GLiNER-BioMed: Active-Learned Clinical Entity & Relation Extractor
-brightgreen){kind=icon}
A high-performance GLiNER-BioMed model fine-tuned on 75,491 gold clinical entity mentions from EHR narratives, specialized for open-vocabulary and zero-shot clinical Named Entity Recognition (NER) across the SNOMED CT taxonomy.
Coupled with a Closed-Loop Active Learning Comparator and Clinical RelEC (Relation Extraction), the model achieves state-of-the-art boundary extraction and clinical relation discovery without requiring GPU hardware for inference.
🌟 Key Capabilities & Updates
- Active Learning Span Resolution: Dual-tier learning loop that ingests character IoU boundary shifts and false negatives from gold standard labels, resolving 78.8% of previously missed entities in real-time.
- Clinical RelEC Integration: Automatically connects extracted entities into directed clinical relations (
treats,caused_by,indicated_for,anatomical_site,evaluates). - MIMIC-IV Zero-Shot Generalization: Successfully extracts and resolves complex clinical shorthand on unseen inpatient notes (e.g., "SBP" $ ightarrow$ Primary bacterial peritonitis, "viral load", "d/c").
- Fast CPU AVX2 Parallelism: Optimized inference latency (~120ms per clinical note on standard multi-core CPUs).
📊 Evaluation & Benchmarks
Validation Partition Performance (Gold-Standard Annotations)
Evaluated across diverse clinical categories against ground truth:
| Metric | Score | Clinical Significance |
|---|---|---|
| Recall | 86.52% | Captures 854/987 ground truth clinical mentions |
| Macro-IoU | 0.8412 | High boundary accuracy on compound multi-word surgical & disorder spans |
| Precision | 69.26% | High sensitivity for exploratory clinical document review |
| F1 Score | 76.93% | Balanced harmonic precision-recall |
Category-Specific Recall
- Clinical Disorders:
91.62%(Diagnoses, diseases, chronic conditions) - Surgical Procedures:
89.88%(Surgeries, interventions, biopsies, endoscopic exams) - Clinical Findings:
82.83%(Symptoms, signs, physical exam observations) - Body Structures:
77.11%(Organs, anatomical locations)
🏥 Supported Entity Classes
clinical disorder: Pathologies, infections, conditions (e.g., acute necrotizing pancreatitis, essential hypertension).clinical finding: Symptoms and signs (e.g., severe epigastric pain, shortness of breath, diaphoresis).surgical procedure: Operations and interventions (e.g., laparoscopic cholecystectomy, common bile duct exploration).body structure: Anatomical sites (e.g., gallbladder, bile duct, myocardium).medication: Therapeutic drugs and dosages (e.g., Lisinopril 20mg, Cefepime 2g IV).diagnostic measurement: Lab tests and measurements (e.g., serum lipase, troponin, WBC count).
🚀 Quickstart: Python Inference
from gliner import GLiNER
# 1. Load model from Hugging Face Hub
model = GLiNER.from_pretrained("gitmodelmujtaba/gliner-snomed-biomed")
# 2. Clinical narrative
text = """
The patient is a 64-year-old female presenting with acute biliary pancreatitis secondary to cholelithiasis.
She was admitted for severe epigastric abdominal pain radiating to the back.
She subsequently underwent emergent laparoscopic cholecystectomy with common bile duct exploration.
Post-operatively, she was initiated on Cefepime 2g IV q8h and Lisinopril 20mg daily for essential hypertension.
"""
# 3. Predict entities
labels = [
"clinical disorder",
"clinical finding",
"surgical procedure",
"body structure",
"medication",
"diagnostic measurement",
]
entities = model.predict_entities(text, labels=labels, threshold=0.30)
for ent in entities:
print(f"[{ent['label'].upper()}] {ent['text']} (score: {ent['score']:.2f})")
🔗 Connected Repositories & Spaces
- Interactive Hugging Face Space: Clinical GLiNER & SapBERT RelEC + Graph-RAG
- Entity Linking Partner Model: SapBERT Tri-Linker
- Author: Mujtaba Hussain (
gitmodelmujtaba) - License: Apache 2.0