Predicting all-cause readmissions using electronic health record data from the entire hospitalization: Model...

Area of Science:

• Health Services Research
• Medical Informatics
• Clinical Prediction Models

Background:

• Predicting 30-day hospital readmissions is crucial for patient care and healthcare costs.
• Current models often rely on limited data, potentially missing key predictive factors.

Purpose of the Study:

• To develop and evaluate a risk-prediction model for all-cause 30-day readmissions using comprehensive electronic health record (EHR) data from the full hospital stay.
• To compare the performance of this "full-stay" model against a "first-day" model and two established validated models (LACE and HOSPITAL).

Main Methods:

• An observational cohort study included medicine discharges from 6 North Texas hospitals (November 2009 - October 2010).
• A "full-stay" EHR-based risk-prediction model was developed and compared to a "first-day" model, LACE, and HOSPITAL models.
• Model performance was assessed using discrimination (C statistic), likelihood ratio, and net reclassification index.

Main Results:

• The study analyzed 32,922 admissions, with a 12.7% readmission rate.
• Significant predictors identified included hospital-acquired Clostridium difficile infection, vital sign instability on discharge, hyponatremia on discharge, and length of stay.
• The "full-stay" model demonstrated modest improvements in discrimination (C statistic 0.69) and reclassification compared to other models.

Conclusions:

• Incorporating granular EHR data from the entire hospital stay offers a modest improvement in predicting 30-day readmissions.
• The limited predictive gains suggest that factors beyond currently captured EHR data, such as psychosocial and behavioral elements, are critical for readmission.
• Future readmission models may need to integrate non-EHR data for enhanced accuracy.