Developing a decision support tool to predict delayed discharge from hospitals using machine learning.

Area of Science:

• Healthcare Management
• Machine Learning in Medicine
• Health Informatics

Background:

• Healthcare systems face challenges in patient flow management due to increasing demand.
• Alternative Level of Care (ALC) patients, who no longer require acute care but face discharge delays, contribute to hospital overcrowding.
• Early prediction of ALC patients at admission is crucial for effective resource planning and improving patient flow.

Purpose of the Study:

• To identify patients likely to be Alternative Level of Care (ALC) upon admission.
• To determine the key features that predict ALC status.
• To develop guidelines for the early identification of ALC patients at the time of hospital admission.

Main Methods:

• Utilized patient data from Nova Scotia Health (2015-2022), including demographics, diagnoses, and clinical information.
• Applied data preprocessing techniques such as outlier management, feature engineering, missing value imputation, and standardization.
• Employed machine learning classifiers including Random Forest, Artificial Neural Network, and eXtreme Gradient Boosting (XGB), with techniques like class weights, random oversampling, and SMOTE to handle data imbalance.

Main Results:

• The XGB model with SMOTE demonstrated superior performance, achieving a recall of 0.95 and an AUC of 0.97 for ALC patient identification.
• Even when restricted to admission-only features, the XGB model with SMOTE maintained strong predictive power (recall 0.91, AUC 0.94).
• The most significant predictors for ALC status were identified as diagnosis 1, patient age, and entry code.

Conclusions:

• Machine learning models effectively predict Alternative Level of Care (ALC) status at admission, supporting real-time decision-making.
• The study provides a framework for early ALC identification, categorizing predictions into probability ranges to guide interventions.
• Implementing these predictive models can significantly improve patient flow and alleviate hospital overcrowding.