Discrimination of Normal and Abnormal Heart Sounds Using Probability Assessment

Abstract

Aims: According to the "2016 Physionet/CinC Challenge", we propose an automated method identifying normal or abnormal phonocardiogram recordings. Method: Invalid data segments are detected (saturation, blank and noise tests). The record is transformed into amplitude envelopes in five frequency bands. Systole duration and RR estimations are computed; 15-90 Hz amplitude envelope and systole/RR estimations are used for detection of the first and second heart sound (S1 and S2). Features from accumulated areas surrounding S1 and S2 as well as features from the whole recordings were extracted and used for training. During the training process, we collected probability and weight values of each feature in multiple ranges. For feature selection and optimization tasks, we developed C# application PROBAfind, able to generate the resultant Matlab code. Results: The method was trained with 3153 Physionet Challenge recordings (length 8-60 seconds; 6 databases). The results of the training set show the sensitivity, specificity and score of 0.93, 0.97 and 0.95, respectively. The method was evaluated on a hidden Challenge dataset with sensitivity and specificity of 0.77 and 0.91, respectively. These results led to an overall score of 0.84.

Recent advances in heart sound analysis

“…5.6. Plesinger et al (2017) Plesinger et al (2017 proposed an algorithm based on fuzzy logic which they termed 'probability assessment' for normal/abnormal heart sound classification, which achieved a M Acc of 0.8411 in the last phase of the challenge, and was ranked 7 th highest (Plesinger et al (2016)). The presented solution produced different results in specific databases.…”

“…This impressive result indicates the potential of CNNs for future use, but also illustrates how enormous volumes of data are likely to be required to outperform well chosen features and standard classification approaches. 2017) Plesinger et al (2017) proposed an algorithm based on fuzzy logic which they termed 'probability assessment' for normal/abnormal heart sound classification, which achieved a MAcc of 0.8411 in the last phase of the challenge, and was ranked 7th highest (Plesinger et al 2016). The presented solution produced different results in specific databases.…”

Classification of heart murmurs using an ensemble of residual CNNs

“…In 2016, most approaches mainly focused on feature-extracting methods that were classified with simple classifiers. For exam-ple, our team used a custom-made feature extraction tool and subsequent probability assessment [3,4], ranked 7th in the following phase. The winner of the 2016 challenge was already using convolutional neural networks [5].…”

Ensemble methods with outliers for phonocardiogram classification

“…Most studies of PCG classification to date have generated specifically selected summary signal features and applied them to a single machine learning algorithm, such as support vector machine (SVM) (Güraksın and Uguz 2011), artificial neural networks (ANN) (Randhawa and Singh 2015), etc. Among PhysioNet Challenge 2016's highest performing entries, commonly chosen features from all works included those derived from wavelet (Kay and Agarwal 2016), PCA (Bobillo 2016), mel frequency cepstral coefficients (MFCCs) (Bobillo 2016, Kay andAgarwal 2016), and fast Fourier transform (FFT)/Hilbert decompositions (Plesinger et al 2016). The challenge's entries also included several examples inspired by recent deep learning research that bypassed the explicit feature selection phase necessary in most previous works, with the use of convolutional networks that automatically determine filters that effectively decompose the signal with discriminating features (Potes et al 2016, Rubin et al 2016, Schölzel and Dominik 2016.…”

“…The machine-learning classification techniques of the highest performing entries included ANNs (Zabihi et al 2016), k-nearest neighbors (Bobillo 2016), generative models (Plesinger et al 2016) and boosting (Potes et al 2016). Other works inspired by deep learning used recurrent neural networks rather than conventional static networks to more naturally process timeseries data (Schölzel and Dominik 2016); and a DropConnect neural net, which incorporates a regularization technique to improve the generalization ability of large neural networks (Kay and Agarwal 2016).…”