INTELLIGENT PATIENT TRIAGE & CARE NAVIGATION ASSISTANT

Abstract

Project Overview This project presents an **Intelligent Patient Triage & Care Navigation Assistant**, a sophisticated multi-agent healthcare system designed to automate patient symptom assessment, urgency classification, and healthcare provider recommendation. Developed as part of an EMBA Seventh Semester Project, the system leverages modern artificial intelligence techniques and workflow orchestration to bridge the gap between patient symptom presentation and appropriate medical care access.

Technical Architecture The system implements a **LangGraph-style multi-agent workflow** comprising five specialized autonomous agents orchestrated through a directed graph with conditional branching logic:

1. **Symptom Intake Agent**: Employs natural language processing to extract structured symptom profiles from free-text patient input, including severity assessment (1-10 scale), duration parsing, and red flag detection for emergency symptoms.

2. **Triage & Reasoning Agent**: Utilizes Retrieval-Augmented Generation (RAG) over medical guideline documents to perform evidence-based urgency classification (Emergency/High/Medium/Low) and specialty mapping, generating comprehensive reasoning traces for transparency.

3. **Doctor Match Agent**: Implements a weighted scoring algorithm (specialty match 30%, availability 30%, patient rating 25%, proximity 15%) to rank and recommend the top three healthcare providers from a mock directory of 15 physicians.

4. **Scheduling & Summary Agent**: Interfaces with a mock REST API to facilitate appointment booking, generates patient-friendly visit summaries, and provides specialty-specific pre-visit instructions.

5

5. **Guardrail Agent**: Enforces critical safety validations including diagnostic claim prevention, mandatory medical disclaimer verification, emergency escalation validation, and prohibited medical language detection.

### Key Technical Features - **Conditional Workflow Branching**: Emergency cases bypass standard doctor matching and route directly to emergency room guidance, ensuring critical cases receive immediate escalation recommendations. - **RAG Implementation**: Medical guideline retrieval system enables evidence-based triage decisions grounded in standardized emergency criteria and symptom specialty mappings. - **Dual Interface Design**: - **Streamlit Web UI**: Interactive web application featuring real-time symptom analysis, color-coded urgency indicators, doctor recommendation cards, session history tracking, and JSON export capabilities. - **Command-Line Interface**: Python-based CLI for programmatic access and testing scenarios. - **Mock Healthcare API**: FastAPI-based REST endpoints (`/api/appointments`) with Swagger/ReDoc documentation, designed for seamless integration with real Electronic Health Record (EHR) systems. - **Comprehensive Logging**: Structured logging system capturing agent reasoning traces, state transitions, and validation results for debugging and audit trails.

### Technology Stack

6 - **Core Language**: Python 3.10+ - **Async Framework**: asyncio for concurrent agent execution - **Web Framework**: FastAPI for REST API, Streamlit for interactive UI - **HTTP Client**: httpx for asynchronous API communication - **Data Validation**: Pydantic for schema enforcement - **Logging**: Python logging module with file and console handlers

### Dataset Architecture

The system operates on three mock JSON datasets:

1. **doctors.json**: 15 physician records with specialties (Cardiology, Neurology, Orthopedics, etc.), ratings, availability schedules, geographic locations, and patient acceptance status.

2. **symptom_specialty_mapping.json**: Comprehensive mappings linking symptom patterns to appropriate medical specialties.

3. **medical_guidelines.json**: Emergency criteria database for RAG based urgency classification, containing symptom-context-urgency-action tuples.

### Safety & Compliance Framework

The system incorporates multiple safety layers:

7 - **Mandatory Medical Disclaimer**: All patient-facing outputs include explicit disclaimers stating the system does not provide medical diagnoses and recommending professional consultation. - **Diagnostic Language Prevention**: Regex-based pattern matching prevents the system from making definitive diagnostic claims. - **Emergency Escalation Validation**: Ensures critical cases receive proper emergency room guidance with 911 contact information. - **Output Validation Pipeline**: All final outputs pass through the Guardrail Agent before patient delivery. ### Workflow State Machine

The system implements an enum-based state machine with the following transitions:

``` INTAKE → TRIAGE → [EMERGENCY | DOCTOR_MATCH] → SCHEDULING → GUARDRAIL → COMPLETE ``` Emergency cases short-circuit the DOCTOR_MATCH state, proceeding directly to emergency guidance generation.

### Extensibility & Future Integration

The architecture is designed for production healthcare system integration: 8

- **Modular Agent Design**: Each agent operates independently with clear input/output contracts, enabling individual component replacement or enhancement. - **API Abstraction Layer**: Mock API endpoints mirror real healthcare API structures (FHIR-compatible), facilitating integration with Epic, Cerner, or other EHR systems. - **Configurable Ranking Weights**: Doctor matching algorithm weights are externalized for domain-specific tuning. - **Logging Infrastructure**: Structured logs support integration with monitoring systems (Prometheus, Grafana) and error tracking platforms (Sentry).

### Project Deliverables

1. **Core Application Files**: - `main.py`: CLI entry point with example patient scenarios - `streamlit_app.py`: Interactive web UI application - `workflow/graph.py`: State machine orchestrator

2. **Agent Modules**: Five specialized agent implementations in `agents/` directory

3. **Mock Healthcare API**: FastAPI server with appointment booking endpoints

9

4. **Mock Datasets**: Three JSON files simulating doctor directories and medical guidelines

5. **Documentation**: - `README.md`: Comprehensive setup and usage guide (501 lines) - `STREAMLIT_GUIDE.md`: Web UI user documentation - `requirements.txt`: Python dependency specifications

6. **Utility Infrastructure**: Logging configuration and module initialization files

### Educational Value & Applications

This project demonstrates advanced software engineering concepts: - **Multi-Agent System Design**: Orchestration of autonomous agents with specialized responsibilities - **Retrieval-Augmented Generation**: Practical implementation of RAG for domain-specific knowledge retrieval - **Conditional Workflow Logic**: State machine design with branching based on runtime conditions - **Healthcare Technology**: Understanding of medical triage processes and safety requirements - **Full-Stack Development**: Integration of backend logic, REST APIs, and interactive web interfaces

### Limitations & Disclaimers 10

**This system is designed exclusively for educational and demonstration purposes.** It is NOT intended for production medical use without: - Comprehensive clinical validation and testing - Regulatory approval (FDA clearance for medical devices, HIPAA compliance) - Integration with certified Electronic Health Record systems - Professional medical oversight and liability coverage - Extensive security auditing and penetration testing

All mock data (doctor information, medical guidelines) are simulated and should not be used for actual medical decision-making. The system does not handle Protected Health Information (PHI) and lacks the security infrastructure required for real patient data.

### Conclusion

The Intelligent Patient Triage & Care Navigation Assistant successfully demonstrates the application of multi-agent AI systems to healthcare workflow automation. By combining natural language processing, retrieval-augmented generation, weighted ranking algorithms, and safety validation pipelines, the project illustrates how modern AI techniques can be architected to support (but not replace) human medical decision-making. The modular design, comprehensive documentation, and extensible architecture provide a foundation for understanding healthcare AI system development while maintaining appropriate safety boundaries for educational use.