Inspiration
Across the globe, clinicians spend nearly two hours on electronic health record documentation for every hour spent with patients. This administrative bottleneck fuels physician burnout, prolongs waiting room times, and dilutes patient care. Language barriers and low health literacy compound the issue, leaving patients struggling to articulate complex symptoms under pressure. We set out to rethink clinical intake: what if patient history could be gathered conversationally in their native tongue and converted into standardized clinical notes before the patient even walks into the consultation room?
What It Does
Our platform serves as an autonomous, multilingual pre-consultation agent:
- Conversational Intake: Engages patients via an intuitive chat interface to elicit chief complaints, onset, severity, medical history, and timeline.
- Multilingual Translation: Allows patients to speak or type in their everyday language or colloquial terms, eliminating miscommunication.
- Structured Clinical Synthesis: Parses unstructured dialogues into standardized clinical formats (such as HPI, ROS, and SOAP notes).
- Interoperable EHR Export: Formats gathered data into FHIR-compliant JSON structures, ready for instant integration into existing hospital management systems.
How We Built It
- Frontend: Built with React/Next.js and Tailwind CSS for a fast, accessible, mobile-first patient and doctor dashboard.
- LLM Engine & Orchestration: Powered by modern large language model APIs structured via LangChain/LlamaIndex to dynamically generate adaptive clinical follow-up questions.
- Schema Validation: Enforced strict JSON schema outputs using Pydantic to ensure all generated notes match standard clinical taxonomy (SNOMED-CT/ICD-10 alignments).
- Backend: FastAPI/Node.js handling data encryption, session management, and API orchestration.
Challenges We Faced
- Hallucination Control & Safety: Ensuring the AI strictly acts as an information gatherer and structuring agent rather than attempting to diagnose or prescribe treatments. We solved this with rigorous system-prompt boundaries and guardrails.
- Medical Entity Extraction: Converting loose, colloquial descriptions of pain and timeline into precise medical terminology without losing subtle context or alarming the patient.
- Latency & Token Efficiency: Optimizing multi-turn context windows to maintain real-time chat responsiveness on low-bandwidth mobile devices.
Accomplishments We're Proud Of
- Building a seamless pipeline that reduces intake documentation time by an estimated 60-70%.
- Achieving high fidelity in mapping non-English patient phrasing into standardized clinical English summaries.
- Designing a clinician-first UI that lets a doctor review an entire intake history in under 30 seconds.
What We Learned
We discovered how critical FHIR/HL7 interoperability standards are for real-world health-tech adoption. We also deepened our understanding of clinical communication—specifically how doctors parse the History of Present Illness (HPI) versus how patients intuitively describe their discomfort.
What's Next
- Direct integration with major open EHR platforms (OpenEMR, Epic on FHIR).
- Ambient voice intake for elderly patients who prefer speaking over typing.
- Offline-first edge deployments for rural clinics with intermittent internet connectivity.
Log in or sign up for Devpost to join the conversation.