FunnelNet

Code for the paper "FunnelNet: An End-to-End Deep Learning Framework to Monitor Digital Heart Murmur in Real-Time".

Abstract

Heart murmurs are abnormal sounds caused by turbulent blood flow in the heart. Existing diagnostic methods, including cardiac auscultation, echocardiography, and phonocardiography (PCG), have limitations related to training requirements, accessibility, cost, and noise interference. This study proposes an end-to-end real-time heart murmur detection approach using traditional and depthwise separable convolutional networks. We applied a Butterworth filter and Continuous Wavelet Transform (CWT) to remove noise and extract meaningful features from PCG data. The proposed FunnelNet consists of three parts: a Squeeze net for compressed feature representation, a Bottleneck layer using depthwise-separable convolutions to reduce computational complexity, and an Expansion net to recover fine details. We evaluated the model on the publicly available CirCor pediatric heart sound dataset. With only ~5.4k parameters, FunnelNet achieved 85% accuracy, 85% sensitivity, and 92% specificity, outperforming several larger models. We further converted the network to TensorFlow Lite for TinyML deployment on resource-constrained devices. The model achieved an average real-time inference accuracy of 91% on a Raspberry Pi 4B and 80% on an Android smartphone.

Overview

Fig: An illustration of the proposed FunnelNet network architecture, divided into three parts: the squeeze net, the bottleneck, and the expansion net. The Squeeze and Expansion networks work like an encoder and decoder to compress and upscale the most relevant input features. A depthwise CNN has been employed at the Bottleneck for reduced computational complexity. A fully connected layer is placed at the end to predict the correct class.

Folder structure

Path	Content
src/	Model, data loading, signal processing and training utilities
data_create.py	Builds the processed dataset (data/<dataset>.npz) from the audio files
tuner.py	Hyperparameter search
train.py	Cross-validated training, evaluation and TFLite conversion
config.yml	All settings (dataset, learning rate, epochs, folds, ...)
run_train.sh	Runs train.py inside the conda environment
eda.ipynb	Notebook used to make the figures
rpi_inference.py	Inference script for the Raspberry Pi
android-src/	Android application
fastapi-server/	Server-side API

Data

The audio comes from the public CirCor DigiScope dataset on PhysioNet. It is not included in this repository. Download it and place the wav files and training_data.csv in dataset/circor/.

How to run

1. Create the environment:

   conda env create -f environment.yml
   conda activate fn

2. Build the processed data:

   python data_create.py

3. Train:

   python train.py

   Settings such as the learning rate, epochs, folds and the SMOTE / outlier-removal switches are in config.yml. Any setting can also be overridden on the command line, for example python train.py --run.epoch 250 --run.smote true.

4. Optional: python tuner.py runs the hyperparameter search.

The trained model is saved under logs/ and, with tflite: true, converted to a TensorFlow Lite model for the Raspberry Pi and Android app.

License

See LICENSE.