Predicting Hospital Readmissions in a Commercially Insured Population over Varying Time Horizons

Area of Science:

• Health Services Research
• Predictive Analytics
• Health Informatics

Background:

• Reducing hospital readmissions is a key federal policy objective.
• Existing predictive models often focus narrowly on 30-day readmissions.
• There's a need to evaluate models across diverse readmission timeframes.

Purpose of the Study:

• To assess a predictive model's performance for hospital readmissions within 14, 15-30, and 31-60 days.
• To analyze the impact of different predictor domains on readmission prediction accuracy.
• To evaluate model performance in a commercially insured population.

Main Methods:

• Retrospective analysis of 2.2 million inpatient admissions (2016-2017).
• Multivariate logistic regression with an 80/20 train/test split.
• Predictor variables included index admission, condition history, demographics, utilization, pharmacy, and environmental factors.

Main Results:

• Model C-statistics ranged from 0.618 to 0.915.
• Predictive accuracy decreased for longer readmission windows (14 days > 15-30 days > 31-60 days).
• Utilization history and condition history significantly improved predictive power.

Conclusions:

• Predicting longer-term readmissions is more difficult than short-term ones.
• Incorporating more data, especially utilization factors, enhances model performance.
• The developed models show potential for operational use in readmission prevention programs.