Comparison of machine learning models for predicting 30-day readmission rates for patients with diabetes

Abstract

Background: Affecting over 25.5 million individuals in the United States, diabetes is linked to increased hospital readmission rates, almost double that of patients without diabetes. Previous research suggested the predictive superiority of deep learning (DL) models, particularly long short-term memory (LSTM) models, compared to traditional machine learning (ML) models, in forecasting diabetes-related hospital readmissions. The objective of this study was to compare one DL and 10 ML methods for predicting 30-day readmission rates in patients with diabetes. Methods: The dataset was obtained from “Diabetes 130-US Hospitals for Years 1999–2008” in the University of California Irvine (UCI) Machine Learning Repository and consisted of 101,766 unique encounters from 130 hospitals and integrated delivery networks across the United States. Encounters were included if they were an inpatient encounter (i.e., hospital admission) during which diabetes was entered in the system as a diagnosis. Predictors of readmission included demographic variables (age, race, gender), admission and discharge type, number of procedures and medications, diagnosis, and hemoglobin A1c (HbA1c) testing, a measure of the last 3 months of glycemic control. The outcome was the rate of readmission within 30 days. This study built upon traditional ML algorithms by incorporating the Grey Wolf Optimizer (GWO), a swarm intelligence-based optimization algorithm, for feature selection. Data preprocessing involved handling missing values, encoding categorical variables, and converting the target variable into a binary classification task. Feature engineering included creating service utilization variables, age group midpoints, and grouping admission types. Categorical features were one-hot encoded, and the dataset underwent standard scaling. Feature selection was conducted using the GWO algorithm, and class imbalance was addressed with synthetic minority over-sampling technique (SMOTE) during group k-fold cross-validation. Eleven ML algorithms, including random forest (RF), extreme gradient boosting (XGBoost), decision tree, support vector machine (SVM), and LSTM were employed for model training and evaluation. Results: RF emerged as the most favorable model, consistently outperforming others in F1 score, accuracy, precision score, and recall score. The RF model achieved the highest F1 score (0.83) and accuracy (0.88), indicating its superior predictive ability, especially in balancing precision and recall. XGBoost showed comparable results, achieving the second highest F1 score (0.84) and accuracy (0.88). Conclusions: The RF and XGBoost models with GWO outcompeted previous DL predictive modeling in diabetes hospital readmission scenarios. However, ML with GWO adoption should be carefully considered based on specific hospital needs, patient populations, and available resources, acknowledging ongoing advancements in predictive analytics for healthcare. Other advanced algorithms, including DL, still play a pivotal role. ML with GWO represents progress in predictive analytics for hospital readmission scenarios.

Topics & Concepts