PreDoc

Inspiration

The clinical "communication gap." Patients often struggle to articulate their symptoms accurately, and doctors have less than 10 minutes per appointment. While millions of people wear devices like the Luna Ring, the data remains "locked" in consumer apps rather than being translated into clinical insights. We wanted to build a bridge that grounds vague symptoms in hard data.

What it does

PreDoc (LunaSync) is a clinical translation layer that turns 14 days of noisy wearable data and messy patient logs into a structured, 30-second SOAP note for doctors.

• Agentic Triage: An AI agent analyzes the patient's "Hinglish" or slang symptoms to decide exactly which biomarkers (HRV, RHR, SpO2, Temp) are relevant to investigate.
• Targeted Analysis: It runs statistical checks—Z-score outliers, Pearson correlations, and medication-response tracking—to find the "signal" in the "noise."
• Dual Mode: It acts as a Clinical Briefing for doctors or a Readiness Report for sports coaches, focusing on recovery vs. overtraining.

How we built it

We built a two-pass agentic pipeline using:

• Python Backend: Leveraged the LLM-as-an-orchestrator pattern.
• Triage Pass: Uses Claude 3 / Llama 3 to intelligently select metrics and analyses based on symptom context.
• Synthesis Pass: Processes row-level data through specialized statistical modules (SciPy/Pandas) before generating the final clinical report.
• Frontend: A responsive Streamlit dashboard featuring a transparency panel to show the AI's reasoning.

Challenges we ran into

• Noisy Wearables: Filtering out "daytime noise" (stress/caffeine) from true clinical signals like overnight resting averages.
• Reliable JSON Parsing: Ensuring the triage agent always returns valid JSON for the analysis engine to consume.
• The "Hinglish" Problem: Training the model to accurately translate slang (e.g., "Mujhe thakan lag rahi hai") into clinical terms like "Generalized fatigue" without losing context.
• Clinical Objectivity: Fine-tuning prompts to ensure the AI acts as a data translation layer only—strictly avoiding unauthorized medical diagnoses.

Accomplishments that we're proud of

• Cross-Signal Correlation: Successfully automating the discovery of patterns, like a drop in HRV correlating with a rise in temperature deviation.
• Medication vs. Metric Tracking: Building a system that can objectively show if a patient’s medication (like Paracetamol) actually lowered their fever while their heart rate remained high.
• Dynamic Analysis: The dashboard isn't static—it actually changes what it looks for based on what the patient says, mirroring a real doctor’s diagnostic thought process.

What we learned

The value of AI in health isn't in "replacing" the doctor, but in pre-processing complexity. By converting 14 days of raw data into a 30-second briefing, we can save hours of clinical time and reduce physician burnout while improving patient outcomes.

What's next for PreDoc

• Multi-Wearable Integration: Expanding beyond the Luna Ring to include Oura, Apple Watch, and WHOOP data.
• EHR Integration: Directly pushing SOAP notes into hospital systems like Epic or Cerner.
• Real-time Voice Triage: A voice interface where patients can speak their symptoms, which are transcribed and triaged in real-time.

Built With

• claude
• streamlit