CS-M — Cardiac Self-Monitoring Tool

Overview

CS-M started as a question my team kept coming back to: could you screen for heart disease with nothing but a microphone? Over two and a half years it grew from that question into custom recording hardware, a signal-processing pipeline, a recurrent neural network, and a React Native app — validated against 500+ patient recordings and recognized by Microsoft, Intel, and Regeneron ISEF along the way.

The Problem

Early detection of cardiac abnormalities usually requires clinical equipment and a trained ear. That puts routine screening out of reach for a lot of people, especially outside major hospitals. We wanted to know whether a low-cost device and a well-trained model could catch the same patterns a cardiologist listens for — cheaply enough to put in more hands.

What I Built

I led the signal-processing side of the project: building the dataset pipeline, preprocessing and denoising raw heart-sound audio, and preparing features for the classification model. In parallel, I helped design and test the recording hardware itself, iterating on device form factor and placement to get cleaner input signal before it ever reached the model.

The team split from there — hardware, AI, mobile, and research validation — and I worked across those boundaries to keep the audio pipeline compatible with whatever the model and app needed next.

Key Features

• Custom heart-sound recording hardware, purpose-built rather than adapted from a stethoscope.
• A denoising and feature-preparation pipeline tuned specifically for cardiac audio.
• An RNN-based classifier reaching 94.7% experimental accuracy on abnormality detection.
• A React Native app that made the whole pipeline usable outside a lab.
• A patent-pending signal-processing approach, validated on 500+ patient recordings.

Challenges

Heart sounds are quiet and easily buried in ambient noise, so most of the early iteration cycles were about the hardware and preprocessing, not the model — a bad recording can't be fixed downstream. Getting a filtering approach that held up across different devices, environments, and patients took a lot of testing before the model work could really start compounding.

What I Learned

This was my first time seeing a project all the way from a rough hardware prototype through to an international competition stage. It taught me that most of the hard problems in applied AI live outside the model — in data quality, hardware design, and the discipline of validating against real patient data instead of a clean benchmark set.