SympVis - Premium Symptom Triage with Explainable AI

SympVis replaces "Dr. Google" panic with AI-powered, explainable symptom triage—analyzing your age, history, and timeline to deliver a trusted risk verdict and a clinical PDF for your doctor.

Comment

Inspiration

We've all spiraled down the rabbit hole of searching symptoms online ("Dr. Google"), only to be met with worst-case scenarios that induce panic rather than clarity. Standard symptom checkers are often rigid, lacking the nuance of a real conversation. We wanted to build a bridge between the user and professional care: a "Triage Coach" that listens like a doctor, understands your specific biological context (age, sex, pregnancy), and—most importantly—explains its reasoning. Our goal was to reduce health anxiety by providing a transparent, evidence-based risk assessment.

What it does

SympVis is an intelligent, privacy-first triage application that acts as your first line of defense.

  1. Contextual Intake: It gathers critical biological markers (Age, Sex, Pregnancy status, Chronic conditions) to tailor its risk model.
  2. Visual Symptom Timeline: Users can map out exactly when symptoms appeared, helping the AI understand disease progression (e.g., "Fever started 2 days ago, Rash today").
  3. Interactive Follow-up: Instead of making a guess on vague info, SympVis asks 2-3 targeted clarification questions (generated dynamically) to rule out strict emergencies.
  4. Explainable Results: It delivers a coded Risk Level (Green/Yellow/Red) accompanied by a "Confidence Score" and a "Confidence Reason," explicitly citing which user-reported symptoms led to the conclusion.
  5. Actionable Output: Users get a generated PDF summary to show their doctor and, in high-risk cases, are immediately shown a directory of nearby emergency facilities.

How we built it

  • The Core (AI): We leveraged Google's Gemini 2.5 Flash for its speed and reasoning capabilities. We used a "Chain of Thought" prompting strategy to force the model to evaluate biological risk factors before generating a final verdict.
  • Frontend: Built with React 19 and Vite for a blazing fast experience. We used Tailwind CSS to create a "Glassmorphism" aesthetic that feels clean, premium, and calming—essential for a health app.
  • State Management: Complex multi-step wizard logic handles the flow from Basic Info -> Timeline -> Triage -> Follow-up -> Results.
  • Security: Integrated Clerk for seamless, secure user authentication.
  • Output: Used jspdf to programmatically generate a professional clinical summary that patients can physically take to a hospital.

Challenges we ran into

  1. The "Uncertainty" Problem: Early versions of the model would confidently guess even when the symptoms were vague. We had to implement a strict "Uncertainty Guardrail" where the AI is instructed to reject the request and ask for more detail if the confidence threshold isn't met.
  2. Structured JSON from AI: Getting the LLM to consistently return valid JSON for not just the risk level, but also the dynamic follow-up questions and array of "Observation Signals", was tricky. We refined our schema definition to ensure robust parsing.
  3. Visualizing Time: Designing a UI that allows users to easily impute a "Timeline" of symptoms (Day 1 vs Day 3) required several iterations to make it intuitive on mobile.

Accomplishments that we're proud of

  • The "Explainability" Layer: We didn't just want a black box. Seeing the AI quote specific symptoms back to the user ("I am concerned because you mentioned shortness of breath in combination with chest pain") builds immense trust.
  • Premium UX/UI: The application feels like a high-end medical device interface. The animations and transitions make the experience less clinical and more comforting.
  • Dynamic Safety: The system automatically adjusts its sensitivity. For example, a fever of 100.4°F triggers a different risk path for a pregnant user versus a general user.

What we learned

  • Context is King: A symptom check without biological context (age/sex) is nearly useless. Adding these parameters improved our AI's accuracy dramatically.
  • Prompt Engineering is a Safety Feature: You cannot rely on the model's default training for medical advice. You must explicitly instruct it on boundaries (e.g., "Do not diagnose, only assess risk").
  • User Trust: Users are more likely to accept an AI's advice if it admits when it's unsure, rather than forcing an answer.

What's next for SympVis Triage Coach

  • Multimodal Analysis: Integrating Gemini's vision capabilities to allow users to upload photos of visible symptoms (rashes, swelling) for analysis.
  • Voice Interface: Adding a "Speak your symptoms" feature for elderly users or those in distress.
  • Live Integration: Connecting the "Emergency Finder" directly to hospital wait-time APIs.
  • Local LLM Support: Exploring on-device models for privacy-focused, offline triage in remote areas.

Built With

Share this project:

Updates