Medical Intelligence Use Cases (NLP)¶
UC-NLP-101: Generate Structured SOAP Notes¶
Purpose: Convert transcript into clinical documentation.
| Property | Value |
|---|---|
| Actor | LLM Service |
| Trigger | Transcript Available |
| Priority | P0 |
Main Success Scenario:
1. Receive transcript + context (Patient Age, Gender)
2. Construct prompt with few-shot examples
3. Call LLM (e.g., GPT-4, Med-PaLM)
4. Parse output into JSON (Subjective, Objective, Assessment, Plan)
5. Validate medical consistency (basic checks)
6. Return structured note object
Acceptance Criteria: 1. [ ] Adheres to SOAP format strictly 2. [ ] Latency < 10s 3. [ ] Handles mixed language (Hinglish) inputs
UC-NLP-102a: Extract Medications (RxNorm)¶
Purpose: Identify drugs, dosages, and instructions.
| Property | Value |
|---|---|
| Actor | NER Model (Rx) |
| Trigger | Note Generation |
| Priority | P1 |
Main Success Scenario:
1. Extract drug entities
2. Normalize to RxNorm CUI (e.g., "Metformin" -> 6809)
3. Parse structured sig: "500mg" (Strength), "BID" (Frequency)
4. Flag contraindications based on patient history
Acceptance Criteria: 1. [ ] Handles brand vs generic names 2. [ ] Extracts duration ("for 5 days")
UC-NLP-102b: Extract Diagnoses (ICD-10)¶
Purpose: Identify conditions for billing and coding.
| Property | Value |
|---|---|
| Actor | NER Model (Dx) |
| Trigger | Note Generation |
| Priority | P1 |
Main Success Scenario:
1. Extract condition terms
2. Map to ICD-10-CM codes (e.g., "T2DM" -> E11.9)
3. Determine status (Active, Chronic, Resolved)
4. Link to supporting evidence in transcript
Acceptance Criteria: 1. [ ] High specificity (avoid unspecified codes if possible) 2. [ ] Confidence score per code
UC-NLP-102c: Extract Procedures & Symptoms (SNOMED)¶
Purpose: Capture clinical findings and interventions.
| Property | Value |
|---|---|
| Actor | NER Model (Clinical) |
| Trigger | Note Generation |
| Priority | P1 |
Main Success Scenario:
1. Extract symptoms ("Chest pain") and procedures ("ECG")
2. Map to SNOMED-CT concepts
3. Capture body site and laterality (e.g., "Left leg")
Acceptance Criteria: 1. [ ] Distinguishes reported vs observed symptoms 2. [ ] Maps negation correctly ("No fever")
UC-NLP-103: Summarization + Noise Filtering¶
Purpose: Clean up the transcript and extract key info.
| Property | Value |
|---|---|
| Actor | NLP Pre-processor |
| Trigger | ASR Completion |
| Priority | P1 |
Main Success Scenario:
1. Identify and remove chit-chat ("How is the weather?")
2. Filter out interruptions (nurse entering, phone ringing)
3. Summarize long patient narratives into concise symptoms
4. Retain clinical negatives ("No chest pain")
Acceptance Criteria: 1. [ ] Does not remove clinically relevant social history 2. [ ] Preserves patient voice/intent 3. [ ] Reduces token count for downstream LLM
UC-NLP-104: EMR Field Mapping¶
Purpose: Map extracted data to specific EMR form fields.
| Property | Value |
|---|---|
| Actor | Mapping Engine |
| Trigger | Entity Extraction Complete |
| Priority | P1 |
Main Success Scenario:
1. Load EMR schema configuration
2. Map extracted entities to fields:
- "Cough for 3 days" -> Chief Complaint
- "Diabetes" -> Past History
- "Metformin 500mg" -> Medications
3. Format data types (Dates, Dropdowns)
4. Generate EMR-ready JSON payload
Acceptance Criteria: 1. [ ] Supports custom EMR templates 2. [ ] Validates mandatory fields 3. [ ] Handles multi-select values