Inspiration

Emergency departments around the world are facing unprecedented strain. A growing nationwide nursing shortage has left hospitals with fewer staff to care for increasing numbers of patients, contributing to longer wait times, provider burnout, and overcrowded emergency rooms. We were particularly impacted by the story of Prashant Shreekumar, an Indian-Canadian man who died after waiting more than eight hours in an emergency department before receiving treatment. While no technology can replace healthcare professionals, we believed AI could help ensure that the sickest patients are identified sooner and that nurses spend less time on repetitive intake tasks that could be automated.

We created PriorityCare to address improve patient triage while streamlining nurse workflows so hospitals can deliver faster, more efficient care.

What it does

PriorityCare is an AI-powered emergency department triage assistant that helps determine the appropriate Emergency Severity Index (ESI) level and recommends the most appropriate level of care.

Patients can describe their symptoms conversationally, after which PriorityCare:

Extracts structured clinical information from free-text conversations.
Assigns an Emergency Severity Index (ESI) score using evidence-based guidelines.
Generates concise clinical summaries for nurses.
Flags high-risk symptoms that require immediate attention.

Beyond patient-facing triage, PriorityCare integrates directly into nurse workflows. Rather than simply producing an ESI score, it reduces documentation burden, speeds patient intake, and helps nurses prioritize patients more efficiently, reducing bottlenecks for both patients waiting to be seen and clinicians managing crowded emergency departments.

How we built it

The frontend began as a Figma Make prototype, which we exported into a React + Vite + shadcn/ui codebase to rapidly iterate on the user experience.

The backend is powered by FastAPI, which exposes three core endpoints:

Extraction: converts patient conversations into structured clinical information. Triage: converts structured data into an ESI score and care recommendation. Chat: combines both processes into a seamless conversational API for the frontend.

To maximize clinical accuracy, we built a hybrid AI pipeline instead of relying solely on an LLM.

Our triage engine combines:

Python for orchestration and clinical logic
The Emergency Severity Index (ESI) Handbook as a retrieval knowledge base
Sentence Transformers for semantic similarity search across triage guidance
Ollama running Llama 2 locally for cases requiring nuanced clinical reasoning instead of straightforward rule matching
Rule-based validation developed and refined with feedback from practicing triage nurses

This layered architecture allows PriorityCare to leverage deterministic clinical rules whenever possible while reserving language-model reasoning for more ambiguous presentations.

Challenges we ran into

Our biggest technical challenge was achieving clinically reliable triage accuracy.

Initially, an LLM-only approach performed poorly, achieving a macro F1 score of just 0.49. Adding retrieval-augmented generation (RAG) improved performance to 0.68, but this still wasn't sufficient for healthcare applications.

The breakthrough came from combining LLMs, retrieval from the ESI handbook, and rule-based validation designed alongside experienced triage nurses, increasing our macro F1 score to 0.98. This reinforced an important lesson: in safety-critical domains like healthcare, hybrid systems outperform purely generative approaches.

Another challenge emerged at the systems level. Improving triage speed alone simply shifted the bottleneck further downstream: patients were classified faster but still waited to receive care. To solve this, we expanded PriorityCare beyond patient triage and designed features that optimize nurse workflows as well, ensuring increased efficiency throughout the emergency department rather than at just one stage of care.

Accomplishments that we're proud of

1st Place at the Stanford Health Care Biomedical Engineering Health Innovation Symposium among 40 teams.
1st Place at the DEDA Entrepreneurship Competition, competing against 150 teams, and awarded $1,600 for Technical Excellence.
Developed a hybrid AI architecture that improved triage performance from a macro F1 score of 0.49 to 0.98.
Built a complete end-to-end clinical decision support platform integrating conversational AI, retrieval, clinical rules, and workflow optimization into a unified system.
Collaborated with real triage nurses to validate and refine our clinical decision-making pipeline. ## What we learned We learned that reliability comes from combining multiple approaches: clinical guidelines, retrieval systems, deterministic rules, and AI reasoning, rather than relying on any single technique.

We also gained a deeper appreciation for systems thinking. Improving one part of a hospital workflow can unintentionally create bottlenecks elsewhere, making it essential to design solutions that enhance the entire care process instead of optimizing only one step.

Finally, working closely with practicing nurses showed us that the best healthcare technology supports clinicians rather than replacing them. Their expertise was invaluable in shaping a system that is both technically accurate and practically useful.

What's next for Priority Care

Our next goal is to:

Integrate directly with Electronic Health Record (EHR) systems.
Expand multilingual support to improve accessibility for diverse patient populations.
Continue refining our triage models using additional real-world clinical data and nurse feedback.
Expand workflow automation features that further reduce administrative burden for emergency department staff.