PoC-SNOMEDCT-extract-BiomedBERT-base
Proof-of-concept -- trained on synthetic data. Every run behind this repo so far fine-tunes on a synthetically generated batch, not real MIMIC-IV-Note. Treat these weights as a proof of concept, not a clinically validated model.
BIO token classifier for SNOMED-CT span extraction (7-class: O + B/I × {finding, procedure, body_structure}),
fine-tuned from microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext.
Metrics below are multiset precision/recall/F1 against held-out validation notes (again, synthetic data).
Run config
| timestamp | 2026-08-10 02:53:54 |
| pipeline | bert-extract |
| model_id | microsoft/BiomedNLP-BiomedBERT-base-uncased-abstract-fulltext |
| max_length | 512 |
| num_epochs | 15 |
| lr | 3e-05 |
| train_batch_size | 8 |
| fast_iter | False |
| n_train_chunks | 563 |
| n_train_notes | 112 |
| n_val_chunks | 132 |
| n_val_notes | 28 |
Metrics
| scale | precision | recall | f1 | gold | pred |
|---|---|---|---|---|---|
| micro | 0.990 | 0.996 | 0.993 | 6641 | 6686 |
| macro | 0.983 | 0.995 | 0.989 | ||
| finding | 0.969 | 0.993 | 0.981 | 1484 | 1520 |
| procedure | 0.996 | 0.998 | 0.997 | 4973 | 4980 |
| body_structure | 0.984 | 0.995 | 0.989 | 184 | 186 |