TermMap - strawhats
Offline mapping of an extracted clinical entity to one ICD-10-CM and one SNOMED CT US code. This is a reproducible competition baseline, not a 95%-accurate solution.
Measured results
Complete supplied three-fold validation, 9,644 retained labels:
| System | Selected method | Exact top-1 |
|---|---|---|
| ICD-10-CM | Learned context ranker | 47.87% |
| SNOMED CT US | Frozen dense retrieval | 38.71% |
| Headline: lower of the two | 38.71% |
The same complete OOF comparison selected dense versus learned ranking per system. This is model-selection validation, not a hidden-test score. The learned SNOMED ranker scored 36.40% and was not promoted. No rows are removed to inflate accuracy. The two invalid ICD source labels excluded by the supplied fold alignment are documented in the source report.
Weights and setup
termmap_model_weights.zip: frozen model/index state, final classifiers, hashed example keys and support vectors. No raw clinical notes or organizer dictionary tables are included.MODEL_WEIGHTS_DOWNLOAD.json: artifact SHA-256 and scope.- The two final
*_final_ranker.joblibfiles are also provided individually. The SNOMED classifier is retained for reproduction of the measured comparison; the selected SNOMED prediction path uses dense retrieval. - Source, full setup instructions and prediction entry point: TeamStrawhats_hilabs.
Setup uses a separate preparation environment with pandas==3.0.1, pyarrow==21.0.0 and openpyxl==3.1.5 to reproduce pinned dictionary bytes; inference uses the resolved runtime dependencies. See the source README for the complete commands.
Use the organizer-supplied original icd10cm_codes.xlsx and SNOMEDCT_US_codes_and_descriptions.xlsx to reconstruct the three excluded static dictionary metadata tables. Raw clinical training text is not needed for deployment.
from huggingface_hub import hf_hub_download
weights = hf_hub_download("Vorun/strawhats-termmap", "termmap_model_weights.zip")
python -m genlink.setup_runtime --weights PATH_TO_ZIP --source PATH_TO_DICTIONARIES --output artifacts/runtime
termmap predict --artifacts artifacts/runtime --input input.csv --output predictions.csv
Input accepts entity or clinical_entity with chunk_text. Use --chunks for a chunk_id lookup, or explicitly --contextless for entity-only data. Outputs preserve order and identifiers and append icd10cm_code,snomed_code. Separate debug records contain ranked candidates. Scores are uncalibrated ranking signals.
Pretrained encoder pointer
The encoder is frozen, not trained by this team:
- cambridgeltl/SapBERT-from-PubMedBERT-fulltext
- Revision:
090663c3ae57bf35ffe4d0d468a2a88d03051a4d - The upstream model card lists Apache-2.0 and UMLS 2020AA training provenance. Its original files and hashes are retained in the weight bundle.
Only the lightweight gradient-boosted classifiers were fitted in the final experiment. Earlier stopped Qwen experiments are excluded. Dictionaries and challenge resources remain subject to their source terms.
Supporting assets
Submission report and 2:30 silent demo walkthrough. The video uses actual captured application screens, paced for reading. Both describe measured behavior and limitations. No API keys are required for prediction or included in the source/model artifacts.