UMCU/CardioMedRoBERTa.nl

Continued, on-premise, pre-training of MedRoBERTa.nl using de-identified Electronic Health Records from the University Medical Center Utrecht, related to the cardiology domain.

Data statistics

Sources:

• Dutch medical guidelines (FMS, NHG)

• NtvG papers

• PMC abstracts translated using GeminiFlash 1.5

• Number of tokens: 1.47B, of which 1B from UMCU EHRs

• Number of documents: 5.8M, of which 3.5M UMCU EHRs

• Average number of tokens per document: 253

• Median number of tokens per document: 124

Training

• Effective batch size: 240
• Learning rate: 1e-4
• Weight decay: 1e-3
• Learning schedule: linear, with 25_000 warmup steps
• Num epochs: 3

Train perplexity: 3.0

Validation perplexity: 4.0

Acknowledgements

This work was done together with the Amsterdam UMC, in the context of the DataTools4Heart project.