Right Care, Right Timing – AI-Powered Subspecialty Triage for Women’s Health

An intelligent clinical triage assistant that automatically converses with patients, collects key medical information, determines urgency, matches the most appropriate OB/GYN subspecialty, and generates appointment summaries with transparent RAG-based references from the textbook Telephone Triage for Obstetrics and Gynecology.

Key Features

• Multi-turn patient dialogue
  • Collects patient name, DOB, contact info, menstrual and pregnancy details, symptoms, allergies, and insurance.
• Intelligent slot-filling
  • Dynamically decides which question to ask next based on missing information.
• RAG-powered clinical reasoning
  • Retrieves relevant pages from Telephone Triage for Obstetrics and Gynecology using FAISS + OpenAI Embeddings.
  • Displays referenced page numbers and excerpts alongside the triage report.
• Automatic urgency & specialty classification
  • Identifies whether the case is Emergency / Urgent / Routine and recommends the appropriate subspecialty.
• Doctor schedule integration
  • Reads an internal .xlsx schedule and assigns the earliest available physician automatically.
• Dual-mode compatibility
  • Works with both Streamlit web chat and Twilio phone call interfaces.

System Architecture

Streamlit (text chat UI)
TriageAgent (core logic)
├── Slot extraction & next-question logic
├── RAG retrieval from obgyn_index/
├── LLM summarization & confirmation
└── Reference display (page numbers + snippets)

Twilio (voice input/output)
└── OpenAI Realtime prompt for phone triage

Project Structure

women-health-triage/
├── app_chat.py               # Streamlit chat interface
├── main.py                   # Twilio voice interface (FastAPI + WebSocket)
├── appointment_db.py         # SQLite appointment booking utilities
├── schedule_loader.py        # Utility to read doctor schedule Excel
├── triage_agent.py           # Core multi-agent logic (slots + RAG + LLM)
├── requirements.txt          # Python dependencies
├── .env.example              # Environment variable template
├── fig1.pdf                  # Project architecture figure
└── README.md

Public Repository Note

This public snapshot intentionally excludes local secrets and generated/private data:

• .env with real API keys
• appointments.sqlite3
• uploaded_schedule.xlsx
• obgyn_index/ FAISS artifacts derived from the clinical reference text
• Python cache files

To run the full RAG workflow, recreate or provide your own local obgyn_index/ directory and point the app to a non-sensitive schedule file.

Quick Start

Install dependencies

pip install -r requirements.txt

Prepare your .env

Create a file named .env in the root folder:

OPENAI_API_KEY=your-openai-api-key

For a full template, copy .env.example and fill in your real values.

Launch the Streamlit App

streamlit run app_chat.py

Then open the displayed local URL (e.g., http://localhost:8501).

Configure the clinic schedule

The patient app reads the clinic schedule from an internal .xlsx file. By default:

uploaded_schedule.xlsx

To use a different file, set:

DOCTOR_SCHEDULE_PATH=/path/to/doctor_schedule.xlsx

Supported schedule formats:

Wide format: Doctor | Specialty | Insurance (optional) | Mon | Tue | Wed ...
Long format: Doctor | Specialty | Insurance (optional) | Day | Time

Day cells can contain Y, Yes, Available, or concrete times such as 09:00, 14:00.

If a day cell contains only Y, the app treats that doctor as available for the whole clinic day and expands it into one-hour slots. By default the clinic day is 09:00 to 17:00, so the generated slots are 09:00, 10:00, …, 16:00. You can override this with environment variables:

CLINIC_DAY_START=08:00
CLINIC_DAY_END=18:00

Appointment Database

The Streamlit app stores booked appointments in a local SQLite database:

appointments.sqlite3

Each (doctor, date, time) can be booked only once. Once a patient selects a slot, that slot is written to the database and disappears from future available slot lists. To place the database somewhere else, set:

APPOINTMENTS_DB=/path/to/appointments.sqlite3

RAG Integration

The RAG pipeline retrieves the most relevant passages from the OB/GYN telephone triage textbook to support model reasoning. These references are displayed at the end of the summary.

Example output:

Triage Summary
- Patient: Alice Chen
- Complaint: Light spotting at 32 weeks
- Urgency: Urgent
- Specialty: Maternal-Fetal Medicine

References from Handbook:
Page 132: Patients presenting with spotting after 30 weeks should be assessed for placental causes...
Page 135: If bleeding is associated with pain, immediate evaluation is recommended...

Optional: Twilio Voice Interface

To enable phone-based triage:

python main.py

• This spins up a FastAPI WebSocket server compatible with Twilio Media Streams.
• Patients can talk to a Realtime API voice triage assistant through a phone call.
• The system uses real-time speech detection and synthesis.

Recommended phone settings in .env:

VOICE_PORT=5050
PUBLIC_BASE_URL=https://your-ngrok-or-domain.example.com
CLINIC_NAME=Women's Health Clinic
OPENAI_REALTIME_MODEL=gpt-realtime
OPENAI_REALTIME_VOICE=verse
OPENAI_REALTIME_TEMPERATURE=0.7

PUBLIC_BASE_URL should be the public HTTPS domain that Twilio can reach. For local development, this is usually your ngrok URL. Do not include /incoming-call in PUBLIC_BASE_URL; the webhook path is added in Twilio.

You can check that the phone server is running at:

http://localhost:5050/health

Change the Twilio phone number

No code change is needed to switch the inbound phone number. In Twilio Console:

1. Buy or select the phone number you want to use.
2. Open that number’s Voice Configuration.
3. Set A call comes in to your public HTTPS URL ending in /incoming-call, for example:

https://your-ngrok-or-domain.example.com/incoming-call

1. Keep the method as POST or GET.

If your public tunnel/domain changes, update both places:

1. .env → PUBLIC_BASE_URL=https://new-domain.example.com
2. Twilio number Voice Configuration → https://new-domain.example.com/incoming-call

Then restart python main.py.

Technologies Used