Abnormal Heart Sound Detection Considering Additive and Convolutional Distortion using Fused Cepstral Features

We propose to use a fusion of Mel-FrequencyCepstral Coefficients (MFCC) and its variants as the input to a 2D-Residual Neural Network architecture to address the presence of channel and additive noise distortion simultaneously.

Parametric Model of the Stethoscope

Multiple Cepstral Features

• fbank
• log-fbank
• mfcc_26
• mfcc_13
• fbank_log-fbank
• fbank_mfcc_13
• log-fbank_mfcc_13
• fbank_log-fbank_mfcc_13
• mfcc_13_d
• mfcc_13_dd

Proposed Model Architechture

Experimental Results

• Validation Scores of Multiple Trainings

• Consequence of Domain Balanced Training (DBT)

Requirements

• Python 3.8.5
• Matlab 2017b

Installation

git clone https://github.com/FarhatBuet14/CepsNET.git
cd CepsNET/codes
pip install -r requirements.txt

How To Run

* Data Preparation

1. 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
   

2. Run extract_segments.m it first then run data_fold.m to create data fold in mat format which will be loaded by the model for training and testing. fold_0.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.

3. For preparing features, run prepare_cepstralFeature.py with passing the name of the input feature (Select from the given list) as an argument. It will prepare an .npz file, saved in the data/feature/train_val_npzs folder. For example,

python prepare_cepstralFeature.py --inp mfcc_13

* Training

For Training run the train.py and provide a feature name i.e. mfcc_13.

python train.py --inp mfcc_13

Other parameters can be passed as arguments.

python train.py --ep 300 --batch 1000 

* Re-Generate Results

Download the logs folder from the Google Drive Link
Replace the folder with the repository folder. It contains all pretrained model files (.npz format).

There is given a validation.ipynb file to validate the pretrained models with different types of input features. pretrained.xlxs is given in the log folder which preserves the hypyer-parameter of all the pretrained models.

Change the feature variable in the Hyper-parameters block and obeserve the validation result.

* McNemar’s Test

We have also provided testbench.ipynb file which contains the codes for comparing the results with McNemar’s Test. Put any input feature in the Selected Feature variable and run the McNemar’s Test block to compare. We have provided a file (baseline_predictions.npz) in the logs folder with the predictions of the baseline models (potes and heartnet) to compare with our proposed model.