Healthcare Intelligence Platform — NRC x FPT Hackathon 2025

NRC x FPT Hackathon 2025 — First Prize Winner

"The future of healthcare analytics lies not in data collection, but in intelligent understanding — where every patient voice becomes a pathway to better care." — NRC x FPT Hackathon 2025

---

🏆 Overview

Project: Healthcare Intelligence Platform with Multi-Agent Knowledge Graph RAG

Achievement: 🥇 First Prize, NRC x FPT Hackathon 2025

Event Theme: AI-Powered Healthcare Analytics

Date: January 2025

Team: Led by Kato (Quan Ngo)

---

🎯 Problem Statement

Healthcare facilities receive thousands of patient feedback comments daily — rich, unstructured text containing critical insights about services, conditions, and experiences. Yet, extracting meaningful patterns from this data remains a monumental challenge.

Traditional analytics approaches fail to:

• Connect related concepts across different comments
• Understand sentiment context for specific healthcare entities
• Enable deep semantic queries that mirror how healthcare professionals think
• Resolve entity variations (e.g., "heart condition" vs "cardiac issue") into unified insights

The challenge for NRC x FPT Hackathon 2025:

Build an intelligent system that transforms raw healthcare feedback into a structured knowledge graph — enabling deep analysis, sentiment-aware entity tracking, and adaptive multi-agent reasoning.

---

🚀 Our Solution — Multi-Agent Knowledge Graph RAG

Our winning solution combines three revolutionary technologies into a unified intelligence platform:

1. Knowledge Graph Ingestion Pipeline — Transforms unstructured feedback into a semantically rich graph
2. Entity Extraction & Resolution — Identifies and unifies healthcare concepts across all comments
3. Multi-Agent Adaptive RAG System — Intelligently routes queries through specialized agents for optimal answers

Multi-Agent Application Architecture

---

🧠 Core Innovation: Knowledge Graph with Entity Resolution

Unique Entity Deduplication

Unlike traditional systems that treat each mention separately, our pipeline ensures one unique entity node per normalized concept across all comments. This means:

• "Heart condition" and "cardiac issue" resolve to the same entity
• Sentiment analysis aggregates at the entity level, not just per comment
• Healthcare professionals can query concepts, not just keywords

Intelligent Entity Extraction

Using AWS Bedrock Claude, we extract healthcare entities from free-text comments with remarkable accuracy:

• Medical Conditions — Diseases, symptoms, diagnoses
• Healthcare Services — Treatments, procedures, consultations
• Facility Attributes — Staff quality, wait times, cleanliness
• Patient Experiences — Emotional states, satisfaction indicators

Each extracted entity is:

• Normalized to a standard form (case-insensitive, standardized terminology)
• Embedded with semantic vectors for similarity search
• Linked to comments via sentiment-labeled relationships

Knowledge Graph RAG Architecture

---

💡 Sentiment-Aware Deep Analysis

Contextual Sentiment Tracking

Our system doesn't just classify comments as positive or negative — it analyzes sentiment at the entity level:

• A single comment mentioning multiple entities gets individual sentiment scores for each
• Sentiment labels include: positive, negative, and neutral
• Confidence scores accompany each sentiment assessment
• Original text spans are preserved for traceability

Rich Relationship Modeling

The knowledge graph connects:

• Patients → Provide feedback → Comments
• Comments → About → Facilities
• Comments → Relate to → Service Lines
• Comments → Mention → Entities (with sentiment properties)

This structure enables queries like:

• "What medical conditions are most frequently mentioned with negative sentiment?"
• "Which facilities receive positive feedback about cardiac care?"
• "How do patient experiences vary across different service lines?"

---

🤖 Multi-Agent Adaptive RAG System

Intelligent Query Routing

Our Coordinator Agent classifies incoming queries and routes them to specialized agents:

Semantic Search Agent

• Handles queries requiring similarity-based retrieval
• Leverages vector embeddings on both comments and entities
• Finds semantically related content even without exact keyword matches

Cypher Query Agent

• Executes graph traversal queries for relationship-based insights
• Navigates the knowledge graph structure efficiently
• Answers questions about connections and patterns

Reasoning Agent

• Performs complex analytical reasoning using Claude Sonnet
• Synthesizes information from multiple sources
• Generates insights that require multi-step logic

LangGraph Integration

The entire system is orchestrated using LangGraph, enabling:

• Dynamic agent selection based on query complexity
• Multi-step reasoning chains that combine different agent outputs
• Adaptive workflows that adjust based on intermediate results

---

📊 Data Flow Architecture

Ingestion Pipeline

1. Data Extraction — Pull structured healthcare data from SQL databases or CSV files
2. Entity Extraction — Use Bedrock Claude to identify healthcare concepts in free-text comments
3. Sentiment Analysis — Analyze sentiment for each extracted entity within its comment context
4. Embedding Generation — Create semantic vectors using Bedrock Titan embeddings
5. Graph Construction — Build Neo4j knowledge graph with unique entity resolution
6. Index Creation — Set up property and vector indexes for fast retrieval

Query & Analysis Flow

1. User Query → Coordinator Agent classifies the query type
2. Agent Selection → Routes to appropriate specialized agent(s)
3. Knowledge Retrieval → Agents query the graph using semantic search or Cypher
4. Reasoning & Synthesis → Combine results with multi-agent reasoning
5. Response Generation → Deliver insights in natural language or structured formats

Export & Reporting

• CSV Generation — Export query results for further analysis
• PDF Reports — Generate formatted reports for stakeholders
• API Integration — Connect with existing healthcare systems

---

🎯 Key Highlights

• 🏆 First Prize Winner — Recognized for innovation in healthcare AI analytics
• 🔗 Unique Entity Resolution — One entity per concept across all data sources
• 💭 Sentiment-Aware Analysis — Entity-level sentiment tracking with confidence scores
• 🧠 Multi-Agent Intelligence — Adaptive routing for optimal query handling
• 🔍 Semantic Search Ready — Vector embeddings enable natural language queries
• 📈 Scalable Architecture — Designed to handle millions of comments and entities

First Prize Award — NRC x FPT Hackathon 2025

---

© 2025 Kato (Quan Ngo) — Team Lead, NRC x FPT Hackathon 2025 First Prize Winner