A multiple-model generalisation of updating clinical prediction models

Area of Science:

• Biostatistics
• Health Informatics
• Clinical Epidemiology

Background:

• Clinical prediction models (CPMs) are crucial for healthcare decision-making.
• Current methods often involve isolated, repetitive model development (de novo derivation) or single-model updating, failing to utilize all available data.
• Existing model updating techniques are limited to adjusting a single existing CPM.

Purpose of the Study:

• To develop a generalized method for updating and aggregating multiple CPMs.
• To create a "hybrid method" that re-calibrates multiple CPMs and revises covariates using individual participant data (IPD).

Main Methods:

• The "hybrid method" employs stacked regression for re-calibration and penalized likelihood for covariate revision with IPD.
• Performance was evaluated against existing methods using a clinical example (mortality risk after transcatheter aortic valve implantation) and two simulation studies.
• Simulation studies assessed the impact of IPD sample size and between-population heterogeneity.

Main Results:

• With small IPD sample sizes, stacked regression and the hybrid method showed comparable and superior performance.
• With large IPD sample sizes, developing a new model de novo and the hybrid method yielded the highest performance.
• The hybrid method demonstrated robust performance across various scenarios.

Conclusions:

• The proposed hybrid method offers a flexible strategy for updating and aggregating multiple CPMs.
• This approach effectively incorporates IPD, existing research, and prior knowledge into model development.
• The findings inform decisions on whether to update existing CPMs or derive new models de novo.