Conception and realization of an IoT-enabled deep CNN decision support system for automated arrhythmia classification

3.1 Remote ECG processing system

It is critical for a heart patient to be monitored regularly to detect even minor variations in the heartbeat and receive proper care as needed. In this article, we propose a remote ECG processing and monitoring system that allows patients and caregivers to monitor heart conditions from the comfort of their own homes. A continuous ECG monitoring device with a wireless module is in charge of collecting data. These data are transmitted to a remote server, where they are analysed and stored for future reference. An ESP8266 Wi-Fi-enabled module along with standard ECG electrodes and instrumentation amplifier is used. The proposed arrhythmia detection model with sliding window implementation can automatically classify the beats into four classes. If an abnormality is observed the caregiver is alerted, thus ensuring continuous monitoring, increasing the patient’s access to expert heart care. Figure 1 illustrates the remote ECG processing system.

Figure 1

Remote ECG processing system.

3.2 MIT-BIH database

The MIT-BIH arrhythmia database contains 48 half-hour ECG excerpts from two-channel recordings of 47 patients collected between 1975 and 1979 by the BIH Arrhythmia Laboratory [50]. The first 23 recordings are chosen at random, and the rest are selected to include rare arrhythmia. The records are digitized with an 11-bit resolution over a 10 mV range at 360 samples per second per channel. The recordings are annotated independently by two or more cardiologists and saved as reference annotations for each beat (approximately 110,000 annotations) included with the database. Table 1 contains an index of the various beat annotations used in the MIT-BIH database [51]. Of the different beats given in Table 1, normal beat (N), left BBB beat (L), right BBB beat (R), and premature ventricular contraction (V) are chosen for classification because the number of samples for the other beat types was negligible compared to the highly sampled N beat, which would have resulted in imbalanced data issue.

To ensure that no part of the ECG is missed, the MIT-BIH data are sampled with a sliding window and an overlap size of 400 and 100, respectively. The sliding window size and overlap size are adjusted to include a complete ECG wave based on the sampling rate of the measured ECG signal. A total of 36,000 samples are obtained by taking 9,000 samples of each beat type, making it a balanced data set. The samples thus obtained are shuffled and then split into training and testing using an 80:20 ratio, as shown in Figure 2. For each beat type, 7,200 samples for training and 1,800 samples for testing are obtained. The training set is subjected to stratified tenfold cross-validation. The resulting accuracy is averaged over the ten folds. Stratified folds ensure that each test set will have the same distribution of classes or as close it can be.

Figure 2

The train-test split and the data flow through the proposed model.

3.3 Deep neural network

A set of algorithms to recognize patterns from unstructured data modelled after the human brain are called neural networks. A deep neural network is a neural network with more than two layers – higher complexity. Each layer of nodes uses the previous layer’s output features and trains on it to extract additional information. The features extracted from each successive layer become more complex as the previous layer’s information is aggregated and recombined. In contrast to the existing methods for beat classification, such as heartbeat segmentation and feature extraction, the proposed method does not require any preprocessing. The data are sampled and the only operation performed on it is normalization and one-hot encoding of the target data. Normalization converts data values to a standard scale, enabling the system to learn the features. One-hot encoding transforms a categorical value into a numerical value that is better suited for machine learning models. The proposed method eliminates preprocessing overhead while having lower complexity.

Deep learning is a time consuming and iterative process in which the various hyperparameters need to be tuned to better optimize the model without penalizing the cost function. Deep neural networks use optimization algorithms to optimize the cost function J , which is defined as:

The value of cost function J is the mean of the loss L between the predicted value y ˆ and actual value y . W is the weight matrix, b is the bias, and m is the number of samples [52]. The main objective of this work is to adopt a computationally efficient optimization algorithm, and works well even with noisy datasets. AdaM (Adaptive Momentum) [53] is a very powerful and fast optimizer and is well suited for the adopted objectives. Another advantage of AdaM is that each network parameter has its learning rate, which gets updated as learning progresses. In addition, the cumulative history of gradients is kept, which accelerates the learning process.

The activation function is the non-linear transformation performed over the input signal. This transformed output is used as input to the next layer of neurons. Activation functions enable back-propagation because the gradients are supplied with the error to update the weights and biases. ReLU makes computation efficient, simple, and it also speeds up training. It is defined as:

ReLU ensures that only a few neurons are activated at a time, resulting in a sparse network. Furthermore, ReLU is idempotent, which solves the vanishing gradient problem.

3.4 Proposed method

The model used for beat classification is of low complexity, with only two 1D convolution layers. Many existing models, such as ResNet and VGG-Net, with multiple layers, might lead to overfitting. The proposed sequential model consists of two layers of convolution, batch normalization, activation, and maxpooling. It is followed by a flatten layer and a dense layer with softmax activation. Though having more than two hidden layers is advantageous for learning complex representations in neural networks, the number of convolutions is limited to two to avoid overfitting. For real-time processing, the sliding window implementation is used. The sliding window size and overlap size can be adjusted according to the sampling rate of the measured ECG signal.

The samples extracted undergo normalization followed by shuffling. Shuffling makes sure that no bias occurs in the model. The data are split into train and test sets in an 80:20 ratio. The training set further undergoes stratified K fold cross-validation with K = 10 . The stratified K fold ensures that all types of samples are present in each fold. Early stopping and model checkpoint callbacks are used, with cross-entropy as the loss function. The training is done for 30 epochs with validation after each epoch. The Adam optimizer is used, with a learning rate of 1 × 1 0 − 3 . Table 2 lists the various parameters chosen for different layers of the proposed model. Figure 3 depicts the model architecture and specifies the output shape of each layer.

Keras is an open-source high-level neural network library written in Python. Keras, along with TensorFlow as the backend is used in the proposed work. Nvidia GeForce GTX 1050Ti GPU with 768 CUDA cores accelerates learning in Ubuntu 16.04 platform. TensorBoard is used to track and visualize parameters such as loss and accuracy.

Figure 3

Proposed model architecture containing the two convolution layers followed by the fully connected layer and softmax.

4.1 Evaluation metrics

The standard metrics used are accuracy, precision, recall, and F 1 score. Classification accuracy is defined as the percentage of correct predictions to total predictions.

When an observation is both positive and predicted to be positive, it is referred to as a true positive (TP). When an observation is positive but is predicted to be negative, this is referred to as a false negative (FN). True Negative (TN) occurs when an observation is negative and is predicted to be negative. False Positive (FP) is characterized by a negative observation but a positive prediction. The total number of correctly classified positive examples divided by the total number of positive examples is defined as recall.

Precision is calculated by dividing the total number of positively classified examples by the total number of positively predicted examples. A high precision value indicates that an example labelled as positive is, in fact, positive.

The F -measure is a metric that represents both precision and recall. The F -measure will always be closer to the smaller among precision and recall values.

4.2 Analysis

A training and testing accuracy of 99.09 and 99.03%, respectively, are achieved. The confusion matrix in Figure 4 shows that the N beat is identified with 100% accuracy while identifying L, R, and V beat with 98, 100, and 99% accuracy. Precision, recall, and F1 score all obtained an overall score of 0.99. Table 3 summarizes the precision, recall, and F1 scores. The four output classes are visualized in Figure 5 using t-distributed stochastic neighbour embedding (t-SNE). The t-SNE method is a dimensionality reduction technique for visualizing high-dimensional data in 2D and 3D maps. t-SNE employs gradient descent to transform the dataset into a low-dimensional space so that the joint probability distribution representing the data points is as similar to the one in the high dimension as possible. In this case, the dense layer output is fed into t-SNE to visualize the four classes.

Figure 4

Confusion matrix for the testing data.

Figure 5

t-SNE visualization of the four classes N, L, R, and V.

Table 4 shows a comparison of similar results with the proposed work using the MIT-BIH database. A 3-layer CNN-LSTM is investigated in ref. [54] to classify normal and abnormal arrhythmia with 83.4% accuracy. Brito et al. used AdaM optimizer with a 12-layer ResNet architecture to classify arrhythmias into four categories and achieve an accuracy of 83.00% [55]. A 9-layer CNN with three convolutions for classifying five different types of beats is studied in ref. [36] and obtained an accuracy of 89.07% with no data augmentations. Fujita et al. employed a similar approach in their work, with no preprocessing but requiring knowledge of feature extraction before training the model and achieving comparable results [56]. To classify ECG beats, Zhou et al. proposed a hybrid CNN - Extreme Learning Machine (ELM) method [57]. They used the Pan–Tompkins algorithm to determine the R peak, followed by CNN for feature extraction and ELM for classification and achieved 98.77% accuracy in distinguishing four classes. An architecture based on hybrid Alexnet-SVM is investigated in ref. [58] for classifying three types of ECG signals with 96.77% accuracy, which is higher than the conventional methods. For arrhythmia detection, Atal et al. used an optimization enabled deep CNN and achieved an accuracy of 93.19%. They examined the model with 50 and 100 convolutional layers since accuracy is proportional to the number of convolutional layers involved [10]. A two-stage hierarchical deep CNN classification system for 16 classes is proposed in ref. [59]. The first stage categorizes the ECG into the five classes N, S, V, F, and Q, while the second stage categorizes the beats into the subclasses specified by the AAMI standards. They achieved an average accuracy of 97.30%. Harrane and Belkhiri used CNN for feature extraction and LSTM for classification across six classes, achieving 98.60% accuracy [60].

The proposed work achieved relatively significant accuracy with an excellent F1 score. The novelty of this work is that it trains a lower complexity CNN model with no preprocessing for use with real-time streamed data. The ECG data are sent to a remote server for analysis. As a result, the patient is no longer required to be admitted to a care facility. In addition, there is no need for P wave or QRS complex position detection, R peak determination, or noise removal to perform the classification. Though the results are specific to GPU training, the same can run with remarkable speed on non-GPU devices. The only disadvantage is that it requires a large amount of data for performing the initial training. The lower number of convolution layers used makes it suitable to deploy on hardware.

An automated ECG heartbeat classifier based on 1D CNNs is proposed in this study. The feature extraction and classification blocks are combined into a single learning body in this method. The need for feature extraction and preprocessing for real-time implementation of heart monitoring is thus eliminated. The proposed method only uses 1D convolutions reducing computational complexity and making it suitable for hardware implementation. The model achieved good classification results with minimal error. The obtained results did not show overfitting, and the test set performed well in terms of accuracy and F1-score. The model is able to classify all four classes remarkably well. Compared with related works in which the authors perform a beat-to-beat classification using CNN on the MIT-BIH arrhythmia database, this work outperforms other approaches in terms of accuracy using a model with lower complexity.