Article type
Abstract
Background: Diabetes, often complicated, significantly contributes to approximately one-fifth of readmissions worldwide. Early identification of patients at heightened risk of readmission during their hospitalization is imperative. Nursing data obtained upon admission hold promise as critical determinants for predicting high-risk patients for readmission. Machine learning–based prediction models are renowned for their heightened accuracy and capacity to furnish robust empirical support in healthcare management regarding patient readmission.
Objectives: This study aims to build a readmission prediction model that scientifically predicts readmission risk for those at high-risk diabetic discharges with complications.
Methods: This retrospective study utilized electronic medical record data from 3005 patients with a primary diagnosis of diabetes mellitus with complications. The data were obtained from 3 hospitals in South Korea over 13 years. We divided into used initial data up to 1 day after hospitalization, Model 1, and used data from the entire hospitalization period, Model 2. We built 6 prediction algorithms for each model and compared their performance. The 6 algorithms are Logistic Regression, Random Forest, Decision Tree, XGBoost (extreme gradient boosting), CatBoost, and Multi-Perceptron Layer. The data imbalance problem was solved using an adaptive synthetic sampling approach, and 10-fold cross validation was used for model evaluation. The analysis in this study was performed using Python 3.11.3.
Results: The performance of predictive models was the Random Forest algorithm with an area under the curve of 0.73 in Model 1 and the CatBoost algorithm with an area under the curve of 0.76 in Model 2, which was the best of all the models. The medications, ward severity, health behavior habits, nursing diagnoses, and length of stay days are the 5 top factors in predicting high risk for readmission.
Conclusion: This study's machine learning–based readmission prediction model for diabetic patients with complications can be used as an evidence-based clinician decision support system for predicting unplanned readmissions.
* This study had no involvement with the public and/or consumers.
Objectives: This study aims to build a readmission prediction model that scientifically predicts readmission risk for those at high-risk diabetic discharges with complications.
Methods: This retrospective study utilized electronic medical record data from 3005 patients with a primary diagnosis of diabetes mellitus with complications. The data were obtained from 3 hospitals in South Korea over 13 years. We divided into used initial data up to 1 day after hospitalization, Model 1, and used data from the entire hospitalization period, Model 2. We built 6 prediction algorithms for each model and compared their performance. The 6 algorithms are Logistic Regression, Random Forest, Decision Tree, XGBoost (extreme gradient boosting), CatBoost, and Multi-Perceptron Layer. The data imbalance problem was solved using an adaptive synthetic sampling approach, and 10-fold cross validation was used for model evaluation. The analysis in this study was performed using Python 3.11.3.
Results: The performance of predictive models was the Random Forest algorithm with an area under the curve of 0.73 in Model 1 and the CatBoost algorithm with an area under the curve of 0.76 in Model 2, which was the best of all the models. The medications, ward severity, health behavior habits, nursing diagnoses, and length of stay days are the 5 top factors in predicting high risk for readmission.
Conclusion: This study's machine learning–based readmission prediction model for diabetic patients with complications can be used as an evidence-based clinician decision support system for predicting unplanned readmissions.
* This study had no involvement with the public and/or consumers.