Smartphone Phonocardiograph Murmur CNN Science Fair

ISEF Category: Biomedical Engineering

Subcategory: Biomedical Devices  ·  Difficulty: Advanced  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

A stethoscope plus a smartphone can replace a $500 cardiologist appointment for screening. A $5 piezo disc with a 3D-printed Helmholtz horn picks up heart sounds. A 1D-CNN trained on PhysioNet recordings classifies normal versus murmur in milliseconds. You can ship the whole thing as a free Flutter app.

What Is It?

A phonocardiograph captures the audio of the heart. A piezoelectric disc is essentially a thin ceramic that produces voltage from vibration.

A Helmholtz horn focuses certain frequency bands. A small 3D-printed cap shaped to match the disc and the phone's mic jack acts as both a holder and an acoustic amplifier.

The PhysioNet CinC Challenge released thousands of labeled heart-sound clips. A 1D-CNN learns time-domain features and outputs a murmur probability. Flutter lets you wrap the model in an app that runs on Android and iOS for free.

Why This Is a Good Topic

Mobile cardiology screening is a real research priority. Hardware is cheap and PhysioNet data is free. You will learn audio signal processing, CNN design, and mobile app deployment.

Research Questions

• How does horn geometry change recording signal-to-noise ratio?
• What is the effect of patient posture on murmur audibility?
• Does the 1D-CNN beat a spectrogram-CNN baseline?
• To what extent does smartphone microphone model affect accuracy?
• Which preprocessing pipeline best matches PhysioNet training?
• How does noise floor change between environments?
• What is the effect of recording length on classifier confidence?

Basic Materials

• Piezo disc (any maker supplier).
• 3D printer and PLA.
• TRRS adapter or USB-C audio dongle.
• Smartphone.
• PhysioNet CinC Challenge dataset.
• Informed-consent form for any live recordings.

Advanced Materials

• Calibrated reference stethoscope.
• Acoustic anechoic chamber.
• Lab spectrum analyzer.
• Cardiology mentor.

Software & Tools

• PyTorch: Trains the 1D-CNN.
• librosa: Preprocesses heart-sound audio.
• Flutter: Wraps the model into a cross-platform app.
• scikit-learn: Computes ROC and calibration.

Experiment Steps

1. Lock the horn geometry and recording chain.
2. Build a subject-wise split on the PhysioNet challenge data.
3. Decide preprocessing and target labels.
4. Train with cross-validation and report calibrated probabilities.
5. Deploy on-device and measure latency.
6. Compare on-device performance to laptop inference.

Common Pitfalls

• Recording at variable mic distances.
• Mixing subjects across train and test.
• Treating any high-frequency artifact as a murmur signature.
• Forgetting to convert mono-stereo channels consistently.
• Reporting only training accuracy.

What Makes This Competitive

Run subject-wise data splits, calibrate the recording chain against a known acoustic source, and report ROC curves with confidence intervals. A competitive entry also runs the model on-device with measured latency and battery impact.

Project Variations

• Add second- and third-heart-sound detection.
• Replace 1D-CNN with a transformer and compare.
• Run a comparison across multiple smartphone models.

Learn More

• PhysioNet CinC Challenge documentation: Free with starter code.
• PubMed: Search smartphone phonocardiography review.
• librosa documentation: Free audio-processing tutorials.
• Flutter documentation: Free cross-platform app guides.
• MIT OpenCourseWare: Course 6.555 Biomedical Signal and Image Processing.