MedHELM's gated MIMIC-IV task: extract ICD-10 codes from an English discharge note and score micro-F1 against gold codes.
unassessed
| Category | domain |
|---|---|
| Subcategory | ICD-10 billing-code extraction from discharge notes |
| Page status | active |
| Metric | mimiciv_billing_code_f1 (micro-F1 over extracted ICD-10 codes) |
| Direction | higher_is_better |
| Publisher | Stanford CRFM and collaborators (MedHELM); notes from MIMIC-IV |
This id is HELM's mimiciv_billing_code scenario, not MIMIC-IV-BHC summarization and not MedConceptsQA. The model reads an English clinical note (column text) and must list ICD-10 billing codes (column target). HELM extracts code-shaped tokens from the generation and compares them to gold as a multi-label set. Inputs are English clinical text. Outputs are code strings.
Zero-shot generation. Instruction: "Given the following clinical note, identify all relevant ICD-10 codes." Input noun Note, output noun Predicted ICD-10 Codes, max_tokens 256, max_train_instances 0. Some MedHELM gated entries raise num_output_tokens to 4000. The scenario tags every row as TEST_SPLIT.
No model card in ModelSpec reports this benchmark yet.