Negative-ResNet: Noisy Ambulatory Electrocardiogram Signal Classification Scheme

Arrhythmia, also known as irregular heartbeat, is an important indicator for most of the heart diseases [4]. Many infrequent arrhythmias are harmless while some may cause serious damage or even death [4].

Electrocardiogram(ECG) is currently widely used for detecting arrhythmia by measuring the electrical activities generated by the heart as contracts [4, 5, 6, 7]. An ECG signal can be characterized by two features: peaks and intervals. Each peak (P, Q, R, S, T) has its normal amplitude, and each interval (P-R, R-R, Q-R-S, S-T) has its duration period. Fig. 2 illustrates an ECG signal, and Table I shows the ECG features and their typical values[8, 9].

However, ECG signal data are usually of large size, and it would be a huge workload for cardiologists to review all of them manually. Hence, over the past few decades, many researchers and engineers have devoted themselves to design robust and accurate equipment and algorithms for ECG based arrhythmia detection and classification.

Conventionally, ECG decomposition and hand-crafted manual features are used for ECG classification. In [10], [1] and [11], the wavelet transform was used to extract features of each QRS complex. In [12] and [13], adaptive filtering was designed and used for arrhythmia detection.

Even though good results were reported, these model-based feature extraction methods are not generic and robust. Individual differences in ECG signals are significant and may lead to poor performance for different individuals.

Neural networks which are able to extract high-level features are proved to be a powerful tool to learn distributions of a dataset and generalize from training data [14, 15]. And the authors of [16] and [17] both presented a machine learning based method which consisted of a 1-D Convolutional Neural Network (CNN) for real-time ECG classification. [17] used a nine-layer CNN to classify five classes of arrhythmia and reached an accuracy of 94% on MIT-BIH dataset. Even though [17] reached a very high accuracy, it failed to consider performance on individual patients. [16] presented a better way of introducing patient-specific training using transfer learning. And it is also based on MIT-BIH database.

Besides CNN, other neural network backbone models are also commonly used in ECG classification. In [18], a 3-layer Recurrent Neural Network was used for 5-type classification and an 85% of accuracy was performed. In [19, 20, 21], long short-term memory(LSTM) showed very promising result in arrhythmia detection and classification tasks.

Even though [12], [16], [10], [1] and [14] showed very significant improvement in classification than using naked eyes, they suffer from either patient-generation ability or adaptability to the real-world data.

In [22] and [23] , patient-specific ECG classifiers were built using transfer learning of 2-D convolutional neural networks. A convolutional neural network was trained and used for feature extraction first. Then an individual classifier was trained for each patient transferring the whole pre-trained model from the training set domain to a specific patient domain.

However, these researches were all built on the MIT-BIH dataset, which is a very classic public and wholesome dataset [24]. The data is clean and collected from a 24-hour monitoring Holter [25, 26], which has a significant difference from the data collected by an ambulatory IoT patch. In many Smart Health applications involving ECG, the collected data will not be as clean and high-quality as the MIT-BIH dataset and therefore requires extra data analysis.

In the era of big data, obtaining large-scale dataset with high quality and precise labels is one of the most important procedures for most of data analysis related projects, especially when data-driven methods such as machine learning and deep learning are applied. However, annotating such a large dataset usually requires tremendous human resources and sometimes even requires domain knowledge. As such, human mistakes occur very often due to the reasons mentioned above, which will lead to a dataset with a certain number of noisy labels. To solve this problem, numerous studies are done in the field of noisy label training in the past several decades. And here, we refer to this survey [27] as a comprehensive review.

Many recent studies have illustrate directly training the noisy labels will impair the training scheme and therefore negatively impact the classification accuracy [28]. To avoid this issue, there are 2 major types of noisy label training schemes: Noisy Model Based Methods and Noisy Model Free Methods.

Our dataset was collected by a device named iRealCare [63] as shown in Fig.3. It’s a portable ECG monitoring device with signals sampled at 250Hz. The patch is lightweight with only 13 grams. In addition, the patch has the characteristic of being able to perform long-term continuous monitoring, which can last 72 hours. As such, it has significant advantages over the traditional ECG Holter.

Our experiments were performed on an ECG signal dataset collected from 65 patients. 4,831,517 heartbeats were identified considering the position of QRS, among which heartbeats were divided into 6 categories,

Normal Heartbeat(N); Ventricular premature beat (V); Supraventricular premature beat(S); Atrial fibrillation(A); Electromagnetic Interference(E) and Motion Interference(Q). In order to construct a balanced dataset, each category contains the same number of data points. Therefore, no bias will be introduced to the neural network during training because of the different numbers of training samples for each category. In other words, each category is equally represented by data samples, which avoids the over-fitting problem for certain categories. Table II shows the number of six types of ECG signals.

As mentioned before, the dataset generated by the patch would be more noisier than the Holter due to the limitation of the IoT patch, which could generate a significant number of Electromagnetic Interference. In addition, a considerable amount of movement interference will be introduced when the patients wearing the patch every day. Besides, there might also be some errors in labelling as well.

A set of possible noisy labels are presented below:

For example, Fig. 4 shows the standard waveform of type V premature ventricular contraction [6].

In our dataset, there are a few signals that have very similar features as shown in Fig. 5. As such, we would consider it as a ”clean label”.

However, there are also a few ”noisy labels” in this dataset, such as the one shown in Fig. 6. It was labelled as type V, but it’s revised to be other type because it doesn’t show the significant features of type V (Ventricular premature beat).

To identify the QRS position, both low pass filtering and high pass filtering are used to mitigate the noise and locate the QRS. Low pass filters can be used to cancel the high pass noise efficiently while the high pass filters can limit the interference of P and T wave. After filtering using low pass and high pass, we used hand-crafted features to identify the QRS wave respectively.

Data normalization is an essential step before feeding data into neural network for training. In deep learning, data are usually normalized to a mean of close to zero, which will usually speed up the learning process and result in a faster convergence. The reasons are as follows.

We define the binary cross-entropy loss as,

$$L=Y\times log(W\cdot X)+(1-Y)\log(1-W\cdot X)$$

(1)

where $Y$ is the training label vector, $X$ is the input vector and $W$ is the weight matrix. If we take the partial derivative with respect to each weight, we have,

$$\frac{\partial L}{\partial w_{i}}=x_{i}\times(a-y)$$

(2)

where $x_{i}$ is corresponding activation output of this weight, $a$ is the network output probability and $y$ is the training label.

Weights update according to the following,

$$w_{i}:=w_{i}-\alpha\times\frac{\partial L}{\partial w_{i}}$$

(3)

where $\alpha$ is the learning rate.

Even though magnitudes are different, weights will be updated with some dependence on each other. And that usually indicates a slow convergence, because independent weight updates ensure all possible combinations of weights, which is crucial for approaching a local minimum.

Therefore, in this work, we normalized our data to the range of $-0.5$ to $0.5$ according to Eq. (4), which also makes sure the scale of our input data is consistent, suppressing the outliers.

$$x_{i}=\frac{x_{i}-\frac{\left\{\max{\left(x_{i}\right)}+\min{\left(x_{i}\right)}\right\}}{2}}{\max{\left(x_{i}\right)}-\min{\left(x_{i}\right)}}$$

(4)

Residual block is the building block of our proposed model. It contains a normal convolutional layer, except for a residual connection from the input of the block to output of the block. It was first proposed in [66] as a solution for vanishing gradient when neural network goes deeper. Similar to [66], let’s consider $M(x)$ as the underlying mapping which the residual block is supposed to learn and $x$ denotes input to the residual block. The idea of residual learning is that rather than learning the underlying mapping $M(x)$, we learn the residual function $F(x)=M(x)-x$. And the underlying mapping can be reconstructed by an identity connection $F(x)+x$, which simply adds input to the output.

It has been proved that residual function F(x) is easier to learn and with extra information passed from input [66], the model can be trained deeper and deeper. If the output dimension is different from the input dimension, for example filter number changes, a linear projection $W_{s}$ needs to perform at the identity connection as $F\left(x\right)+W_{s}\times x$ (where $W_{s}$ is a matrix with a dimension of a$\times$b), ensuring the identity connection’s dimension matches the output dimension, such that we can perform an element-wise addition.

Fig. 9 shows the structure of our proposed residual block. It consists of two 1-D convolution layers followed by Batch Normalization, ReLu activation and Dropout which improve the convergence of the training process. As can be seen, an identity connection is established from the input to the output with a simple addition operation.

This backbone is a residual block based deep 1-D Convolution neural network which consists of 12 convolutional layers as shown in Fig. 10. Residual blocks which consist of a novel residual connection from the input to the output propagating information to deeper layers are basic building blocks of this backbone model [66]. The number of filters in the residual block is double every four residual block and a max-pooling operation is performed on the residual connection such that the input is sub-sampled by half before adding to its output for every two residual blocks.

We have constructed a baseline model with the same structure but normal positive training scheme. The result of baseline model is shown in 11. The highest accuracy achieved at convergence is $71\%$.

The whole training process of the system is shown in Fig. 12. The raw data is firstly feed into the CNN. Later on, the dataset is divided into clean data and noisy data based on the confidence level. Here, the confidence level refers to the output probability of the model. If the output confidence level is higher than a certain threshold, it is considered as clean data, vice versa.

The clean data (high confidence level) would be trained positively, and the noisy data (low confidence level) would go through negative learning process. Loss calculated from both positive and negative learning will be added together, and then passed back for gradient descent. Weights will be updated at the same time.

Different confidence levels are tested in our experiments. As can be seen from the results of Table III, the best result can be achieved when the confidence level is set to $80\%$, which is around $10\%$ higher than the baseline model. The confidence level controls how many data samples should be trained positively and how many data should be trained negatively. Therefore, it reflects the level of noisy labels contained in the dataset. And the best confidence level varies depending on different datasets.

Note that originally there are six labels which consist of two interference labels: Type Q and Type E.

As such, the accuracy reaches $81\%$ when training with $80\%$ confidence level. The accuracy and loss plot of training with $80\%$ of confidence level is shown .

Note: The confidence level for Type A is set to be 50% such that few Type A sample will go through the Negative Learning. It is because Type A in our dataset is confirmed by our medical expert to be of high labeling correctness.

The confusion matrix for the baseline model is shown in Fig. 14. The precision and recall of the baseline model are shown in Table IV.

The confusion matrix, precision and recall of optimal Negative ResNet are shown in Fig. 15 and Table V, respectively.

Generally speaking, the result of negative training is $4-11\%$ better than the normal training process, especially for interference, where an increment of $11\%$ accuracy can be achieved.

However, for type A, the accuracy remains the same when going through negative training. This is mainly because for type A (Atrial fibrillation), the original labelling process is relatively accurate. As such, using negative learning in this case will not have a significant improvement in the accuracy of type A. It proves our idea that negative learning performs better in a noisy dataset.