Model Card for GENIE
Model Details
Model Size: 8B (English)
Max Tokens: 8192
Base model: Llama 3.1 8B (English)
Model Description
GENIE (Generative Note Information Extraction) is an end-to-end model designed to structure free text from electronic health records (EHRs). It processes EHRs in a single pass, extracting biomedical named entities along with their assertion statuses, body locations, modifiers, values, units, and intended purposes, outputting this information in a structured JSON format. This streamlined approach simplifies traditional natural language processing workflows by replacing all the analysis components with a single model, making the system easier to maintain while leveraging the advanced analytical capabilities of large language models (LLMs). Comparing with general-purpose LLMs, GENIE does not require prompt engineering or few-shot examples. Additionally, it generates all relevant attributes in one pass, significantly reducing both runtime and operational costs. GENIE is co-developed by the groups of Sheng Yu (https://www.stat.tsinghua.edu.cn/teachers/shengyu/), Tianxi Cai (https://dbmi.hms.harvard.edu/people/tianxi-cai), and Isaac Kohane (https://dbmi.hms.harvard.edu/people/isaac-kohane).
Usage
from vllm import LLM, SamplingParams
model = LLM(model='THUMedInfo/GENIE_en_8b', tensor_parallel_size=1)
#model = LLM(model=path/to/your/local/model, tensor_parallel_size=1)
PROMPT_TEMPLATE = "Human:\n{query}\n\n Assistant:"
sampling_params = SamplingParams(temperature=temperature, max_tokens=max_new_token)
EHR = ['xxxxx1','xxxxx2']
texts = [PROMPT_TEMPLATE.format(query=k) for k in EHR]
output = model.generate(texts, sampling_params)
res = json.loads(output[0].outputs[0].text)
An example:
Input:
EHR = ["""Unit No:___
Admission Date:___
Discharge Date:___
Date of Birth:___
Sex: F
Service: MEDICINE
Allergies:
Sulfur / Norvasc
Attending:___
Addendum:
See below
Chief Complaint:
abdominal pain
Major Surgical or Invasive Procedure:
none
History of Present Illness:
84 F with PMHx of Renovascular HTN c/b NSTEMI now s/p renal
stents, Gout and h/o Crohn's disease who presented to the ED on
___with RLQ pain for approx 2 days. She denies any
nausea/vomiting/diarrhea or constipation but has not been taking
po well and felt dehydrated."""]
Output:
res = [{'phrase': 'allergies',
'semantic_type': 'Disease, Syndrome or Pathologic Function',
'assertion_status': 'title',
'body_location': 'null',
'modifier': 'null',
'value': 'not applicable',
'unit': 'not applicable',
'purpose': 'not applicable'},
{'phrase': 'sulfur',
'semantic_type': 'Chemical or Drug',
'assertion_status': 'conditional',
'body_location': 'not applicable',
'modifier': 'not applicable',
'value': 'null',
'unit': 'units: null',
'purpose': 'null'},
{'phrase': 'norvasc',
'semantic_type': 'Chemical or Drug',
'assertion_status': 'conditional',
'body_location': 'not applicable',
'modifier': 'not applicable',
'value': 'null',
'unit': 'units: null',
'purpose': 'null'},
{'phrase': 'abdominal pain',
'semantic_type': 'Sign, Symptom, or Finding',
'assertion_status': 'present',
'body_location': 'Abdominal',
'modifier': 'null',
'value': 'not applicable',
'unit': 'not applicable',
'purpose': 'not applicable'},
{'phrase': 'surgical or invasive procedure',
'semantic_type': 'Therapeutic or Preventive Procedure',
'assertion_status': 'title',
'body_location': 'null',
'modifier': 'not applicable',
'value': 'not applicable',
'unit': 'not applicable',
'purpose': 'null'},
{'phrase': 'renovascular hypertension',
'semantic_type': 'Disease, Syndrome or Pathologic Function',
'assertion_status': 'present',
'body_location': 'renal',
'modifier': 'null',
'value': 'not applicable',
'unit': 'not applicable',
'purpose': 'not applicable'},
{'phrase': 'non-st elevation myocardial infarction',
'semantic_type': 'Disease, Syndrome or Pathologic Function',
'assertion_status': 'present',
'body_location': 'null',
'modifier': 'null',
'value': 'not applicable',
'unit': 'not applicable',
'purpose': 'not applicable'},
{'phrase': 'gout',
'semantic_type': 'Disease, Syndrome or Pathologic Function',
'assertion_status': 'present',
'body_location': 'null',
'modifier': 'null',
'value': 'not applicable',
'unit': 'not applicable',
'purpose': 'not applicable'},
{'phrase': "crohn ' s disease",
'semantic_type': 'Disease, Syndrome or Pathologic Function',
'assertion_status': 'present',
'body_location': 'not applicable',
'modifier': 'not applicable',
'value': 'not applicable',
'unit': 'not applicable',
'purpose': 'not applicable'},
{'phrase': 'emergency department',
'semantic_type': 'Therapeutic or Preventive Procedure',
'assertion_status': 'present',
'body_location': 'null',
'modifier': 'not applicable',
'value': 'not applicable',
'unit': 'not applicable',
'purpose': 'null'},
{'phrase': 'pain',
'semantic_type': 'Sign, Symptom, or Finding',
'assertion_status': 'present',
'body_location': 'right lower quadrant',
'modifier': 'null',
'value': 'not applicable',
'unit': 'not applicable',
'purpose': 'not applicable'},
{'phrase': 'nausea',
'semantic_type': 'Sign, Symptom, or Finding',
'assertion_status': 'absent',
'body_location': 'null',
'modifier': 'null',
'value': 'not applicable',
'unit': 'not applicable',
'purpose': 'not applicable'},
{'phrase': 'vomiting',
'semantic_type': 'Sign, Symptom, or Finding',
'assertion_status': 'absent',
'body_location': 'null',
'modifier': 'null',
'value': 'not applicable',
'unit': 'not applicable',
'purpose': 'not applicable'},
{'phrase': 'diarrhea',
'semantic_type': 'Sign, Symptom, or Finding',
'assertion_status': 'absent',
'body_location': 'null',
'modifier': 'null',
'value': 'not applicable',
'unit': 'not applicable',
'purpose': 'not applicable'},
{'phrase': 'constipation',
'semantic_type': 'Sign, Symptom, or Finding',
'assertion_status': 'absent',
'body_location': 'null',
'modifier': 'null',
'value': 'not applicable',
'unit': 'not applicable',
'purpose': 'not applicable'}]
Citation
If you find our paper or models helpful, please consider cite:
BibTeX:
@misc{ying2025geniegenerativenoteinformation,
title={GENIE: Generative Note Information Extraction model for structuring EHR data},
author={Huaiyuan Ying and Hongyi Yuan and Jinsen Lu and Zitian Qu and Yang Zhao and Zhengyun Zhao and Isaac Kohane and Tianxi Cai and Sheng Yu},
year={2025},
eprint={2501.18435},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2501.18435},
}