AEGIS Model Training
Back to HomeAEGIS uses local large language models (LLMs) for clinical classification. Your expert reviews create high-quality training data that improves these models over time, resulting in faster and more accurate classifications tailored to your institution.
How It Works
LLM Classification
Clinical cases are processed by our 70B parameter model running locally. The model extracts relevant information from clinical notes and makes initial classifications.
Expert Review
IP/ASP team members review each classification, confirming correct decisions or providing corrections with clinical reasoning.
Model Fine-Tuning
After collecting ~500 expert-reviewed cases, we fine-tune a smaller, faster model specifically for your institution's patterns.
Training Data Collection
HAI Detection
Collecting DataWhat we capture:
- Clinical note text - The relevant portions of notes used for classification
- LLM extraction - Extracted clinical findings (symptoms, device info, lab values)
- LLM classification - The model's initial HAI determination (Confirmed/Suspected/Ruled Out)
- Expert decision - Your confirmed classification
- Override reason - When you correct the LLM, why it was wrong:
- Extraction error - Model missed or misread clinical data (symptoms, dates, sources)
- Rules error - Model applied NHSN criteria incorrectly
This feedback helps the model learn both what to extract and how to apply surveillance definitions.
Antibiotic Indications
Collecting DataWhat we capture:
- Clinical context - Relevant notes, labs, and diagnoses at time of antibiotic order
- LLM extraction - Extracted clinical syndrome and supporting evidence
- Syndrome decision - Your validation of the indication:
- Confirm syndrome - LLM identified the correct clinical syndrome
- Correct syndrome - You specify the actual syndrome (training signal)
- No indication - No documented reason for antibiotics
- Viral illness / ASB - Common inappropriate indications
- Agent appropriateness - Whether the specific antibiotic was appropriate for the syndrome
Syndrome corrections are particularly valuable - they teach the model to recognize clinical patterns it missed.
Guideline Adherence
Collecting DataWhat we capture:
- Clinical notes - Relevant notes used for clinical appearance assessment
- Tiered extraction - Fast triage (7B model) and full analysis (70B model) results:
- Triage decision - Clear well, clear ill, or needs full analysis
- Escalation reasons - Why triage escalated to full model
- Response times - Performance metrics for optimization
- Appearance decision - Your validation of the clinical appearance:
- Confirm - LLM assessment was correct
- Override to Well/Ill/Toxic - Correct the appearance classification
- Missed findings - Specific clinical signs the LLM failed to identify (lethargy, mottling, poor feeding, etc.)
- Override reasons - Why the LLM was wrong (missed lethargy, overinterpreted, documentation ambiguous, etc.)
This feedback trains the model to better recognize clinical appearance indicators in febrile infant documentation.
Why Local Models?
Data Privacy
PHI never leaves your network. All processing happens on local infrastructure.
Speed
Fine-tuned smaller models run faster than general-purpose large models.
Accuracy
Models trained on your data learn your institution's documentation patterns.
Cost
No per-query API costs. Run unlimited classifications on your hardware.
Training Progress
Progress updates automatically as you review cases. Fine-tuning begins after reaching 500 expert-reviewed cases per module.