Palakdwivedi1706/glaucoma-rag-agent

OculiRAG: Glaucoma Agentic RAG & Decision Framework

OculiRAG is an interactive clinical decision support web application and API prototype. It demonstrates how Retrieval-Augmented Generation (RAG) can be integrated into an agentic Chain-of-Thought (CoT) reasoning loop to solve complex diagnostic tasks in ophthalmology.

Specifically, it ingest patient physiological parameters (measured intraocular pressure (IOP), central corneal thickness (CCT), cup-to-disc ratio (CDR), family history, and age), performs a local TF-IDF semantic query against standard clinical guidelines (EGS & AAO consensus guidelines), runs a multi-agent logic flow to correct IOP measurements, weights risk markers, and predicts a safe Target IOP range and Glaucoma Progression Risk.

🏗️ Architecture & Reasoning Pipeline

The system models a clinical care team composed of four cooperative agents:

[Patient Data & Retinal Disc]
        │
        ▼
 ┌──────────────┐
 │ Triage Agent │ ──► Extracts structural parameters & establishes risk alerts
 └──────┬───────┘
        │ Keywords
        ▼
 ┌─────────────────┐
 │ Retrieval Agent │ ──► Executes TF-IDF Search on Guidelines Document Store (RAG)
 └──────┬──────────┘
        │ Guidelines Citations
        ▼
 ┌───────────────────────────┐
 │ Clinical Reasoning Agent  │ ──► Chain-of-Thought math (CCT correction) & risk scoring
 └──────┬────────────────────┘
        │ Corrected IOP & Score
        ▼
 ┌─────────────────┐
 │ Consensus Agent │ ──► Formulates clinical report, Target IOP, and risk class
 └────────┬────────┘
          │
          ▼
   [Web Dashboard]

🔬 Core Clinical Rules & Formulations

1. 1.Central Corneal Thickness (CCT) Correction: Goldmann Applanation Tonometry assumes a baseline corneal thickness of 545 µm. Thinner corneas artificially underestimate pressure (masking glaucoma risk), while thicker corneas overestimate it. The framework applies a linear correction: $$\text{IOP}{\text{corrected}} = \text{IOP}{\text{measured}} + \frac{545 - \text{CCT}}{10} \times 0.6 \, \text{mmHg}$$

1. 1.Progression Risk Score ($R \in [0, 15]$): Integrates corrected IOP, vertical Cup-to-Disc ratio, CCT boundaries, family history, and systemic vasospastic indicators (e.g. Migraines, Raynaud's).

1. 1.Target IOP Decision Bounds: Target IOP thresholds are determined dynamically:
2. 2.High Risk ($R \ge 8$): Target IOP is set to at least 30% below corrected baseline, capped at 14 mmHg.
3. 3.Moderate Risk ($5 \le R < 8$): Target IOP is set to 20-25% below baseline, capped at 18 mmHg.
4. 4.Mild/Low Risk ($R < 5$): Monitor baseline target (<20 mmHg).

🛠️ Getting Started

📋 Prerequisites

Ensure you have Python 3.8+ installed. Install the backend dependencies:

pip install fastapi uvicorn scikit-learn numpy pydantic

(Note: These dependencies are typically already present in science and medical environments).

🚀 Running the Application

1. 1.Navigate to the project root directory:

   cd E:\glaucoma-rag-agent

1. 1.Start the FastAPI backend server:

   python backend/main.py

The server will start running on [http://localhost:8000](http://localhost:8000).

1. 1.Access the Web Dashboard: Open your web browser and navigate to:

   http://localhost:8000/

(FastAPI automatically serves the `index.html` dashboard directly from the root port).

📊 Developer & Research Resources Panel

A slide-up resources panel is built directly into the bottom footer of the web dashboard. Developers, students, or clinicians can:

• —View and copy the Mermaid.js Flowchart code for the multi-agent decision pipeline.
• —Retrieve LaTeX Formulas for medical manuscript typesetting or documentation.
• —Read descriptive commentary describing the logical difference between traditional heuristic clinical rules and modern LLM-driven agent reasoning pipelines.