Can We Do More With Less While Building Predictive Models? A Study in Parsimony of Risk Models for Predicting Heart...

Area of Science:

• Cardiology
• Health Informatics
• Biostatistics

Background:

• Hospital readmissions for heart failure (HF) represent a significant financial and clinical burden.
• Accurate prediction of 30-day HF readmission risk is crucial for effective post-discharge interventions.
• Increasing data complexity necessitates efficient risk prediction models.

Purpose of the Study:

• To explore and demonstrate predictor reduction methods for HF readmission risk models.
• To assess the impact of reducing predictors on predictive performance and model-building time.
• To identify optimal models balancing accuracy and efficiency.

Main Methods:

• Applied three predictor reduction techniques to a real-world dataset of 1210 patients with 57 variables.
• Compared models derived from reduced predictors against a full 57-predictor model.
• Evaluated predictive performance using the C-statistic and measured model-building time.

Main Results:

• Predictive performance (C-statistic) ranged from 0.630 to 0.840.
• Model-building time varied significantly, from 10 minutes to 10 hours.
• A final model with 16 predictors achieved a C-statistic of 0.832, comparable to the full model (0.840), with a 66-fold improvement in building time.

Conclusions:

• Predictor reduction methods can significantly decrease model-building time for heart failure readmission risk.
• Efficient models with fewer predictors maintain high predictive accuracy.
• Optimized models facilitate timely and effective post-discharge care planning.