• No results found


ilarity in real and synthetically generated beats. The training curves of DCCGAN depicted stable training, and the change in evaluation metrics followed a similar pat- tern. Generated synthetic heartbeats resemble real beats as they encompass essential characteristics present in beats and follow the intricate structure present in the dif- ferent classes of beats.



Penalty Induced Prototype-Based eXplain- able ResNet for Heartbeat Classification

An abnormality in heart rhythm or deviation of electrical impulses from normal sequence leads to irregular heartbeats or cardiac arrhythmias. The heartbeats con- sidered for this work are non life threatening heartbeats such as normal beat (N), supraventricular ectopic beat (SVEB), and ventricular ectopic beat (VEB) as these are the major categories of beats recommended by the Association for Advancement of Medical Instrumentation (AAMI) [7]. This chapter focuses on classifier design and explanation generation aspect of CAD as it highly influences the classifier perfor- mance [17]. To achieve state-of-the-art performance, the developed models require huge amount of labeled, diverse, and realistic-looking heartbeats. Therefore, heart- beat synthesis of regular and irregular heartbeats is performed as described in Section 3.3. The beats are classified by deep learning models [16]. Although deep learning models achieve good performance for heartbeat classification, their black box nature inhibits real world deployment [236]. A Penalty Induced Prototype based eXplain- able Residual Neural Network (PIPxResNet) is proposed for heartbeat classification that provides explanations along with improved diagnostic performance. PIPxRes- Net is adopted from explainable deep neural network [31] due to its non-iterative and non-parametric nature that eliminates hyperparameter tuning. PIPxResNet employs pretrained residual neural network that extracts features from decluttered heartbeats and minimizes resource consumption (eliminating the need for GPU, reduced train- ing and inference time) following the concept of task transfer learning. The features similar to other class features are penalised and their contribution towards the corre- sponding class is reduced. The selected encoded features represent prototypes. The TH-2764_156201001


prototypes are combined to make rules that explain model predictions and enhance trust in predictions. In addition, the encoded and actual prototypes are utilised to generate a reliability score that depends on nearest prototype support and correlation between the actual prototype and heartbeat. PIPxResNet is verified on four publicly available standard databases.

The chapter organization is as follows. Section 4.1 provides the literature review of the previous heartbeat classification methods. Section 4.2 describes the proposed PIPxResNet technique. Section 4.3 describes experimental setup and results. Section 4.4 summarizes the chapter.

4.1 Literature Review

Heartbeat classification has been performed using a combination of feature extrac- tion, feature selection, and classification methods. Several time domain, frequency domain, and morphological features have been proposed in the literature. The most commonly used features include RR interval (RRI), defined as the interval between two adjacent R-peaks, that achieves the best performance [229]. Other features in- clude, Morphological features such as position, amplitude, and interval features [237], wavelet based features [17, 238–244], discrete cosine transform [245], Hilbert trans- form [246], Kalman filter-based features [247], stockwell transform (ST) [248], higher order statistical (HOS) features [249], Hermite polynomial based features (HBF) [250], mixture of features [251] and complex heartbeat representations such as mean connec- tivity degree, maximum connectivity degree, joint entropy, joint energy, mean joint degree extracted from temporal vectorcardiogram (TVCG) [252]. Redundant features are removed through dimensionality reduction techniques such as principal compo- nent analysis (PCA), linear discriminant analysis (LDA), independent component analysis (ICA) [243]. Feature selection has been performed through particle swarm optimization (PSO) [252, 253], genetic algorithm (GA) [254], and bacteria foraging optimisation (BFO) [248] have been employed to reduce feature redundancy.

The classification of extracted features has been performed using weighted con- ditional random fields (CRF) [237], support vector machines (SVM) [93, 96, 245, 248, 252, 255], mixture-of-experts [256, 257], and K-nearest neighbors [240]. Clustering methods [250, 258] such as maximum margin clustering with immune evolution [259], fuzzy-entropy-based clustering [260], K-means clustering [260] have also been em- ployed for classification. The handcrafted features sometimes fail to capture essential information present in data and limit the learning capability of machine learning

82 TH-2764_156201001


models and reduce the performance.

Therefore, end to end Deep Neural Networks have been applied that have achieved exceptional performance for heartbeat classification. These include recurrent neural networks (RNN) [239, 261] based long short term memory (LSTM) [239], fuzzy neu- ral networks [241], convolution neural networks (CNN) [262–265], multiscale wavelet CNN [244], feedforward neural network (FFNN) [245], probabilistic neural network (PNN) [245], multilayer perceptron (MLP) [251], modular neural networks [256], hybrid neural network [266], deep belief networks [267], artificial neural network (ANN) [93], learning vector quantization NN [268], and radial basis function net- work [242, 269].

The developed classifiers lack explainability, leading to reduced confidence in model diagnosis. The intrinsic interpretation methods create interpretable models and explain the prediction using data points from training dataset. The selected data points or prototypes represent data in a condensed format that encompasses basic characteristics of the respective class [31]. Several such methods have been developed in the past including prototype with autoencoders [265], deep sequence model con- sisting of sequence encoder and prototype layer for classification and reasoning [270], and explainable deep neural network (xDNN) [31].

The methods developed in the past have achieved excellent performance but the limitations are: (i) Only a few methods follow the ANSI/AAMI EC57:2008 [7]

recommended beat classes; (ii) Extraction and selection of handcrafted features; (iii) existing models perform inconsistently when tested on other datasets as their testing was limited to a single or couple of databases; (iv) the black box nature of neural networks limits real-world deployment. The aforementioned issues are addressed by performing heartbeat classification and prediction explanation through PIPxResNet.

4.2 Penalty Induced Prototype-Based eXplainable