August 15, 2025JAMIA OpenOpen Access

Benchmarking of pre-training strategies for electronic health record foundation models

View Full Paper

Ask AI

Bookmark

Task-specific supervised pre-training achieved superior performance with AUROC of 0.70 and AUPRC of 0.23 on adverse cardiac events.
The model evaluation showed that while supervised methods excel, self-supervised models outperformed in mortality prediction.
Three pre-training strategies were analyzed using a transformer architecture, including baseline, self-supervised, and supervised approaches.
Selection of pre-training strategies should align with application needs, data availability, and desired generalization capabilities.

Discussion

No takes yet. Share an insight, caveat, or question.

Implication

Benchmarking reveals task-specific supervised pre-training excels at predicting adverse cardiac events and mortality in patient cohorts, indicating tailored approaches yield better outcomes.

Cite This Study

Mataraso et al. (2025) studied this question.

synapsesocial.com/papers/68a3654c0a429f797332ae4c https://doi.org/10.1093/jamiaopen/ooaf090

View Full Paper

Ask AI

Bookmark

Implication

Benchmarking reveals task-specific supervised pre-training excels at predicting adverse cardiac events and mortality in patient cohorts, indicating tailored approaches yield better outcomes.

Benchmarking of pre-training strategies for electronic health record foundation models

Key Points

Discussion

Implication

Cite This Study

Benchmarking of pre-training strategies for electronic health record foundation models

Key Points

Discussion

Implication

Cite This Study

Also Consider

Also Consider