A three-stage model to predict hospital length of stay and death using machine learning algorithms

Abstract

Globally, healthcare administrators are working to reduce healthcare costs without sacrificing the quality of service provided. The most significant portion of health expenses is related to hospitalization. Length of stay directly impacts the hospitalization cost and resource management. Researchers have been trying to find the relevant factors to accurately identify LOS for the last few years. The early detection of the death of a patient in the hospital is also an essential aspect of hospitalization to manage the hospital resources better to give ultimate care to severe patients. This research developed a strategy to forecast the number of hospitalization days and deaths in a community using the USA Sparcs de-identified data repository. The data is not individually identifiable. The dataset consists of 2343569 instances and 34 features. This study proposes a three-stage model for improved hospital length of stay prediction: admission, post-admission, and discharge stage. These stage-wise models will help the patient and hospital management correctly predict LOS from the beginning of hospitalization to discharge. This study suggests a paradigm for predicting patient length of stay (LOS) using various machine learning (ML), Deep learning (DL), and ensemble techniques. This stage-wise prediction will be performed using a weighted probability averaging ensemble model. The weights of the individual classifiers will be selected based on their performance through a grid search method. The proposed ensemble model has outperformed the individual base ML models. The proposed ensemble model has improved balanced classification performance across stages, from admission to discharge. For death prediction, we propose a single-stage admission model.This model has a unique seven days death prediction implementation.This prediction will be performed based on Machine learning and deep learning models. The death-to-survived ratio is highly imbalanced in the sparcs dataset. We have evaluated our models using both synthesized data and original data. The performance of the models in the actual data is not so good, and the models performed better in the synthesized environment. We have implemented few imbalanced data models in the original data, and their performance is quite good compared to generic ML models. The death prediction in the post-admission phase can be implemented using a combination of ensemble and synthesized or imbalanced data models.