A Novel Method for Identifying a Parsimonious and Accurate Predictive Model for Multiple Clinical Outcomes.

Area of Science:

• Biostatistics
• Clinical Epidemiology
• Health Services Research

Background:

• Traditional prognostic models focus on single outcomes, limiting their utility when multiple patient outcomes are of interest.
• Predicting simultaneous outcomes, such as nursing home admission and mortality in older adults, is clinically relevant but methodologically challenging.

Purpose of the Study:

• To propose and evaluate a novel computational method for selecting common predictors for multiple clinical outcomes.
• To provide open-source code for implementing the proposed predictor-selection algorithm.

Main Methods:

• Developed the Best Average BIC (baBIC) algorithm to select a parsimonious subset of common predictors minimizing average normalized Bayesian Information Criterion (BIC) across outcomes.
• Compared baBIC against Union, Intersection, and Full (no selection) methods using Harrell's C-statistic (predictive accuracy) and predictor count (parsimony).
• Validated the method using Health and Retirement Study (HRS) data and a comprehensive simulation study.

Main Results:

• The baBIC and Union methods yielded comparable predictive accuracy (average Harrell's C-statistic) across outcomes.
• baBIC produced more parsimonious models than the Union method, while Intersection models were most parsimonious but least accurate.
• Simulation results demonstrated baBIC's robustness in identifying relevant predictors.

Conclusions:

• The proposed baBIC method offers a superior balance between parsimony and predictive accuracy for models predicting multiple clinical outcomes.
• This approach enhances the development of clinical prognostic models for complex patient scenarios.