research should seek to develop predictive models that are explainable and remain cost-sensitive while being accurate and broad in their application.

Statistics