Generative Model Framework for External Validation in Predictive Modeling
A new framework distinguishes model deficiencies from case-mix effects in external validation of predictive models. It uses autoencoders to quantify each external patient's similarity to development data, enabling validation without sharing original data. Demonstrated on synthetic data and the Netherlands Heart Registration for mortality prediction.
Key facts
- External validation assesses transportability of predictive models.
- Interpretation confounded by differences between external and development populations.
- Framework quantifies patient-level similarity to development data.
- Uses autoencoders as generative models for similarity estimation.
- Offers flexible alternative to traditional linear approaches.
- Enables validation without sharing original development data.
- Demonstrated using synthetic data.
- Applied to Netherlands Heart Registration data for mortality prediction.
Entities
Institutions
- Netherlands Heart Registration
Locations
- Netherlands