Ensemble of Feature:based and Deep learning:based Classifiers for Detection of Abnormal Heart Sounds

Abstract

The goal of the 2016 PhysioNet/CinC Challenge is the development of an algorithm to classify normal/abnormal heart sounds. A total of 124 time-frequency features were extracted from the phonocardiogram (PCG) and input to a variant of the AdaBoost classifier. A second classifier using convolutional neural network (CNN) was trained using PCGs cardiac cycles decomposed into four frequency bands. The final decision rule to classify normal/abnormal heart sounds was based on an ensemble of classifiers combining the outputs of AdaBoost and the CNN. The algorithm was trained on a training dataset (normal= 2575, abnormal= 665) and evaluated on a blind test dataset. Our classifier ensemble approach obtained the highest score of the competition with a sensitivity, specificity, and overall score of 0.9424, 0.7781, and 0.8602, respectively. IntroductionHeart auscultation is the primary tool for screening and diagnosis in primary health care [1]. Availability of digital stethoscopes and mobile devices provides clinicians an opportunity to record and analyze heart sounds (PCG) for diagnostic purposes. The goal of the 2016 PhysioNet/CinC Challenge is the development of algorithms to classify normal/abnormal heart sound recordings [2]. We proposed an ensemble of a feature-based classifier and a deep learningbased classifier to boost the classification performance of heart sounds. Method and MaterialA block diagram of the proposed approach to classify normal/abnormal PCG is shown in Fig. 1. Challenge DatabaseThe challenge database provided PCG recordings of healthy subjects and pathological patients collected at either a clinical or non-clinical environment. Details about the challenge dataset can be found in [2]. For algorithm development, in-house training and test sets were generated by randomly taking 80% and 20% of the records from each database, while keeping the same prevalence of abnormal classes. In-house training set was used for training and cross-validation of different models, and in-house test set was used for evaluation of the classification performance independently from the blind test dataset. Pre-processingEach PCG was resampled to 1000 Hz, band-pass filtered between 25 Hz and 400 Hz, and then pre-processed to remove any spikes in the PCG [3]. Furthermore, preprocessed PCGs were segmented into four heart sound states using a segmentation method proposed by Springer et al. [4]. Each PCG is comprised of more than one cardiac cycle (beat), and each beat is comprised of four heart sound states (i.e. S1, systole, S2, and diastole). Feature-based ApproachIn this approach, a variant of AdaBoost classifier [5] was trained for classification of normal/abnormal PCGs using time and frequency-domain features. Time-domain FeaturesMean and standard deviation (SD) of the following parameters were used as time-domain features (36 features): 1. PCG intervals: RR intervals, S1 intervals, S2 intervals, systolic intervals, diastolic intervals, ratio of systolic interval to RR interval of each heart beat, ratio of diastolic...

Short-segment Heart Sound Classification Using an Ensemble of Deep Convolutional Neural Networks

“…Applications of DL to cardiac signals are introduced very recently [6][7][8]. CNNs have been used for normal/abnormal PCG classification using input features such as spectrogram and Mel-frequency cepstrum coefficients (MFCCs) in [9] on 5-second windowed segments, and MFCC heatmaps of 3-second segments in [10].…”

Towards Domain Invariant Heart Sound Abnormality Detection Using Learnable Filterbanks

“…We use two different baseline models for performance comparison: (i) the top-scoring method from the INTERSPEECH 2018 ComParE Heart Beats Sub-Challenge by Gabor et al [16] as a traditional machine learning baseline, and (ii) the best performing system in the Physionet 2016 CinC Challenge developed by Potes et al [13] as a deep learning baseline.…”

“…The BoAW ensemble and the Fusion framework are denoted by Gabor-BoAW-SVC and Gabor-Fusion-SVC, respectively. The branched CNN model by Potes et al [13] (Potes-CNN) is implemented as our deep learning baseline system. It has a static front-end FIR filterbank as the input and provides inferences for each segmented cardiac cycles.…”

“…6 and 7 show the front-end FIR filter coefficients from different methods and their magnitude/phase responses, respectively. All of the learnable filterbanks are initialized with static FIR coefficients [13] except for the gammatone tConv units. The gammatone tConv is initialized with α k = 10 5 and η k = 4, while f k ∼ U(10, 400) and β k ∼ N (30, 6 2 ) are randomly sampled from a uniform and a normal distribution, respectively.…”

“…Notable features reported for this dataset includes, time, frequency and statistical features [8], Mel-frequency Cepstral Coefficients (MFCC) [9], and Continuous Wavelet Transform (CWT) [10]. Classifiers like SVM [11], k-Nearest Neighbor (k-NN) [9], Multilayer Perceptron (MLP) [10], [12] and Random Forest [8], deep learning approaches with 1D & 2D CNNs [13], [14], and Recurrent Neural Network (RNN) [15] based architectures were employed in the challenge submissions. The winning algorithm, similar to a good number of other submissions, proposed an ensemble; a static filter front-end 1D-CNN model combined with an Adaboost-Abstain classifier using a threshold-based voting algorithm.…”