Towards Stethoscope Invariant Heart Sound Abnormality Detection using Learnable Filterbanks

Published in IEEE JBHI, arXiv link

Objective: Cardiac auscultation is the most practiced non-invasive and cost-effective procedure for the early diagnosis of heart diseases. While machine learning based systems can aid in automatically screening patients, the robustness of these systems is affected by the phonocardiogram (PCG) acquisition device, i.e., the stethoscope. This paper studies the adverse effect of stethoscope/sensor variability on PCG signal classification and develops strategies to address this problem.

Methods: We propose a novel Convolutional Neural Network (CNN) layer, consisting of time-convolutional (tConv) units, that emulate Finite Impulse Response (FIR) filters. These filter coefficients can be updated via backpropagation and be stacked in the front-end of the network as a learnable filterbank. The filters can incorporate properties such as linear/zero phase-response and symmetry while ensuring robustness towards stethoscope variations by attenuating sensor dependent patterns.

Results: Our methods are evaluated using multi-source heart sound recordings obtained from the 2016 PhysioNet/CinC Challenge Dataset and the 2018 INTERSPEECH ComParE Heart Beats Sub-Challenge Dataset. The proposed learnable filterbank CNN architecture surpasses the top-scoring methods from both of these challenges on our multi-domain evaluation tasks. Our systems achieved relative improvements of up to 11.84% in terms of modified accuracy (Macc), compared to state-of-the-art methods.

Conclusion: The results demonstrate the effectiveness the proposed learnable filterbank CNN architecture in achieving robustness towards sensor variations in PCG signals.

Significance: To the best of our knowledge, this is the first research work that addresses the domain variability challenge for heart sound classification.

Requirements

• Python 3.6.3
• Matlab 2017b
• Keras 2.2.4
• Tensorflow 1.12.0
• Sklearn 0.19.1
• Tensorboard

How To Run

Data Preparation:

First download the data folder from this GoogleDrive Link
Place Physionet dataset (not included in the provided data folder) in the corresponding folders inside the data/physionet/training folder. The csv files containing the labels should be put inside the corresponding folders inside the labels folder and all of them should have the same name, currently 'REFERENCE_withSQI.csv'. If you change the name you'll have to rename the variable labelpath in extract_segments.m and extract_segments_noFIR.m
Run extract_segments_noFIR.m it first then run data_fold_noFIR.m to create data fold in mat format which will be loaded by the model for training and testing. fold0_noFIR.mat is given inside data/feature/folds for convenience, so that you don't have to download the whole physionet dataset and extract data for training and testing.

Training:

For Training run the trainer.py and provide a dataset name (or fold name) i.e. fold0_noFIR. The command should be like this :

python trainer.py fold0_noFIR

Other parameters like epochs, verbose, batch_size, pre-trained model path can be passed as arguments.

python trainer.py fold0_noFIR --epochs 300 --batch_size 1000 

Re-Generate Results:

Run the heartnet testbench.ipynb on Jupyter Notebook from the beginning until the block named Model.Predict . Select a log_name by uncommenting one from the LOG name block. The trained models for "heartnet type2 tconv" and "potes algorithm" is given in the log and model directory. These models are trained on fold0_noFIR which is included in the data folder.
To do the McNemer test read the instruction given in the LOG name block of the notebook. To plot roc curve run the ROC curve block.