Diagnose Diseases Anytime, Anywhere! A Proposed Model For Diagnosing Arrhythmias, Utilizing ECGs From Wearable Devices!
3 main points
✔️ Diagnosis of arrhythmias based on electrocardiogram (ECG) signals using wearable devices has received increasing attention in clinical time series data in the medical field.
✔️ In this study, we propose a lightweight neural network - KecNet - to design a suitable DL for wearable devices with resource constraints.
✔️ The results of the evaluation showed that the correct answer rate - ACC-, sensitivity - SEN-, and fit rate - PRE- were 99.31%, 99.45%, and 98.78 %, respectively.
KecNet: A Light Neural Network for Arrhythmia Classification Based on Knowledge Reinforcement
(Submitted on 24 Apr 2021)
Comments: J Healthc Eng
The images used in this article are from the paper, the introductory slides, or were created based on them.
Can wearable devices predict arrhythmias with high accuracy?
In this study, we aim to construct a model for arrhythmia classification under resource constraints based on the assumption that ECG is acquired by wearable devices.
Among cardiac diseases with high mortality risk, arrhythmia is the most common cause of stroke and cardiac death. In this context, the electrocardiogram (ECG) has become a popular tool for arrhythmia detection because it is non-invasive and easy to measure; however, the random nature of arrhythmias requires long-term patient monitoring On the other hand, arrhythmias are characterized by randomness, which requires monitoring patients over a long period - resulting in significant costs for the processing of a large number of ECGs. Wearable devices are therefore attracting attention because they are less costly, and arrhythmia diagnostic systems using these devices are expected to reduce processing costs and improve the accuracy of arrhythmia classification.
In this context, ECG classification algorithms based on DLs have attracted much attention. Automatic arrhythmia recognition is prone to errors due to the use of heartbeat features that are highly influenced by individual differences - DL, which can construct a classification model that takes these features into account, is attracting attention; on the other hand, DL has the problem that the model performance depends on the network size, On the other hand, DL has a problem that the model performance depends on the network size - the larger the network size, the better the performance; therefore, existing DL algorithms are not likely to be suitable for wearable devices with constraints on both computational power and memory capacity. Therefore, there is a need to construct lightweight networks that can work in resource-constrained environments.
In this work, we propose a lightweight DL architecture - KecNet - based on ECG signals, assuming the use of wearable devices. The features of the proposed method are the introduction of a CNN network -Sinc convolution- based on digital signal processing to fit each application and; the exploitation of additional parameters based on clinical knowledge of ECG features.
What is a wearable device?
In this section, we briefly describe the wearable devices that are the subject of analysis in this study.
The wearable device is a generic term for information devices that subjects wear and use. Wearable means wearable or wearable, and there are wristband/wristwatch-type, eyeglass-type, and other types. Since these devices can be worn, they can be used in many aspects of daily life, from exercising to bathing and sleeping - they are thought to be effective in recording exercises such as jogging and swimming, and in acquiring health information by capturing heartbeat, pulse, sleep time, etc. It is thought to be effective in obtaining health information by capturing exercise records such as jogging and swimming, heart rate, pulse, and sleeping time.
In this section, we describe the proposed model in this study.
In the proposed model, the learning step is carried out by adjusting the parameters for the optimization problem - see equation below.
f(∗, θf): a function that simulates the mapping between data and labels; θ: a parameter associated with the mapping f; L: a function describing the loss of assigning a prediction category to a sample x(i) with label y(i)
In related work, the main approach to improve the performance of classification models has been to increase the number of layers in the network and add nonlinear operations - increasing the expressive power; however, there are three problems with this approach: first, due to the increase in the number of layers, the number of parameters in the Second, the increase in the depth of the model may cause gradient vanishing - resulting in ineffective parameter updating; third, a large amount of training data is required to prevent over-training of the model. training data is required to prevent over-training of the model. Therefore, it is possible to achieve better classification performance based on shallow networks without increasing model complexity and training data. In this work, relevant domain knowledge is introduced to the design process of convolutional neural networks - CNNs - by exploiting the amplitude-frequency characteristics of bandpass filters to filter noise in the ECG signal; and the Extraction of additional parameters for features in ECG data.
The proposed model - see the figure below - consists of segmenting and normalizing the data; feeding the segmented data into KecNet - the proposed model; integrating the feature vectors into and inserting them into the softmax classifier.
The ECG is a mixture of electrical activity at various sites in the myocardium - and therefore, depending on the quality of the data is likely to contain multiple types of noise, including baseline drift, motion artifacts, and electromyographic interference. Against this background, optimization in the first convolutional layer of the CNN is important - this layer is responsible for directly processing the original ECG and assisting the subsequent convolutional layers to perform complex nonlinear representations of the data. In this work, we introduce a Sinc-convolution layer developed for speech recognition: this layer is constructed based on the parameterized cardinal sine-Sinc-function for bandpass filter design. While traditional CNNs learn all the parameters of the filter, Sinc-convolution - the lower formulation - defines and learns a pre-tunable filter bank g.
Sinc-convolution is characterized by a high selectivity of frequency response compared to CNN - it extracts components in a specific frequency range from a complex signal, which has the effect of improving robustness and readability.
Symbolic Representation in Rhythmic Features
In the analysis of discrete time series, sequences are often converted into practical symbols to simplify the analysis process. Among them, the coefficient of variation -CV- represents the degree of dispersion of the RR interval and is used to measure the regularity of the RR interval: since the R peak is also prominent in ECGs, the RR interval features are noise tolerant. Therefore, the optimization in the proposed model adds CVs to the network as a symbolic representation of rhythm features in addition to the CNN structure designed to extract spatial morphological features - this allows ECGs with abnormal rhythms to be more easily identified.
This section describes the evaluation of this study.
The performance of the proposed method - from the table below - confirms that the performance of Sinc convolution is better than standard convolution. Moreover, the parameters were reduced by about 80% compared to CNNs with the same structure. The model performance was also improved by 1-1.5% depending on the rhythm variation coefficient.
The parameter gain - equation below - was utilized as an indicator for resource reduction.
where PC: the number of parameters. In standard CNNs, PC increases with increasing L - the length of the filter - while PC is always constant in the Sinc-convolution layer; therefore, the longer L is, the more PC can be reduced by KecNet. The PC of KecNet is reduced by 80% compared to standard CNNs.
We also compared the classification performance of KecNet with classical CNNs - GoogleNet, MobileNet, and SqueezeNet - in the table below. The results showed that the classification performance of KecNet outperformed SqueezeNet and MobileNets, while it was lower than GoogleNet; on the other hand, in the PC evaluation, KecNet was about 50% better than SqueezeNet and MobileNet, and about 80% better than GoogleNet. KecNet was about 50% less than SqueezeNet and MobileNet, and about 80% less than GoogleNet.
Here, we added white noise to the data to investigate the robustness of KecNet: the change in model accuracy concerning the signal-to-noise ratio - SNR - from 0 to 60 dB was targeted.
The results showed that the accuracy of all models tended to increase with higher SNR; on the other hand, as the SNR decreased, the detection accuracy decreased for models other than the proposed model - thus indicating that the proposed method is more robust than the conventional CNN.
The objective of this study is set to design a DL for the utilization of resource-constrained wearable devices, and a lightweight neural network based on domain knowledge - KecNet - is constructed.
The number of patients with arrhythmia has been increasing in recent years, and diagnostic models using wearable devices have been attracting much attention. In this study, we proposed a learning model for the classification of arrhythmia based on the assumption that wearable devices have such limitations. The features of this model are twofold: introduction of physically interpretable Sinc-convolutional, reduction of the number of parameters in CNN; addition of rhythm variability coefficients to the network; clarification of correlated features in the Shallow-CNN. Improvement of network performance and clinical usefulness. The evaluation results were trained and tested on the MIT-Bih arrhythmia dataset with ECG data: ACC, SEN, and PRE were 99.31%, 99.45%, and 98.78%, respectively. The size of the neural network was reduced and the robustness against noise was improved.
On the other hand, the following issues need to be addressed: the need for validation with real data; the implementation of ECGs in wearable devices. In this study, the evaluation was based on the dataset obtained from the database - therefore, the effectiveness of the system in real clinical situations is unclear. Therefore, in the future, ECG recordings should be collected and annotated from real patients to target more types of disease classifications. In addition, since there are currently only a limited number of wearable devices with ECGs, it is necessary to develop, test, and improve the performance of ECG systems that can be used with these devices.
Categories related to this article