PharmaCare AI

An AI-powered medication care plan generator for clinical pharmacists. Accepts patient diagnoses and medication data, generates structured care plans via Claude LLM, detects drug-drug interactions using TF-IDF similarity, stratifies patient risk with a PyTorch neural network, and audits every claim for hallucinations using an LLM-as-Judge pipeline.

---

Tech Stack

Tech	Version	Purpose
FastAPI	0.115	REST API framework
Claude (claude-sonnet-4-6)	—	Care plan generation + LLM-as-Judge
PyTorch	2.x	Patient risk stratification MLP
scikit-learn	1.4+	TF-IDF drug name fuzzy matching
SQLAlchemy	2.x	ORM for care plan persistence
PostgreSQL	16	Primary database (SQLite for local dev)
pypdf	4.x	PDF patient record extraction
Docker Compose	—	Local Postgres + app orchestration

---

Features

• Care Plan Generation — Claude generates a structured 6-section medication care plan tagged with source labels ([PATIENT], [GUIDELINE], [INFERRED]) for full clinical transparency
• LLM-as-Judge — A second Claude call audits every claim in the generated plan against the original patient data; reports Precision / Recall / F1 and flags hallucinations
• Drug-Drug Interaction (DDI) Scorer — scikit-learn TF-IDF character n-gram model fuzzy-matches patient medications against a curated 36-pair knowledge base; returns HIGH / MODERATE / LOW risk pairs with mechanism descriptions
• Patient Risk Stratification — PyTorch MLP (6 → 32 → 16 → 3) classifies patients into LOW / MODERATE / HIGH risk using ICD-10 diagnosis complexity, polypharmacy burden, and drug risk weights; MC Dropout used for uncertainty estimation
• PDF Upload — Patient records can be submitted as a PDF; text is extracted with pypdf
• Frontend UI — Single-page HTML/CSS/JS interface with risk score badge, DDI warnings panel, and collapsible verification report

---

Project Structure

PharmaCare/
├── main.py                # FastAPI app — /generate/ endpoint
├── models.py              # SQLAlchemy CarePlan model
├── database.py            # Engine + session factory
├── judge.py               # LLM-as-Judge pipeline (P/R/F1)
├── verify_careplan.py     # Rule-based code verification (ICD-10, dose regex)
├── ml/
│   ├── __init__.py
│   ├── ddi_scorer.py      # TF-IDF drug interaction scorer (36-pair KB)
│   └── risk_score.py      # PyTorch MLP risk stratification
├── public/
│   └── index.html         # Frontend UI
├── docker-compose.yml     # Postgres 16 on port 5433
├── Dockerfile
└── requirements.txt

---

ML Architecture

Drug-Drug Interaction Scorer (ml/ddi_scorer.py)

• TfidfVectorizer with character n-grams (2–4) fits over all known drug names
• Each token in the patient's medication string is matched via cosine similarity (threshold 0.45)
• All matched drug pairs are checked against the 36-pair knowledge base
• Results sorted: HIGH → MODERATE → LOW

Patient Risk Stratification (ml/risk_score.py)

Input features (6):
  num_diagnoses · num_medications · icd_risk_sum
  drug_risk · has_records · polypharmacy

Architecture:
  Linear(6 → 32) + ReLU + Dropout(0.3)
  Linear(32 → 16) + ReLU + Dropout(0.3)
  Linear(16 → 3)  → Softmax → LOW / MODERATE / HIGH

Trained on 800 synthetic patients with domain-weighted labels (drug risk weighted 55%). MC Dropout (30 forward passes averaged) provides calibrated confidence scores.

LLM-as-Judge (judge.py)

Each claim in the generated plan receives one of three verdicts:

• VERIFIED — directly supported by the patient input
• UNVERIFIED — clinically plausible but not confirmable
• HALLUCINATION — contradicts or fabricates patient-specific facts

Metrics: Precision = VERIFIED / (VERIFIED + HALLUCINATION) · Recall = ground-truth facts covered · F1 = harmonic mean

---

Getting Started

1. Clone and install dependencies

pip install -r requirements.txt

2. Configure environment

# .env
ANTHROPIC_API_KEY=sk-ant-...
DATABASE_URL=postgresql://pharmcare:pharmcare@localhost:5433/pharmcare

3. Start Postgres (Docker)

docker-compose up -d

4. Run the server

uvicorn main:app --reload   # http://localhost:8000

Local dev without Docker (SQLite)

DATABASE_URL=sqlite:///./pharmcare.db uvicorn main:app --reload

---

API

POST /generate/

Field	Type	Required	Description
patient_first_name	string	✓
patient_last_name	string	✓
patient_mrn	string	✓	6-digit MRN
referring_provider	string	✓
referring_provider_npi	string	✓	CPSO number
primary_diagnosis	string	✓	ICD-10 code + description
medication_name	string	✓	Current medication
additional_diagnoses	string	—	Newline-separated ICD-10 codes
medication_history	string	—	Newline-separated past medications
patient_records_text	string	—	Clinical notes
patient_records_file	file	—	PDF upload (alternative to text)

Response

{
  "id": 42,
  "plan": "1. Medication Review...",
  "risk": {
    "tier": "HIGH",
    "score": 0.99,
    "factors": ["Polypharmacy (4 medications)", "High-risk medication in regimen"],
    "color": "#dc2626"
  },
  "drug_interactions": [
    {
      "drug_a": "warfarin",
      "drug_b": "aspirin",
      "risk": "HIGH",
      "mechanism": "Additive anticoagulation — major bleeding risk",
      "confidence": 1.0
    }
  ],
  "verification": {
    "status": "verified",
    "metrics": { "precision": 1.0, "recall": 0.875, "f1": 0.933 },
    "claims": [ { "text": "...", "verdict": "VERIFIED", "reason": "..." } ]
  }
}