Development of Machine Learning Model for Predicting Length of Stay and Mortality Among Hospitalized Patients with Type 2 Diabetes Mellitus and Hypertension

Background:
Type 2 diabetes mellitus (T2DM) and hypertension (HTN) are major global health issues, particularly in lower-middle-income countries, where they contribute significantly to mortality rates. Inadequate treatment and management of these diseases lead to higher healthcare usage, increased medical costs, and significant societal burdens, especially in countries like Indonesia. Accurate prediction of clinical outcomes can guide treatment decisions, optimize healthcare costs, and improve patient care. This study focuses on predicting the length of stay (LoS) and mortality among hospitalized patients in Taiwan using the National Health Insurance (NHI) dataset.
Methods:
We analyzed insurance claims data of 245,643 patients with both T2DM and HTN hospitalized in around 500 hospitals in Taiwan. We used several prediction models based on machine-learning algorithms to predict LoS about disease severity, physician-in-charge, room type, comorbidities, and types of procedures performed. The primary outcomes in our study were mortality and LoS. Mortality was predicted as a categorical variable, with 1 for "alive" and 0 for "death," while LoS was expected as a continuous variable.
Results:
XGBoost was the best-predicted inpatient with T2DM and LoS, achieving an R2 score of 0.5349. The other metrics obtained by the model were root mean square error (RMSE) of 0.4020, mean absolute error (MAE) of 0.3065, and mean square error (MSE) of 0.1616. The best results were observed for inpatient mortality using GBC (AUROC 0.7591).
Conclusions:
The best model to predict the LoS was XGBoost, and the best model to predict mortality was GBC. Patients with primary diseases such as T2DM or HTN may have comorbidities that can prolong inpatient LoS. k
Keywords: artificial intelligence, type 2 diabetes mellitus, hypertension, mortality, length of stay, comorbidity