Deep Learning Models for Cardiac Abnormality Detection from ECG Signals: An Interpretability Perspective

165 7.10 Proposed 8-layer ResNet architecture with different con- RNNs. figurations and the module of the attention mechanism. Several interpretability approaches have been proposed in the literature that aim to provide explanations or interpretations of the model to stakeholders.

Figure 1.1: Number of Cardiovascular Deaths from 1990 to 2019 [1].

Motivation

The training data is crucial in developing a generalized classifier that performs well when tested on new subjects with CVD. Thus, this work aims at developing a good classifier that achieves good performance and model interpretation.

Challenges

Large segments may not focus on a specific part of the segment responsible for classifying the rhythm as arrhythmic. In addition, the explanation of the classification is also required to support the model decisions.

Contributions

In this work, BW is removed using VMD, and results are shown for normal sinus rhythm and ventricular tachycardia. The performance of VMD is evaluated using the percentage root mean square difference, Pearson correlation, maximum absolute error and signal decomposition time.

Thesis Outline

Electrophysiology behind Heartbeat Genesis

First, the left and right atrium are depolarized due to the firing of the SA node. Then, the firing of the AV node transmits the heart's electrical activity from the atria to the ventricles.

A Brief History of Clinical Electrocardiogram

Thus, the Thoracic Leads/Precordial Leads/Candid Leads representing the horizontal plane of the heart were born. The invention of unipolar leads completed the major advance toward the 12-lead ECG.

Figure 2.2: Conduction System of Heart [2].

Multichannel Electrocardiogram and its Acquisition

In 1942, Emanuel Goldberger used Wilson's center terminal, constructed unipolar plugs with the center (zero) terminal, and attached additional positive unipolar plugs to each of the left and right arms and left leg [67] to provide more detailed coverage of the frontal plane. Lead V1 and lead V2 are positioned in the fourth intercostal space just on the right sternum and left sternum, respectively [69].

Figure 2.5: Standard 12-Lead ECG Placement on Human Body [4].

Overview of a Typical Electrocardiogram Classification Framework . 22

Heartbeat Segmentation from Single Lead ECG
Synthesising ECG Signals
Classification using Deep Learning Models
Interpreting Model Diagnosis

Low- and high-frequency noises distort the ECG signal during acquisition and make the subsequent processing challenging [74]. QRS detection is difficult due to the physiological variability, and various noises present in the ECG signal [85].

Figure 2.7: Computer Aided Disease Diagnostic System.

Interpretability Landscape

Defining Interpretability and Explainability
Interpretability Criteria
Taxonomy of Interpretability Methods
Implicit vs Posthoc Techniques
Model Interpretability in Cardiology

Nature of interpretability techniques: Techniques can be applied to a model after or during model construction. Explainable Prototypes: The approaches are applied in a cascading fashion to a black-box model that provides prototypes as explanations for improving the transparency of the model's internal mechanisms [118]. They can be applied to a pre-trained interpretable model to analyze the input-output relationship to describe the internal mechanism of the model.

Figure 2.9: Taxonomy of interpretability techniques.

Summary

Brief Overview of Baseline Wander Removing Techniques
Investigated Techniques
Dataset Description
Methodology
Results and Discussion

BW can be represented as a sinusoidal component at the respiratory rate added to the ECG signal, see Figure 3.1f. Midrange filters produce a very smooth baseline due to the presence of the midrange filter. From Figure 3.6 it can be observed that due to the change in the baseline, the performance metrics worsened.

Figure 3.1: Types of noises corrupting ECG signals.

R-Peak Detection from Single Lead ECG

Brief Overview of Beat Detection Techniques

However, the technique was not sensitive to the change of QRS integral information, leading to errors in case of low amplitude QRS, sudden change of amplitude and high P waves and T waves. 203] used EMD to decompose the ECG signal into high- and low-frequency components, followed by nonlinear transformation to enhance the QRS complex. 204] utilized the dual criteria of amplitude and QRS complex duration using finite impulse response filter, differentiation and threshold.

Proposed Fractal Based R-Peak Detection

The procedure consists of a pre-processing step, fractal analysis and a post-processing step as described in Figure 3.10. Any set that is a subset of R satisfies the property partial ordering by the relation ≤ for any pair of elements xi, xj ∀i, j,∈ N. The second property is satisfied in x[n] due to the existence of peaks and valleys in the signal. The steps to calculate area are described in Algorithm 5. The recursive nature of Algorithm 5 was further optimized using the approach of dynamic programming with memoization, which reduced the time complexity of the algorithm and made it computationally efficient.

Figure 3.10: Different stages of the proposed approach.

Results and Discussion

Maximum sensitivity was obtained using area average as threshold and frame size of 10 timestamps. Increasing threshold with less frame size also enabled the algorithm to detect significant spikes, minimizing DER. The figures show only those combinations of threshold and frame size where the algorithm was able to detect a significant number of peaks in the MIT-BIH records.

Figure 3.12: Radar Plot of Sensitivity and Predictivity with varying threshold and frame size using Triangular Structuring Element.

Heartbeat Synthesis using Generative Models

Brief Overview of ECG Synthesis Techniques
Proposed Deep Convolution Conditional GAN
Dataset Description and Preprocessing
Evaluation Metrics
Results and Discussion

DCCGAN performance is measured by the quality and diversity of the generated signals obtained from the generator model. The beats generated by the generator model at these groups depict a perfect match with the actual beats present in the training data set. The high error in evaluation metrics of normal beats may be due to irregularity or slight problem in the generation of R wave of QRS complex, that is, the ventricular depolarization of the heart.

Figure 3.17: Illustration of regular and irregular beats.

Summary

A Penalty Induced Prototype-based Explainable Residual Neural Network (PIPxResNet) is proposed for heartbeat classification, which provides explanations along with improved diagnostic performance. Heartbeat classification was performed using a combination of feature extraction, feature selection, and classification methods. Therefore, end-to-end deep neural networks have been applied that have achieved outstanding performance for heartbeat classification.

Penalty Induced Prototype-Based eXplainable Residual Neural Network 83

Data Description and Preprocessing
Evaluation Metrics
Feature extraction using Neural Networks
Pretrained Neural Network Performance
Evaluation of Prototype-Based Techniques
Prototype Interpretation
Comparison with Existing Methods

The training and testing times were reduced due to the reduction of the data dimension from 186 to 3. The introduction of penalties helps to obtain better prototypes that are more representative of the training dataset rather than affecting the number of prototypes. It is characterized by an abnormal P-wave (missing or negative P-wave) and a very short duration of the QRS complex, especially the missing P-wave and a narrow QRS-complex describe SVEB compared to N. SVEB prototypes also correctly represent the ideal SVEB and thus provide a good explanation.

Figure 4.2: Heartbeat classification and explanation.

Summary

Deep learning models have performed better, but the black-box aspect limits real-world implementation. In the past, deep learning models (DLMs) were used to detect VT and VF from a single-channel electrocardiogram (ECG) signal. Studies performed to detect ventricular tachyarrhythmias can be classified into two broad categories: conventional feature-based techniques and deep learning-based techniques.

Methodology

Data Description and Preprocessing
Residual Neural Network Classifier
Posthoc Interpretability Techniques
Sanity Check Mechanism

The score and gradient are used to calculate GBP, Grad CAM and Guided Grad CAM client maps. 45] is used to verify the validity of saliency maps by performing model parameter randomization and data permutation test. Model weight randomization test: The weights of convolution layer filters are modified and the effect on saliency maps is observed.

Figure 5.1: Block diagram of the proposed framework.

Experiments and Results

Quantitative Evaluation
Comparison with State of the Art Techniques
Qualitative Evaluation
Ventricular Arrhythmia Origin
Correctly Diagnosed Rhythms
Incorrectly Diagnosed Rhythms
Interpretability Technique Comparison

The origins of NSR, VT and VF with their respective ECG waveforms are shown in Figure 5.11. The saliency maps obtained from the last convolution layer of the ResNet model for misclassified rhythms are illustrated in Figure 5.15. In Figure 5.15c, coarse VF is predicted as NSR, but GC focuses on narrow QRS complexes present in NSR.

Figure 5.6: Model performance variation for increasing number of ResNet layers for two second ECG segment.

Summary

325] published a review of deep learning-based AF detection that coherently describes recent advances in neural network models for AF detection. 323] used a combination of expert features including statistical features, medical features, center-wave frequency features extracted using a dynamic time warping technique, and features extracted using a deep neural network for AF detection. 145] used a bidirectional recurrent neural network followed by a three-level attention mechanism, a wave level to calculate wave weights, a beat level to calculate beat weights, and a window level (multiple heartbeats) to calculate window weights responsible for AF detection.

Figure 5.15: Saliency maps of incorrectly classified NSR, VT, and VF rhythms.

Methodology

Dataset Description and Preprocessing
Baseline Convolution Neural Network
Residual Neural Network
Attentive Convolution Neural Network
Transformer Neural Network
Interpretation

The segment length plays a major role in evaluating the performance of the model and it has not been investigated much in the literature as per the TH. The input ECG is supplied to the encoder block, and the output of the last convolution filter followed by BN and ReLU activation is added with positional encoding (PE) output and supplied to the transformer block. The output of the MHA module is supplied to the dropout layer, and the output of the dropout layer is added with MHA output through an additive residual connection and passed through a normalization layer.

Figure 6.2: Illustration of different rhythms [6].

Experiments and Results

Quantitative Performance Evaluation

The performance of all the models using different segment length (ranging from 1 to 10 seconds duration) ECG signals is shown in Figure 6.6 using Precision (Pr), Accuracy (Acc), Sensitivity (Se), Specificity (Sp) and F1 Score (F1). It can be observed that ResNet achieved the highest performance among all the models for all length ECG signals followed by TNN, BCNN and ACNN as can be seen in Figures 6.6b, 6.6d, 6.6a and 6.6c respectively. The results show a clear improvement in performance when the signal length was increased from 1 to 3 seconds and the performance deteriorated for longer length ECG signals.

Table 6.2: TNN parameter selection for three second ECG signal.

Comparison with Existing Methods

Qualitative Prediction Interpretation

The TNN model focuses heavily on the presence of P-waves and the ST segment, as shown in Figure 6.9d. TNN focuses more on the f-waves and less on the QRS complex, and only the single-beat QRS complex is highlighted, as shown in Figure 6.10b, making TNN the better choice. Models focus on f-waves without focusing on QRS complexes, while classifying segments as AF.

Figure 6.7: Confusion Matrix of all models for 2 second ECG signal.

Summary

In the first stage, the classification of the ECG with a 12-lead label is performed through the examination of convolutional, repetitive and attention-based patterns. Several studies aimed at solving the 12-lead ECG classification problem have been published. Traditional machine learning (ML) models such as K-Nearest Neighbor [339] and Support Vector Machines [340] combined with hand-crafted features have achieved significant performance for 12-lead ECG classification .

Figure 7.1: Overall Stage Wise Workflow.

Experimental Setup

Problem formulation

The loss represents the negative mean log of corrected predicted probabilities, whereσ(xi) is the probability that the sample xi belongs to the respective cardiac pathology.

Dataset Description

Left bundle branch block LBBB 1st degree AV block IAVB Sinus rhythm NSR Complete left bundle branch block CLBBB Prolonged PR interval LPR Bradycardia Brady Complete right bundle branch block CRBBB Prolonged QT interval LQT Atrial flutter AFL. Ventricular premature contractions PVC Q wave abnormal QAb T wave abnormal TAb Incomplete right bundle branch block IRBBB Right axis deviation T wave inversion RAD TInv Supraventricular premature beats SVPB Right bundle branch block RBBB Sinus arrhythmia SA. Left anterior fascicular block LANFB Bundle branch block BBB Left axis deviation LAD Nonspecific intraventricular conduction NSIVCB Sinus bradycardia SB Sinus tachycardia STach.

Table 7.3: Diagnosis with Abbreviations (Abbrev).

Evaluation Metrics

The quantity |xk∪yk|is the number of distinct classes with positive label and/or classifier output for record k. CRBBB and RBBB are considered similar, so a predicted output in one of these classes is considered a positive label or classifier output for all. The following class pairs: PAC and SVPB; PVC and VPB; CRBBB and RBBB are considered similar, so a predicted output in one of these classes is considered a positive label or classifier output for all.

Figure 7.4: Illustration of the Reward Matrix for the scored diagnoses with rows and columns labeled by diagnoses abbreviations.

Model Investigation for Single Label Twelve Lead ECG Classification 161

Recurrent Neural Network Models
RNN with Attention Mechanism
ResNet with RNN
ResNet with RNN and Attention Mechanism
Observations

A classifier that returns only positive results will usually receive a negative score, i.e. a lower score than a classifier that returns only negative results, reflecting the harm of false alarms.