Spaces:
Runtime error
A newer version of the Gradio SDK is available: 6.20.0
title: Heartline
emoji: 🫀
sdk: gradio
sdk_version: 5.33.0
app_file: app.py
pinned: true
short_description: Client-side cardiac home monitoring
tags:
- track:backyard
- sponsor:openbmb
- achievement:offgrid
- achievement:welltuned
- achievement:offbrand
- achievement:llama
- achievment:field-notes
Heartline 🫀
Cardiac home monitoring that runs entirely in your browser. No cloud. No data leaving your device. No GPU required.
Heartline lets patients log blood pressure readings, upload phonocardiogram audio from their phone, and generate a physician-ready clinical SOAP note — all processed locally using on-device AI.
Read the blogpost here, honestly its better than the video or even the readme for explanations https://dev.to/pranay_narula_bdc02868409/huggingface-build-small-hackathon-my-first-health-sprint-39g5
What it does
1. Blood pressure tracking
Log systolic, diastolic, and heart rate readings with timestamps and symptoms. A live trend chart and summary statistics (mean, SD, pulse pressure) update as you type.
2. ECG rhythm classification
Upload a or audio recording captured by pressing your phone microphone to the left sternal border (4th-5th intercostal space). A custom-trained 5-class classifier runs entirely in the browser via ONNX Runtime Web with WebGPU acceleration and WASM fallback. It classifies:
- Normal sinus rhythm (NSR)
- Tachycardia
- Bradycardia
- Atrial fibrillation (AFib)
- Premature ventricular contractions (PVC)
3. AI-generated SOAP note
A ** Finetuned Version of MiniCPM 1B Q4_K_M GGUF using constitutional finetuning with deepseek v4 pro based on soap clinical notes: https://huggingface.co/SlitherCode/heartline-soap
opensource on huggingface ** model runs in the browser via wllama (WebAssembly port of llama.cpp). It streams a physician-style Objective section incorporating all BP readings, heart rates, rhythm classifications, and patient history — tokens appear as they generate, no server involved.
Tech stack
| Component | Technology |
|---|---|
| ECG classifier | Custom PyTorch model exported to ONNX, runs via onnxruntime-web (WebGPU + WASM fallback) |
| Clinical LLM | heartline-soap which is a finetuned version of MiniCPM 1B Q4_K_M GGUF via wllama v3.5.0 (WebAssembly llama.cpp) |
| Frontend | Vanilla JS, zero framework, fully custom UI |
| Backend | Gradio Server used as static file server only — zero backend inference |
| Audio | AudioContext.decodeAudioData() → mel spectrogram → classifier |
Zero backend inference. The Gradio server only serves index.html, heartline.onnx, and model_Q4_K_M.gguf. All AI runs client-side.
Models used
- ECG classifier: Custom 5-class CNN trained on PhysioNet phonocardiogram data. ~94 MB ONNX.
- Clinical summarizer: MiniCPM 1B Q4_K_M quantized GGUF (~688 MB). Streams the SOAP Objective section directly in the browser via WebAssembly.
Medical disclaimer
Heartline is a research prototype and has not been evaluated or cleared by the FDA. It is not a substitute for professional medical advice, diagnosis, or treatment. Cardiac rhythm classifications are automated and may be inaccurate. If you are experiencing a medical emergency, call 911 or your local emergency number immediately.
Demo
Demo video: https://youtu.be/xRb6cS7FfWY
Social post: https://www.linkedin.com/feed/update/urn:li:activity:7472434733075869696/
How to use
- Fill in patient background (age, sex, conditions, medications)
- Log blood pressure readings with heart rate
- Upload a .wav audio recording from your phone (or use the provided synthetic samples)
- Wait for the ECG classifier to run (a few seconds)
- Generate the SOAP note — MiniCPM streams the clinical summary live in your browser
Built for the Build Small Hackathon — Backyard AI track