• 

• 

• 

TriageAI: AI-Powered Medical Triage & Healthcare Navigation

Inspiration

The inspiration for TriageAI emerged from witnessing firsthand the critical inefficiencies in the U.S. healthcare system that disproportionately affect vulnerable populations. Our team members had experienced situations where friends and family struggled to navigate appropriate care options—resulting in either delayed treatment or unnecessary emergency room visits that created financial strain.

We were particularly moved by statistics showing that 13-27% of ER visits are for non-urgent conditions, costing the healthcare system billions annually while simultaneously creating access barriers for true emergencies. At the same time, 36% of Americans report skipping or postponing needed care due to cost concerns, and medical debt remains a burden for 41% of U.S. adults.

The convergence of these problems with recent advancements in AI—particularly the accessibility of large language models and computer vision technologies—inspired us to create a solution that could democratize access to medical guidance and healthcare resource navigation. We envisioned TriageAI as a bridge between anxious patients and appropriate care, leveraging technology for tangible societal benefit.

What It Does

TriageAI is an intelligent medical triage and healthcare navigation platform that guides users from symptom identification to the most appropriate care options. The application:

1. Accepts multi-modal symptom input through text description, voice messages, or image uploads of visible conditions (rashes, injuries, eye issues)
2. Engages in conversational assessment using a fine-tuned LLM that asks clarifying questions based on medical triage protocols
3. Provides risk stratification categorizing conditions into four urgency levels: Emergency Care, Urgent Visit, Primary Care, or Self-Care
4. Generates personalized care recommendations based on location, insurance status, financial constraints, and accessibility needs
5. Curates local healthcare options with detailed information about clinics, urgent care centers, and hospitals—prioritizing affordability (sliding-scale clinics), insurance acceptance, and proximity
6. Facilitates next steps with direct links to navigation, contact information, and telehealth options when available

The entire experience is designed to be empathetic, educational, and action-oriented, always emphasizing that the AI provides guidance rather than diagnosis and encouraging professional medical consultation.

How We Built It

Architecture Overview

We implemented a modular microservices architecture to ensure scalability, maintainability, and flexibility in integrating various AI components:

Frontend Layer:

• React Native with Expo for cross-platform mobile development
• React Navigation for seamless screen transitions
• React Native Maps for geolocation services
• Gifted Chat for conversational UI components
• Custom components built with Styled Components for consistent branding

Backend Services:

• Python Flask API gateway handling request routing
• Authentication service with JWT token management
• Session management for maintaining conversation context
• Geospatial query service for location-based recommendations

AI/ML Components:

• Fine-tuned Llama 3.1 model using parameter-efficient fine-tuning (PEFT) on medical dialogue datasets
• Google Cloud Vision API for image analysis (with fallback to TensorFlow Lite model for offline capability)
• Custom triage classification algorithm combining LLM output with structured symptom assessment
• Context management system to maintain conversation history and user preferences

Data Layer:

• PostgreSQL with PostGIS extension for geospatial queries of healthcare providers
• Curated database of healthcare resources incorporating:
  • HRSA-funded health centers
  • Medicaid/Medicare acceptance data
  • Sliding-scale clinic information
  • Real-time availability via Zocdoc API (where available)
• Redis for session caching and temporary data storage

DevOps & Deployment:

• Docker containerization for all services
• AWS ECS/Fargate for container orchestration
• CloudFront CDN for static asset delivery
• Automated testing pipeline with GitHub Actions
• Monitoring via CloudWatch and Datadog

Challenges We Ran Into

Medical AI Safety

The most significant challenge was ensuring the AI provided conservative, medically appropriate guidance without overstepping into diagnosis. We addressed this through:

• Implementing multiple safety layers in prompt engineering
• Incorporating disclaimers at every appropriate interaction point
• Building a fallback mechanism to default to higher acuity recommendations when uncertainty thresholds were exceeded
• Consulting with medical professionals during development to validate approach

Data Quality and Integration

Creating a comprehensive, accurate database of healthcare providers proved more challenging than anticipated:

• Public data sources were often incomplete or outdated
• Insurance acceptance information was particularly difficult to verify automatically
• We implemented a hybrid approach combining automated scraping with manual verification for our pilot region

Real-time Conversation Management

Maintaining context-aware conversations while managing API costs required careful engineering:

• Developed a session-based caching system to minimize redundant LLM calls
• Implemented conversation summarization to manage context window limits
• Balanced response quality with latency requirements through careful model selection

Privacy and Security Considerations

Handling potentially sensitive health information necessitated:

• Implementing end-to-end encryption for all health data
• Comprehensive anonymization of user interactions for model improvement
• Strict data retention policies with automatic deletion of sensitive information

Accomplishments That We're Proud Of

1. Developing a clinically validated triage algorithm that aligns with established medical protocols while maintaining approachability for end-users

2. Creating an intuitive user experience that reduces rather than exacerbates healthcare anxiety—verified through usability testing with diverse participant groups

3. Building a comprehensive healthcare provider database for our pilot region that includes affordability metrics and accessibility information often missing from commercial solutions

4. Achieving sub-second response times for most AI interactions despite the complexity of the models involved

5. Developing a robust architecture that successfully integrates multiple AI modalities (NLP and computer vision) with traditional healthcare data systems

6. Receiving positive feedback from healthcare professionals who tested the application and recognized its potential to improve appropriate care utilization

What We Learned

Technical Insights

• Fine-tuning strategies for domain-specific LLMs require careful balance between specialization and maintaining general conversational ability
• Multi-modal AI systems introduce unique challenges in synchronization and error handling
• Geospatial queries for healthcare resources need to consider multiple dimensions beyond simple distance (affordability, availability, appropriateness)

Domain Knowledge

• The U.S. healthcare system's complexity exceeds technical challenges—navigating insurance landscapes, regulatory requirements, and institutional practices requires deep domain expertise
• Healthcare accessibility encompasses more than physical proximity—financial, cultural, and language barriers significantly impact care utilization
• Effective triage requires understanding both clinical factors and human emotional states during health concerns

Project Development

• Building healthcare technology demands longer iteration cycles due to necessary safety considerations
• Partnering with domain experts early significantly improves product relevance and safety
• Pilot testing in limited geographies allows for more thorough validation before scaling

What's Next for TriageAI

Short-Term Priorities (0-6 months)

1. Expand geographical coverage to 5 additional states with particular focus on healthcare deserts
2. Implement insurance verification API integration to provide more personalized cost estimates
3. Develop provider portal allowing clinics to update their information directly
4. Add multilingual support starting with Spanish and Mandarin Chinese
5. Conduct clinical validation study to measure impact on appropriate care utilization

Medium-Term Goals (6-18 months)

1. Integrate with electronic health record systems using FHIR standards for better personalization
2. Develop partnerships with health systems to implement TriageAI as a formal triage tool
3. Create specialized assessment modules for mental health, pediatric care, and chronic condition management
4. Implement advanced features such as wait time prediction and appointment scheduling
5. Expand to additional countries with similar healthcare access challenges

Long-Term Vision (18+ months)

1. Develop predictive analytics capabilities to identify community-level healthcare access patterns
2. Create a public health dashboard for policymakers to identify service gaps
3. Build a research platform for studying healthcare navigation behaviors and outcomes
4. Establish a nonprofit entity to ensure the technology remains accessible to those who need it most
5. Contribute to open-source healthcare AI initiatives to advance the field collectively

Through this phased approach, we aim to transform TriageAI from a hackathon project into a sustainable initiative that genuinely improves healthcare access and reduces systemic inefficiencies—one conversation at a time.

Built With

• taskade