Artificial intelligence informing clinical decision making on the risk of hospital readmissions in multi-morbid patients: a systematic review

Abstract

Abstract Introduction Multi-morbid patients have complex care needs and are more likely to utilise healthcare services, which represents a financial burden on healthcare organisations. 30-day readmission of patients is a method of measuring the quality of the services provided within healthcare.[1] Artificial intelligence (AI) models can flag multi-morbidity patients at increased risk of readmission.[2] However, it is unclear which common predictors have been used to develop these models. Aim To systematically review the medical literature to identify predictors that have been used to develop AI models that predicted unplanned 30-day hospital readmissions of multi-morbid patients, and whether these predictors were modifiable or non-modifiable. Methods Four large databases: Medline, Embase, Web of Science and Cumulative Index to Nursing and Allied Health Literature (CINAHL) were searched on 15th November 2022. Only publications that developed a machine learning algorithm to predict unplanned 30-day hospital readmission for patients with multi-morbidity were included. Articles published in English language were eligible for inclusion. A narrative synthesis of all eligible studies was undertaken. Key findings were identified through inductive strategy, which includes simplifying raw data in the included articles into a comprehensive format. This systematic review was registered with the PROSPRO database (CRD42022373937) and followed PRISMA guidelines. Results There were 1,906 articles extracted from all databases, 18 of which met our inclusion criteria. A total of 669 predictors of hospital readmission were found, and were divided into 103 modifiable (i.e., action could be taken to change the outcome) and 566 non-modifiable (i.e. age, gender, number of previous emergency admissions. The average number of modifiable predictors per prediction model was six, with the most common being length of hospital stay, multi-morbidity, obesity, hypertension, diabetes, depression, and anaemia. The most commonly used AI algorithm was the gradient boosting algorithm. This algorithm generally showed a better performance than other models reported. The performance of all models in the included studies was compared; 13 studies showed an average sensitivity of 72%. The average specificity was calculated at 73% from 11 studies. The highest AUC reported was above 0.9, which demonstrates an excellent performing model, while five models did not report a value for the AUC. Conclusion We identified modifiable and non-modifiable predictors used in previously developed AI algorithms that predict 30-day hospital readmissions of multi-morbid patients. Modifiable predictors in particular can help guide clinical decision-making by allowing early actions to be taken in primary care to potentially reduce the risk of readmission. However, one limitation of this review is that some included studies used balanced (i.e. synthetic) data which could have introduced a level of bias.