Medical Specialty Matchmaker: World Telehealth Initiative

The Arizona State University Artificial Intelligence Cloud Innovation Center, powered by Amazon Web Services (AWS), collaborated with World Telehealth Initiative to develop Medical Specialty Matchmaker - an AI-enabled clinical intake and specialty classification experience designed to help resource-constrained hospitals and clinics connect with the most appropriate volunteer specialists.

Built on a serverless AWS architecture, the solution supports both conversational triage and direct classification across 30+ medical specialties and 200+ subspecialties. The chatbot guides clinicians through targeted follow-up questions, classifies the case with confidence scoring, and stores structured request details in a centralized database for retrieval and follow-up.

The World Telehealth Initiative supports clinicians in low-resource settings by connecting them to volunteer medical specialists globally through telehealth. However, clinicians requesting support may be unfamiliar with how medical specialties and subspecialties are classified in the U.S. and other healthcare systems—resulting in requests being routed to the wrong specialists and delaying care.

World Telehealth Initiative needed a tool that could:

• Help clinicians describe the medical need in a structured way
• Identify the correct specialty and subspecialty quickly and consistently
• Ask focused, clinically relevant follow-up questions when information is incomplete
• Preserve clinician information across consultations to reduce repetitive entry
• Store request details for downstream matching and follow-up workflows

The CIC team built a serverless, Bedrock-powered Medical Specialty Matchmaker on AWS with three main flows: Clinician Intake, Triage + Classification, and Case Review + Submission. 

• AI (Bedrock): Dual-model strategy using Claude 3.5 Haiku for conversational intake + 2–3 focused follow-up questions, and Amazon Nova 2 Lite for structured extraction and specialty/subspecialty classification across 30+ specialties and 200+ subspecialties. 
• Confidence-based routing: The system only finalizes subspecialty when the classifier meets a 90% confidence threshold, otherwise it continues clarifying to reduce misrouting. 
• API layer & orchestration: Amazon API Gateway (REST) routes requests to AWS Lambda functions that manage conversation state, invoke Bedrock models, and assemble the response payload used by the UI. 
• Clinical reference grounding (no RAG ingestion): Specialty/subspecialty taxonomy is derived from the ABMS specialty list and used to guide classification behavior (treated as a controlled reference input rather than a crawled knowledge base). 
• Frontend: Next.js (React) web app hosted on AWS Amplify delivers a clinician-first workflow (saved clinician info → chat → review → submit) optimized for repeat requests. 
• Data storage: Amazon DynamoDB stores each structured request (specialty, subspecialty, reasoning, symptoms, age group, urgency, timestamps, and clinician context needed for follow-up) to support retrieval and downstream matching workflows. 
• Infrastructure: AWS CDK (TypeScript) deploys the serverless stack (Amplify, API Gateway, Lambda, DynamoDB, Bedrock integration) for repeatable environments and simple handoff. 

Medical Specialty Matchmaker improves telehealth request quality and routing accuracy by standardizing how clinicians describe needs and how those needs map to specialist categories.

"The ASU team delivered an elegant, scalable, and affordable solution with minimal input from us despite significant complexity in the functional requirements of translating international and diverse requests for medical support.  They created an impressive full stack solution, with an engaging and intuitive user experience that harnesses very naturally multiple powerful AI engines. The team was terrific to work with, and we enthusiastically rate their professional and technical skills as five stars."

 Dave Kosmo, Director of Technology, World Telehealth Initiative

This solution can be extended to additional global health and care coordination contexts, including:

• Referral intake workflows for regional hospital networks
• Telemedicine triage and routing support for NGOs and humanitarian settings
• Clinical decision support intake layers that standardize symptom capture
• Multi-language intake experiences for cross-border clinical coordination

Future enhancements may include multilingual support (e.g., Spanish, French, Ukrainian), improved token efficiency, expanded admin capabilities, and analytics dashboards for monitoring usage and model performance.