Predicting readmission risk with institution-specific prediction models.

Area of Science:

• Health Informatics
• Clinical Prediction Models
• Healthcare Management

Background:

• Accurate prediction of patient readmission risk is crucial for hospitals, particularly under programs like the Hospital Readmission Reduction Program.
• Existing readmission risk prediction models often lack sufficient accuracy due to variations in hospital patient populations.
• There is a need for adaptable and precise models tailored to individual healthcare institutions.

Purpose of the Study:

• To develop a generic, institution-specific framework for predicting patient readmission risk.
• To create flexible models optimized for individual hospital characteristics and specific conditions.
• To enhance the accuracy and clinical utility of readmission risk prediction.

Main Methods:

• Developed a framework for institution-specific readmission risk prediction using patient data from a single institution.
• Experimented with classification (support vector machines) and prognosis (Cox regression) methods.
• Compared the proposed framework against industry-standard methods like the LACE model.

Main Results:

• The proposed framework demonstrated significantly higher prediction accuracy (AUC) compared to the LACE model across various conditions (e.g., AMI, HF, PN) and all-cause readmissions.
• Institution-specific models leveraging discharge-time features outperformed those using only admission-time features.
• Admission-time models from the proposed framework already surpassed the performance of the discharge-time LACE model.

Conclusions:

• The institution-specific readmission risk prediction framework is more flexible and effective than one-size-fits-all models.
• The proposed models offer significant improvements in prediction accuracy, sometimes yielding two to three times greater effectiveness.
• Admission-time models provide early warnings, enabling timely clinical interventions while patients are still hospitalized.