Machine Learning-Enabled 30-Day Readmission Model for Stroke Patients.

Area of Science:

• Neurology
• Medical Informatics
• Data Science

Background:

• Hospital readmissions represent a significant financial and clinical challenge.
• Stroke patients have a high readmission rate, impacting care quality.
• Predicting and reducing stroke readmissions is crucial for healthcare improvement.

Purpose of the Study:

• To identify predictors of 30-day readmission after ischemic stroke.
• To develop and compare machine learning models for predicting stroke readmissions.
• To enable targeted interventions for high-risk stroke patients.

Main Methods:

• Utilized patient data from electronic health records (EHR).
• Applied five machine learning algorithms: random forest, gradient boosting machine, extreme gradient boosting (XGBoost), support vector machine, and logistic regression (LR).
• Employed data-driven feature selection and adaptive sampling techniques.

Main Results:

• Included 3,184 ischemic stroke patients; key predictors identified include NIH Stroke Scale score > 24, indwelling urinary catheter, hypercoagulable state, and percutaneous gastrostomy.
• XGBoost with ROSE-sampling achieved an AUC of 0.74 for predicting 30-day readmission.
• Logistic Regression with feature selection and ROSE-sampling demonstrated the highest sensitivity (0.53).

Conclusions:

• Machine learning models effectively predict 30-day stroke readmissions using EHR data.
• XGBoost and LR models offer distinct advantages in performance (AUC vs. sensitivity).
• Identified clinical variables and developed models can guide personalized interventions to reduce readmissions.