Background: The application of machine learning (ML) to predict hospital length of stay (LoS) displays promise for advancements in healthcare management and patient care. LoS is frequently seen as a metric that can help determine the severity of sickness, cost of care, and resource use. Furthermore, individualized discharge planning has been linked to quantifiable outcomes including lower rebound admission rates and increased patient satisfaction. However, these results depend on best practice standards being in place for working health professionals. The goal of this study is to determine ML’s viability to assist the healthcare process in predicting LoS. If ML can help predict the LoS then this would help improve patient care and reduce hospital penalizations from the Hospital Readmissions Reduction Program.
Methods: A systematic review was conducted using mainly PubMed articles (with two exceptions, one article from ACM and the other from MDPI) using the keywords Machine Learning and Length of Stay. We found 24 studies from January 2020 to January 2024 that highlight the efficacy of different ML models/algorithms in various LoS medical contexts, looking specifically at their potential to improve operational decisions, resource allocations, and clinical outcomes. The scope of this systematic review is to evaluate recent developments that are related to the use of ML in LoS prediction.
Results: We found in our systematic review that a majority of studies highlight the accuracy of ML models being very high when predicting LoS. The accuracy rating cited by multiple sources was often greater or equal to 89%, however many did mention the limitations of needing a larger data scope and potential biases in the program. The algorithm’s predictions can be classified as short-term or long-term LoS predictions for patients, with the former usually being a prediction of fewer than 7 days and the latter being anything past short-term. However, the prediction quality often decreases near the beginning of the long-term marker. This (short and long) marker varies between studies due to factors such as data quality and the ML algorithm/model used. Many studies reference the Random Forest model as having the highest or near highest accuracy when predicting LoS. Furthermore, this accuracy was consistent across different sectors of medicine, with missing or uncategorized data.
Conclusions: ML models hold substantial promise in predicting hospital LoS, potentially transforming healthcare operations by enabling more informed decision-making, early intervention, efficient scheduling, decreased workload on physicians, a more informed patient, and optimized resource allocation. This would ultimately help create better-personalized discharge plans that would benefit the patient, physician, and hospital administrators. Programming advancements and further validation are crucial to realize the possible benefits fully.