Finding a constrained number of predictor phenotypes for multiple outcome prediction

Area of Science:

• Medical informatics
• Clinical prediction modeling
• Biostatistics

Background:

• Prognostic models are crucial for medical decision-making but often predict single outcomes.
• Existing models require unique predictors, limiting their integrated use.
• A need exists for a unified clinical tool predicting multiple outcomes from common predictors.

Purpose of the Study:

• To identify a constrained, outcome-agnostic set of predictors.
• To develop a method for creating versatile prognostic models.

Main Methods:

• A novel technique aggregated standardized mean differences across numerous outcomes.
• This method identified a constrained set of predictors relevant to multiple outcomes.
• Model performance was evaluated on eight prediction tasks against various benchmarks.

Main Results:

• A constrained set of 67 predictors, plus age and sex, was identified.
• These predictors encompass cardiovascular, kidney/liver, mental health, gastrointestinal, infectious, and oncologic conditions.
• Models using the constrained set showed comparable or slightly lower discrimination than models with extensive predictors, outperforming existing clinical models.

Conclusions:

• A small, common set of predictors can support models for numerous outcomes.
• This facilitates the implementation of multiple prediction models via a single interface.
• The identified predictor set serves as a foundation for future prognostic model research.