Neural architecture search for real-time quality assessment of wearable multi-lead ECG on mobile devices
Introduction
Wearable electrocardiogram (ECG) devices are of great significance to cardiovascular health, as they help medical experts achieve remote monitoring on the user’s cardiovascular events for the timely diagnosis of anomalies. Users can use wearable devices to record ECG signals, upload the signals to a cloud platform, and choose to pay for the medical diagnostic services. Wearable ECG signals are generally recorded by users in their daily lives with Mason-Likar system which are different from the hospital-recorded ECGs by cardiovascular experts with Wilson-Likar system [1]. The non-professional operation of users and the non-resting state of daily life may lead to complicated interferences of wearable ECG signals. Some of the complicated interferences would strongly affect the diagnosis of cardiovascular experts or the inference of artificial intelligence models, causing high false alarm rates and invalid healthcare costs. ECG signals are annotated as unacceptable cases if the complex interferences in these signals prevent physicians from obtaining valuable information for accurate diagnosis. In order to filter out the unacceptable ECG signals on mobile devices (as shown in Fig. 1), reliable and real-time quality assessment for all leads of wearable ECG data is essential. In this way, the users can be promptly reminded of adjusting wearing posture, thereby facilitating the diagnosis of cardiovascular diseases [2].
Hand-crafted features based algorithms and convolutional neural network (CNN) based methods are predominantly employed for the quality assessment of ECG data. Most of them have a same disadvantage, namely that only one overall label for a multi-lead signal is unable to provide the quality assessment result for a specific lead. Such a deficiency affects the acquisition of ECG signals with diagnostic value, as the users are often incapable of correctly adjusting the specific position.
Hand-crafted features based algorithms of ECG quality assessment depend on the features designed by cardiologists. The features can be divided into fiducial features [3], [4], [5], [6], [7], [8] and non-fiducial features [9], [10], [11]. However, the hand-crafted features are sensitive to the complex interferences of wearable ECG data. Besides, the task of classifying the noisy ECG data with useful diagnostic value (NU data, as shown in Fig. 2 (b)) and unacceptable ECG data (as depicted in Fig. 2 (c)) is tough due to the high similarities between them.
CNN-based methods in the quality assessment of ECG data [12], [13], [14], [15] achieve superior performance to hand-crafted features based algorithms. However, most of the existing CNNs are the cumbersome architectures which are difficult to run on mobile devices for real-time assessment. The model compression [16], [17], [18], [19] is commonly used for obtaining a lightweight architecture; however, it degrades the performance of quality assessment. Therefore, a compact CNN with promising performance needs to be designed manually, which is error-prone and time-consuming.
To search a network architecture with superior performance automatically, neural architecture search (NAS) methods [20], [21], [22], [23] have been proposed. In the field of ECG analysis, the Darts [24] has been used in arrhythmia classification [25]. However, in terms of the real-time quality assessment of wearable ECG data, the Darts is not lightweight enough and likewise suffers from high computational cost.
In this study, we use a resource-saving modified ProxylessNAS [26] to search a model with fewer parameters and lower computational cost for the real-time signal quality assessment of wearable ECG on mobile devices (as shown in Fig. 1).
The main contribution of this study is that we propose to apply a resource-saving modified NAS technique for the automatic and fast search of a lightweight and robust model, named as ECGQAnet, to enable the real-time quality assessment of wearable ECG on mobile devices. Furthermore, we propose an any-lead strategy to assess the quality of all leads in a multi-lead ECG signal using only a single model, with high accuracy.
Section snippets
Related work
Existing relevant works for the quality assessment of ECG data include the hand-crafted features based methods and the CNN-based methods.
Materials and method
Motivated by the need of assessing the quality of all leads in a wearable multi-lead ECG signal, three frameworks are presented (as shown in Fig. 4), including multiple single-channel-input models for the quality assessment of multiple leads, one multi-label classification model for the quality assessment of all leads, and one single-channel-input model for the quality assessment of any lead. To reduce the manual cost involved in obtaining the quality assessment model, ProxylessNAS [26] is
Training details
All models were implemented in Python 3.6 on a server with two Titan RTX 2080Ti GPUs. The L2 weight decay coefficient was 0.0001. The label smoothing rate was 0.1. The batch size was 256. The learning rate for SGD optimizer was initialized as 0.01 to search stage and 0.0001 to final training stage. The total searching epochs was 320, for finding the excellent architecture from the whole supernet. The total training epochs was 80, for fine-tuning the parameters of the searched model. Resulting
Result
In this section, we conducted three aspects of comparisons to evaluate the performance of our ECGQAnet, including: 1) ablation study, 2) performance comparison on wearable ECG dataset, and 3) performance comparison on PICC 2011 dataset.
Discussion
In this study, a lightweight and efficient network, ECGQAnet, is constructed by leveraging the automatic search ability under limited computation resource of ProxylessNAS [26] for the real-time quality assessment of wearable multi-lead ECG data. Contributed to the lightweight architecture and low computational cost, ECGQAnet can easily run on mobile devices to discard signals without diagnostic information by the quality assessment of all leads in multi-lead ECG data. Additionally, our ECGQAnet
Conclusion
In this study, we implement a resource-saving NAS method to construct ECGQAnet with a compact structure and low latency automatically for the quality assessment of wearable ECG data, at a low searching cost. ECGQAnet can be employed on mobile devices for quality assessment of wearable ECG signals, which is essential for reminding users of adjusting the wearing posture in real time, saving medical cost for patients, and reducing the burden on cardiovascular practitioners. It is noteworthy that
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgment
This work was supported by a grant from the National Key R&D Program of China (No. 2018YFC2001203)
References (50)
- et al.
Artefact detection and quality assessment of ambulatory ECG signals
Comput. Methods Programs Biomed.
(2019) - et al.
A machine learning approach to multi-level ECG signal quality classification
Comput. Methods Programs Biomed.
(2014) - et al.
SQI quality evaluation mechanism of single-lead ECG signal based on simple heuristic fusion and fuzzy comprehensive evaluation
Front. Physiol.
(2018) - et al.
ECG quality assessment based on hand-crafted statistics and deep-learned S-transform spectrogram features
Comput. Methods Programs Biomed.
(2021) - et al.
Self-supervised ECG pre-training
Biomed. Signal Process. Control
(2021) - et al.
Fundamental differences between the standard 12–lead electrocardiograph and the modified (Mason—Likar) exercise lead system
Eur. Heart J.
(1987) - et al.
A deep learning approach for featureless robust quality assessment of intermittent atrial fibrillation recordings from portable and wearable devices
Entropy
(2020) - et al.
Real-time biosignal quality analysis of ambulatory ECG for detection of myocardial ischemia
- et al.
Signal quality and false myocardial ischemia alarms in ambulatory electrocardiograms
- et al.
Signal quality indices and data fusion for determining clinical acceptability of electrocardiograms
Physiol. Meas.
(2012)
Electrocardiogram signal quality measures for unsupervised telehealth environments
Physiol. Meas.
ECG quality assessment for patient empowerment in mHealth applications
Real-time signal quality assessment for ECGs collected using mobile phones
ECG quality assessment based on multi-feature fusion2017 13th International Conference on Natural Computation
ECG quality assessment using 1D-convolutional neural network
Deep learning-based electrocardiogram signal noise detection and screening model
Healthcare Inf. Res.
Densely connected convolutional networks and signal quality analysis to detect atrial fibrillation using short single-lead ECG recordings
Noise rejection for wearable ECGs using modified frequency slice wavelet transform and convolutional neural networks
IEEE Access
On compressing deep models by low rank and sparse decomposition
Channel pruning for accelerating very deep neural networks
Quantized neural networks: Training neural networks with low precision weights and activations
J. Mach. Learn. Res.
Learning transferable architectures for scalable image recognition
Nas-bench-101: towards reproducible neural architecture search
Cited by (10)
A deformable CNN architecture for predicting clinical acceptability of ECG signal
2023, Biocybernetics and Biomedical EngineeringFGSQA-Net: A Weakly Supervised Approach to Fine-Grained Electrocardiogram Signal Quality Assessment
2023, IEEE Journal of Biomedical and Health InformaticsToward efficient neural architecture search with dynamic mapping-adaptive sampling for resource-limited edge device
2023, Neural Computing and Applications
- 1
Huixin Tan and Jiewei Lai contributed equally to this work.