Docs / Raw

Medical Features

Sourced from docs/MEDICAL_FEATURES.md

Edit on GitHub

Medical Features Documentation

Overview

VoiceAssist includes specialized medical capabilities designed for healthcare professionals, with a focus on evidence-based information retrieval, clinical decision support, and privacy-conscious handling of medical data.

Core Medical Features

1. Medical Textbook Knowledge Base

Concept

Pre-loaded medical textbooks are indexed and available for semantic search with precise citations.

Supported Textbooks (Planned)

  • Harrison's Principles of Internal Medicine
  • Robbins and Cotran Pathologic Basis of Disease
  • Williams Obstetrics
  • Nelson Textbook of Pediatrics
  • Specialty-specific textbooks (customizable)
  • UpToDate (if subscription available)

How It Works

  1. PDF textbooks are uploaded to the admin panel
  2. Text is extracted and OCR'd if needed
  3. Content is chunked by section/paragraph with page tracking
  4. Embeddings generated and stored in vector database
  5. Metadata includes: book name, edition, chapter, page number, section title

Example Queries

  • "What does Harrison's say about diabetic ketoacidosis management?"
  • "According to Robbins, what are the pathological features of atherosclerosis?"
  • "What's the recommended treatment for preeclampsia in Williams Obstetrics?"

Response Format

According to Harrison's Principles of Internal Medicine, 21st Edition,
Chapter 420 (Diabetes Mellitus), page 2987:

"Diabetic ketoacidosis (DKA) is characterized by hyperglycemia,
metabolic acidosis, and increased total body ketone concentration..."

[Full relevant excerpt]

Would you like me to read more from this section or explore related topics?

Features

  • Exact page citations
  • Multi-book cross-referencing
  • "Read more" option to get additional context
  • Voice narration of text sections
  • Bookmark frequently referenced sections

2. Medical Journal Search & Retrieval

Databases Supported

  • PubMed/MEDLINE: Primary source for biomedical literature
  • PubMed Central (PMC): Open-access full-text articles
  • OpenEvidence: Evidence summaries and clinical questions
  • Direct journal access (with institutional credentials if available)

Search Capabilities

Natural Language Queries:

  • "Find recent papers on GLP-1 agonists for heart failure"
  • "What's the latest evidence on early goal-directed therapy for sepsis?"
  • "Show me systematic reviews about omega-3 fatty acids and cardiovascular outcomes"

Advanced Filters:

  • Publication date range
  • Article type (RCT, meta-analysis, review, case report)
  • Journal impact factor
  • Study population
  • Sample size

Automatic Ranking:

  • Relevance to query
  • Study quality (based on type and journal)
  • Recency
  • Citation count

PDF Download & Processing

Workflow:

  1. User requests journal article
  2. System searches PubMed for matches
  3. Displays results with abstracts
  4. User selects article(s) to download
  5. System attempts download:
    • PMC open access (direct)
    • Institutional access (via credentials)
    • Direct journal link
    • Alternative sources (as legally permitted)
  6. PDF is downloaded, processed, and indexed
  7. Full text becomes searchable

Processing Steps:

  • OCR for image-based PDFs
  • Extract text, figures, tables
  • Parse sections (abstract, methods, results, discussion)
  • Generate embeddings
  • Store with metadata (DOI, authors, journal, year, type)

Storage:

  • PDFs saved on Ubuntu server
  • Optionally backed up to Nextcloud
  • Organized by topic/specialty folders

Example Use Cases

Research Question: "What's the current evidence on dapagliflozin for heart failure?"

System Response:

  1. Searches PubMed for relevant papers
  2. Finds DAPA-HF trial, meta-analyses, recent reviews
  3. Downloads key papers
  4. Synthesizes findings:
Based on 15 recent publications including 3 large RCTs:

DAPA-HF Trial (McMurray et al., NEJM 2019):
- 4,744 patients with HFrEF
- Dapagliflozin reduced CV death or HF hospitalization by 26% (HR 0.74)
- NNT = 21 over 18 months

[Additional studies summarized]

Key Takeaway: Strong evidence supports dapagliflozin for HFrEF,
regardless of diabetes status. Class 1A recommendation in
2021 ESC guidelines.

Would you like me to download the full DAPA-HF paper or review the guideline?

3. Clinical Guidelines Access

Guideline Sources

  • CDC: Disease prevention and control guidelines
  • WHO: International health recommendations
  • Specialty Societies:
    • American Heart Association (AHA)
    • American Diabetes Association (ADA)
    • American College of Cardiology (ACC)
    • Infectious Diseases Society of America (IDSA)
    • Many others
  • National guidelines (NICE, SIGN, etc.)

Indexing Strategy

  • Scrape official guideline PDFs/web pages
  • Index by disease/condition
  • Track guideline updates
  • Flag when new versions are released
  • Compare old vs new recommendations

Query Examples

  • "What's the current AHA guideline for hypertension management?"
  • "CDC recommendations for COVID-19 post-exposure prophylaxis"
  • "IDSA guidelines for community-acquired pneumonia"

Response Features

  • Guideline year and version
  • Strength of recommendation (1A, 2B, etc.)
  • Quality of evidence
  • Key changes from previous version
  • Link to full guideline

4. OpenEvidence Integration

What is OpenEvidence?

AI-powered clinical decision support system trained on medical evidence.

Integration Approach

  • API calls to OpenEvidence for clinical questions
  • Supplement with local knowledge base
  • Compare OpenEvidence summary with direct literature review
  • Provide both synthesized answer and source citations

Example Query

"Is anticoagulation indicated for atrial fibrillation with a CHA2DS2-VASc score of 1?"

OpenEvidence Response:

Evidence Summary (via OpenEvidence):
For males with CHA2DS2-VASc score of 1, anticoagulation may be considered
but is not mandated. Individualize based on bleeding risk and patient preference.

Direct Evidence (from local knowledge base):
2020 ESC AF Guidelines recommend:
- CHA2DS2-VASc ≥ 2 (males): Oral anticoagulation recommended (Class I)
- CHA2DS2-VASc = 1 (males): Consider anticoagulation (Class IIa)

Supporting Studies:
[List of relevant RCTs with outcomes]

Recommendation: Discuss risks/benefits with patient. Consider if
bleeding risk (HAS-BLED) is low.

5. Medical Calculators & Tools

Integrated Calculators (Planned)

  • CHA2DS2-VASc (stroke risk in AF)
  • HAS-BLED (bleeding risk)
  • ASCVD Risk Calculator
  • GFR/CKD-EPI
  • APACHE II (ICU mortality)
  • CURB-65 (pneumonia severity)
  • Wells Score (DVT/PE probability)
  • Many others via MDCalc integration

Usage

Voice command: "Calculate CHA2DS2-VASc for a 68-year-old male with hypertension and diabetes"

System prompts for missing variables, calculates score, provides interpretation and recommendations.

6. Drug Information & Interactions

Features (Planned)

  • Drug dosing information
  • Renal/hepatic adjustments
  • Drug-drug interactions (via API or database)
  • Pregnancy/lactation categories
  • Common adverse effects

Sources

  • FDA labels
  • Lexicomp or Micromedex API
  • Open drug databases

7. Differential Diagnosis Assistant

Concept

Present clinical scenario, get differential diagnosis list with reasoning.

Implementation

  • Leverage GPT-4's medical reasoning
  • Cross-reference with UpToDate/textbooks
  • Present likelihood and key differentiating features
  • Suggest diagnostic workup

Example: "42-year-old male with acute chest pain radiating to left arm, diaphoresis, no trauma"

Response:

Differential Diagnosis (in order of urgency):

1. Acute Coronary Syndrome (MI/Unstable Angina) - HIGH PRIORITY
   - Classic presentation
   - Immediate: ECG, troponin, aspirin

2. Aortic Dissection
   - Less common but life-threatening
   - Check BP both arms, CXR, consider CT angio

3. Pulmonary Embolism
   - Consider if risk factors present
   - Wells score, D-dimer if low probability

[Additional differentials with reasoning]

Recommended Immediate Actions:
- 12-lead ECG
- IV access
- High-flow oxygen if hypoxic
- Aspirin 325mg if no contraindications
- Serial troponins
- Activate cath lab if STEMI

Would you like the full workup algorithm or medication dosing?

8. Case-Based Learning (Future)

Features

  • Present clinical cases
  • User works through diagnosis/management
  • System provides feedback and teaching points
  • Reference to guidelines and evidence

Privacy & Compliance

HIPAA Considerations

Safe Practices:

  • Never include patient names, MRNs, or identifiable information in queries
  • Use "de-identified" case presentations
  • Local processing of any PHI
  • Audit logs for compliance

Example Safe Query: "Management of 55-year-old with new-onset atrial fibrillation and CKD stage 3"

Example UNSAFE Query: "What should I do for John Smith, MRN 123456, who has AF?"

Data Flow Rules

  1. General medical knowledge queries → Cloud APIs OK
  2. Patient-specific with PHI → Local processing ONLY
  3. De-identified cases → Cloud OK with caution
  4. Personal medical info → Local only

User Training

  • Clear documentation on privacy
  • Warning prompts if PHI detected
  • Option to force local processing

Medical Knowledge Update Cadence

Continuous Updates

  • PubMed alerts for specified topics
  • Weekly scan for new high-impact publications
  • Automatic download and indexing of relevant papers

Periodic Reviews

  • Monthly guideline update checks
  • Quarterly textbook supplement review
  • Annual major textbook edition updates

User-Triggered

  • Manual search and add
  • Topic deep-dives on demand
  • Specialty focus area curation

Quality Assurance

Citation Verification

  • Every medical claim must have source
  • Page numbers for textbooks
  • DOI/PMID for journals
  • URL for guidelines
  • No "hallucinated" references

Accuracy Checks

  • Cross-reference multiple sources
  • Flag conflicting information
  • Provide date of information
  • Disclaimer: "Verify with primary sources for clinical decisions"

Disclaimers

System should include:

This system provides information for educational and reference purposes.
It is not a substitute for professional medical judgment. Always verify
critical information with primary sources and current guidelines. Medical
knowledge evolves rapidly - confirm latest evidence before clinical application.

Use Case Scenarios

Scenario 1: Pre-Clinic Preparation

"I have a patient with resistant hypertension coming in. Remind me of the workup for secondary causes."

System provides:

  • Checklist of secondary causes
  • Recommended lab tests and imaging
  • Algorithm from JNC-8/AHA guidelines
  • Recent papers on approach

Scenario 2: In-Clinic Quick Reference

"What's the dosing for apixaban in atrial fibrillation with CKD?"

System provides:

  • Standard dosing (5mg BID)
  • Renal adjustments (2.5mg BID if Cr >1.5 and age >80 or weight <60kg)
  • Source citation
  • Monitoring recommendations

Scenario 3: Literature Review for Grand Rounds

"Find the top 10 papers on immunotherapy for melanoma from the past 2 years"

System:

  • Searches PubMed with filters
  • Ranks by impact and relevance
  • Downloads PDFs
  • Generates summary of key findings
  • Creates bibliography

Scenario 4: After-Hours Question

Voice: "Hey assistant, I have a patient with suspected NMS from haloperidol. What's the management?"

System (voice response):

  • Confirms neuroleptic malignant syndrome features
  • Management algorithm (stop drug, supportive care, dantrolene/bromocriptine)
  • ICU admission criteria
  • Prognosis information
  • Offers to text detailed protocol

Future Enhancements

Clinical Decision Support

  • Integration with EMR (FHIR API)
  • Real-time alerts (drug interactions, contraindications)
  • Order set suggestions

Continuing Medical Education

  • CME credit tracking
  • Quiz generation from literature
  • Spaced repetition for retention

Research Assistant

  • Literature review automation
  • Data extraction from papers
  • Meta-analysis support
  • Bibliography generation

Teaching Tool

  • Medical student/resident education mode
  • Socratic questioning
  • Board exam preparation

Technical Implementation Notes

Embedding Models

  • Use specialized medical embeddings if available (BioGPT, PubMedBERT)
  • Or fine-tune OpenAI embeddings on medical corpus
  • Benchmark retrieval accuracy

Chunking Strategy

  • Textbooks: By section/subsection (preserve context)
  • Journal articles: By paragraph with section labels
  • Guidelines: By recommendation statement
  • Overlap chunks to avoid boundary issues

Metadata Schema

{
  "id": "uuid",
  "type": "textbook|journal|guideline",
  "source": "Harrison's 21st Ed",
  "title": "Diabetic Ketoacidosis",
  "chapter": "420",
  "page": "2987",
  "section": "Treatment",
  "date": "2022",
  "specialty": ["Endocrinology", "Internal Medicine"],
  "keywords": ["DKA", "diabetes", "ketoacidosis"],
  "embedding": [0.123, ...],
  "content": "Full text chunk"
}

Performance Optimization

  • Cache common queries
  • Pre-compute embeddings for frequently accessed content
  • Hybrid search (vector + keyword) for best results
  • Pagination for large result sets

Ethical Considerations

  • Ensure equity in knowledge representation (not just Western medicine)
  • Acknowledge limitations of AI in medical decision-making
  • Maintain human physician as ultimate authority
  • Transparent about sources and confidence levels
  • Regular bias audits of recommendations

Regulatory Compliance

  • FDA consideration: Not a medical device (information only)
  • HIPAA: No PHI in cloud
  • malpractice insurance: Confirm coverage for AI-assisted decision-making
  • Document limitations prominently

Note: Medical features should be developed and validated carefully with emphasis on accuracy and patient safety. Consider consulting with medical informaticists and legal advisors during implementation.

← All documentationHome
Beginning of guide
End of guide