Learning Front-end Filter-bank Parameters using Convolutional Neural Networks for Abnormal Heart Sound Detection

Abstract

Automatic heart sound abnormality detection can play a vital role in the early diagnosis of heart diseases, particularly in low-resource settings. The state-of-the-art algorithms for this task utilize a set of Finite Impulse Response (FIR) band-pass filters as a front-end followed by a Convolutional Neural Network (CNN) model. In this work, we propound a novel CNN architecture that integrates the front-end bandpass filters within the network using time-convolution (tConv) layers, which enables the FIR filter-bank parameters to become learnable. Different initialization strategies for the learnable filters, including random parameters and a set of predefined FIR filter-bank coefficients, are examined. Using the proposed tConv layers, we add constraints to the learnable FIR filters to ensure linear and zero phase responses. Experimental evaluations are performed on a balanced 4-fold cross-validation task prepared using the PhysioNet/CinC 2016 dataset. Results demonstrate that the proposed models yield superior performance compared to the state-of-the-art system, while the linear phase FIR filterbank method provides an absolute improvement of 9.54% over the baseline in terms of an overall accuracy metric.

Towards Domain Invariant Heart Sound Abnormality Detection Using Learnable Filterbanks

“…Cross-entropy loss is optimized using the Adam optimizer. The learning rate and hyperparameters are set according to [3] via 4-fold crossvalidation. Codes are available at Github 2 .…”

“…data acquired from different stethoscopes (especially of lower cost) and diverse environments. The frequency characteristics of the stethoscope or sensor used for recording can cause machine learning models to be biased towards majority sources of training data [3], [4]. The visible clusters in Fig 1-(a) corresponding to different stethoscope models prove that feature distributions can be significantly different depending on which domain the data is from.…”

“…In this work, we extrapolate on our findings of stethoscope dependent variances and quantify domain mismatch for PCG signals. We expand the scope of our learnable filterbank approach originally proposed in [3] incorporating novel variants of tConv layers including Type I-IV FIR and Gammatone auditory filterbanks. In addition, we propose a training regime within each mini-batch termed Domain Balanced Training (DBT), to specifically address domain differences.…”

FilterNet: A Many-to-Many Deep Learning Architecture for Time Series Classification

“…We choose the name FilterNet in order to emphasize a key and distinguishing property of this class of architectures-namely, that like the FIR filters common in signal processing [41], a FilterNet model can be applied to time series of arbitrary length, and it will infer an output time series of length proportional to the input length. This is true for FilterNet because it is true for all of its constituent building blocks -1D CNNs, LSTMs, pooling layers, interpolation layers, batch normalization layers, etc.…”

FilterNet: A Many-to-Many Deep Learning Architecture for Time Series Classification

“…We choose the name FilterNet in order to emphasize a key and distinguishing property of this class of architectures-namely, that like the FIR filters common in signal processing [36], a FilterNet model can be applied to time series of arbitrary length, and it will infer an output time series of length proportional to the input length. This is true for FilterNet because it is true for all of its constituent building blocks -1D CNNs, LSTMs, pooling layers, interpolation layers, batch normalization layers, etc.…”