Time-to-Event Pretraining for 3D Medical Imaging

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Biomedical Informatics

Background:

• Medical foundation models and imaging data are growing.
• Current self-supervised methods for 3D imaging models capture local features but miss temporal context for long-term outcome prediction.
• A missing context problem limits linking imaging biomarkers to disease progression.

Purpose of the Study:

• To introduce time-to-event pretraining for 3D medical imaging models.
• To leverage longitudinal electronic health records (EHRs) for temporal supervision.
• To identify imaging biomarkers predictive of future disease risk.

Main Methods:

• Developed a pretraining framework using paired, longitudinal EHR data with 18,945 CT scans.
• Utilized time-to-event distributions across thousands of EHR-derived tasks for supervision.
• Evaluated performance on 8 benchmark outcome prediction tasks.

Main Results:

• Achieved an average AUROC increase of 23.7% in outcome prediction.
• Demonstrated a 29.4% gain in Harrell's C-index across benchmark tasks.
• Maintained diagnostic classification performance.

Conclusions:

• Time-to-event pretraining effectively integrates longitudinal EHR and 3D imaging data.
• The method enhances clinical risk prediction by identifying imaging biomarkers linked to long-term health outcomes.
• This approach advances the development of medical foundation models for disease risk assessment.