hilabs-termmap models

Trained models for hilabs-termmap, a context-aware clinical term mapping engine: given a pre-extracted clinical term (plus optional chart context), predict one ICD-10-CM code and one SNOMED CT concept.

All models are initialized from cambridgeltl/SapBERT-from-PubMedBERT-fulltext and fine-tuned on the hackathon training split (9,646 rows; per-system).

Contents

subfolder	task	pooling
biencoder_ICD10CM	ICD-10-CM dense retrieval (bi-encoder)	mean
biencoder_SNOMEDCT_US	SNOMED CT dense retrieval (bi-encoder)	mean
reranker_ICD10CM	ICD-10-CM cross-encoder reranker (1 logit)	CLS
reranker_SNOMEDCT_US	SNOMED CT cross-encoder reranker (1 logit)	CLS

The engine also uses a frozen second dense source, FremyCompany/BioLORD-2023 (for recall only), and TF-IDF/char/fuzzy indexes built from the dictionaries.

Usage

from transformers import AutoModel, AutoTokenizer, AutoModelForSequenceClassification

# bi-encoder (mean pooling over tokens)
model = AutoModel.from_pretrained("d3let3/hilabs-termmap", subfolder="biencoder_ICD10CM")
tok = AutoTokenizer.from_pretrained("d3let3/hilabs-termmap", subfolder="biencoder_ICD10CM")

# cross-encoder reranker: (term + chart context, candidate description) -> relevance logit
rer = AutoModelForSequenceClassification.from_pretrained(
    "d3let3/hilabs-termmap", subfolder="reranker_ICD10CM"
)
tok = AutoTokenizer.from_pretrained(
    "d3let3/hilabs-termmap", subfolder="reranker_ICD10CM"
)

Full engine, indexes and reproduction instructions: see the project repository (hilabs-termmap).

Results (held-out dev, entity-level 85/15 split)

system	exact match
ICD-10-CM	67.9%
SNOMED CT	51.6%

Headline (lower of the two): 51.6%. Long-tail (gold code unseen in training): 36.8% / 40.6%.

Notes

• These weights are research artifacts from a one-day hackathon; they are not a medical device and must not be used for clinical decisions.
• No patient data is contained in the weights; training used a public hackathon dataset of clinical terms.