Deep Learning Approaches to Detect Atrial Fibrillation Using Photoplethysmographic Signals: Algorithms Development Study

Learning From Alarms: A Robust Learning Approach for Accurate Photoplethysmography-Based Atrial Fibrillation Detection Using Eight Million Samples Labeled With Imprecise Arrhythmia Alarms

Atrial fibrillation (AF) is a common cardiac arrhythmia with serious health consequences if not detected and treated early. Detecting AF using wearable devices with photoplethysmography (PPG) sensors and deep neural networks has demonstrated some success using proprietary algorithms in commercial solutions. However, to improve continuous AF detection in ambulatory settings towards a population-wide screening use case, we face several challenges, one of which is the lack of large-scale labeled training data. To address this challenge, we propose to leverage AF alarms from bedside patient monitors to label concurrent PPG signals, resulting in the largest PPG-AF dataset so far (8.5 M 30-second records from 24,100 patients) and demonstrating a practical approach to build large labeled PPG datasets. Furthermore, we recognize that the AF labels thus obtained contain errors because of false AF alarms generated from imperfect built-in algorithms from bedside monitors. Dealing with label noise with unknown distribution characteristics in this case requires advanced algorithms. We, therefore, introduce and open-source a novel loss design, the cluster membership consistency (CMC) loss, to mitigate label errors. By comparing CMC with state-of-the-art methods selected from a noisy label competition, we demonstrate its superiority in handling label noise in PPG data, resilience to poor-quality signals, and computational efficiency.

Photoplethysmography based atrial fibrillation detection: a continually growing field

Objective. Atrial fibrillation (AF) is a prevalent cardiac arrhythmia associated with significant health ramifications, including an elevated susceptibility to ischemic stroke, heart disease, and heightened mortality. Photoplethysmography (PPG) has emerged as a promising technology for continuous AF monitoring for its cost-effectiveness and widespread integration into wearable devices. Our team previously conducted an exhaustive review on PPG-based AF detection before June 2019. However, since then, more advanced technologies have emerged in this field.Approach. This paper offers a comprehensive review of the latest advancements in PPG-based AF detection, utilizing digital health and artificial intelligence (AI) solutions, within the timeframe spanning from July 2019 to December 2022. Through extensive exploration of scientific databases, we have identified 57 pertinent studies.Significance. Our comprehensive review encompasses an in-depth assessment of the statistical methodologies, traditional machine learning techniques, and deep learning approaches employed in these studies. In addition, we address the challenges encountered in the domain of PPG-based AF detection. Furthermore, we maintain a dedicated website to curate the latest research in this area, with regular updates on a regular basis.

Continuous Atrial Fibrillation Monitoring From Photoplethysmography: Comparison Between Supervised Deep Learning and Heuristic Signal Processing

BACKGROUND

Continuous monitoring for atrial fibrillation (AF) using photoplethysmography (PPG) from smartwatches or other wearables is challenging due to periods of poor signal quality during motion or suboptimal wearing. As a result, many consumer wearables sample infrequently and only analyze when the user is at rest, which limits the ability to perform continuous monitoring or to quantify AF.

OBJECTIVES

This study aimed to compare 2 methods of continuous monitoring for AF in free-living patients: a well-validated signal processing (SP) heuristic and a convolutional deep neural network (DNN) trained on raw signal.

METHODS

We collected 4 weeks of continuous PPG and electrocardiography signals in 204 free-living patients. Both SP and DNN models were developed and validated both on holdout patients and an external validation set.

RESULTS

The results show that the SP model demonstrated receiver-operating characteristic area under the curve (AUC) of 0.972 (sensitivity 99.6%, specificity: 94.4%), which was similar to the DNN receiver-operating characteristic AUC of 0.973 (sensitivity 92.2, specificity: 95.5%); however, the DNN classified significantly more data (95% vs 62%), revealing its superior tolerance of tracings prone to motion artifact. Explainability analysis revealed that the DNN automatically suppresses motion artifacts, evaluates irregularity, and learns natural AF interbeat variability. The DNN performed better and analyzed more signal in the external validation cohort using a different population and PPG sensor (AUC, 0.994; 97% analyzed vs AUC, 0.989; 88% analyzed).

CONCLUSIONS

DNNs perform at least as well as SP models, classify more data, and thus may be better for continuous PPG monitoring.

Performance of Single-Lead Handheld Electrocardiograms for Atrial Fibrillation Screening in Primary Care: The VITAL-AF Trial

Background

Handheld single-lead electrocardiographic (1L ECG) devices are increasingly used for atrial fibrillation (AF) screening, but their real-world performance is not well understood.

Objectives

The purpose of this study was to quantify the diagnostic test characteristics of 1L ECG automated interpretations for prospective AF screening.

Methods

We calculated the diagnostic test characteristics of the AliveCor KardiaMobile 1L ECG (AliveCor, US) algorithm using unblinded cardiologist overread as the gold standard using single 30s tracings administered by medical assistants among individuals aged ≥65 years participating in the VITAL-AF trial (NCT03515057) of population-based AF screening embedded within routine primary care.

Results

A total of 14,230 individuals (mean age 74 ± 7 years, 60% women, 82% White) had 31,376 tracings reviewed by 13 cardiologists. A total of 24,906 (79.6%) tracings had an AliveCor interpretation of normal, 5,046 (16.1%) were unclassified, 797 (2.5%) were possible AF, and 573 (1.8%) were no analysis. Cardiologists read 808 (2.6%) tracings as AF. AliveCor possible AF had a PPV of 51.7% (95% CI: 47.8%-55.6%). AliveCor normal had an NPV of 99.8% (95% CI: 99.7%-99.8%). The AliveCor algorithm had an overall sensitivity of 51.0% (95% CI: 47.1%-54.9%) and a specificity of 98.7% (95% CI: 98.6%-98.9%). AliveCor tracings interpreted as unclassified (PPV 5.9%, 95% CI: 5.1%-6.7%) and no analysis (PPV 6.5%, 95% CI: 4.6%-8.9%) had low predictive values for AF and were increasingly prevalent at older ages (13.7% for age 65-69 years to 28.1% for age ≥85 years, P < 0.01).

Conclusions

In an older primary care population undergoing AF screening with handheld 1L ECGs, automated algorithm interpretations were sufficiently accurate to exclude the presence of AF but not to establish an AF diagnosis.

Continuous cuffless blood pressure monitoring using photoplethysmography-based PPG2BP-net for high intrasubject blood pressure variations

Continuous, comfortable, convenient (C3), and accurate blood pressure (BP) measurement and monitoring are needed for early diagnosis of various cardiovascular diseases. To supplement the limited C3 BP measurement of existing cuff-based BP technologies, though they may achieve reliable accuracy, cuffless BP measurement technologies, such as pulse transit/arrival time, pulse wave analysis, and image processing, have been studied to obtain C3 BP measurement. One of the recent cuffless BP measurement technologies, innovative machine-learning and artificial intelligence-based technologies that can estimate BP by extracting BP-related features from photoplethysmography (PPG)-based waveforms have attracted interdisciplinary attention of the medical and computer scientists owing to their handiness and effectiveness for both C3 and accurate, i.e., C3A, BP measurement. However, C3A BP measurement remains still unattainable because the accuracy of the existing PPG-based BP methods was not sufficiently justified for subject-independent and highly varying BP, which is a typical case in practice. To circumvent this issue, a novel convolutional neural network(CNN)- and calibration-based model (PPG2BP-Net) was designed by using a comparative paired one-dimensional CNN structure to estimate highly varying intrasubject BP. To this end, approximately [Formula: see text], [Formula: see text], and [Formula: see text] of 4185 cleaned, independent subjects from 25,779 surgical cases were used for training, validating, and testing the proposed PPG2BP-Net, respectively and exclusively (i.e., subject-independent modelling). For quantifying the intrasubject BP variation from an initial calibration BP, a novel 'standard deviation of subject-calibration centring (SDS)' metric is proposed wherein high SDS represents high intrasubject BP variation from the calibration BP and vice versa. PPG2BP-Net achieved accurately estimated systolic and diastolic BP values despite high intrasubject variability. In 629-subject data acquired after 20 minutes following the A-line (arterial line) insertion, low error mean and standard deviation of [Formula: see text] and [Formula: see text] for highly varying A-line systolic and diastolic BP values, respectively, where their SDSs are 15.375 and 8.745. This study moves one step forward in developing the C3A cuffless BP estimation devices that enable the push and agile pull services.

Log-Spectral Matching GAN: PPG-based Atrial Fibrillation Detection can be Enhanced by GAN-based Data Augmentation with Integration of Spectral Loss

Photoplethysmography (PPG) is a ubiquitous physiological measurement that detects beat-to-beat pulsatile blood volume changes and hence has a potential for monitoring cardiovascular conditions, particularly in ambulatory settings. A PPG dataset that is created for a particular use case is often imbalanced, due to a low prevalence of the pathological condition it targets to predict and the paroxysmal nature of the condition as well. To tackle this problem, we propose log-spectral matching GAN (LSM-GAN), a generative model that can be used as a data augmentation technique to alleviate the class imbalance in a PPG dataset to train a classifier. LSM-GAN utilizes a novel generator that generates a synthetic signal without a up-sampling process of input white noises, as well as adds the mismatch between real and synthetic signals in frequency domain to the conventional adversarial loss. In this study, experiments are designed focusing on examining how the influence of LSM-GAN as a data augmentation technique on one specific classification task - atrial fibrillation (AF) detection using PPG. We show that by taking spectral information into consideration, LSM-GAN as a data augmentation solution can generate more realistic PPG signals.

Intra Prediction Method for Depth Video Coding by Block Clustering through Deep Learning

In this paper, we propose an intra-picture prediction method for depth video by a block clustering through a neural network. The proposed method solves a problem that the block that has two or more clusters drops the prediction performance of the intra prediction for depth video. The proposed neural network consists of both a spatial feature prediction network and a clustering network. The spatial feature prediction network utilizes spatial features in vertical and horizontal directions. The network contains a 1D CNN layer and a fully connected layer. The 1D CNN layer extracts the spatial features for a vertical direction and a horizontal direction from a top block and a left block of the reference pixels, respectively. 1D CNN is designed to handle time-series data, but it can also be applied to find the spatial features by regarding a pixel order in a certain direction as a timestamp. The fully connected layer predicts the spatial features of the block to be coded through the extracted features. The clustering network finds clusters from the spatial features which are the outputs of the spatial feature prediction network. The network consists of 4 CNN layers. The first 3 CNN layers combine two spatial features in the vertical and horizontal directions. The last layer outputs the probabilities that pixels belong to the clusters. The pixels of the block are predicted by the representative values of the clusters that are the average of the reference pixels belonging to the clusters. For the intra prediction for various block sizes, the block is scaled to the size of the network input. The prediction result through the proposed network is scaled back to the original size. In network training, the mean square error is used as a loss function between the original block and the predicted block. A penalty for output values far from both ends is introduced to the loss function for clear network clustering. In the simulation results, the bit rate is saved by up to 12.45% under the same distortion condition compared with the latest video coding standard.

Machine learning for atrial fibrillation risk prediction in patients with sleep apnea and coronary artery disease

Background

Patients with sleep apnea (SA) and coronary artery disease (CAD) are at higher risk of atrial fibrillation (AF) than the general population. Our objectives were: to evaluate the role of CAD and SA in determining AF risk through cluster and survival analysis, and to develop a risk model for predicting AF.

Methods

Electronic medical record (EMR) database from 22,302 individuals including 10,202 individuals with AF, CAD, and SA, and 12,100 individuals without these diseases were analyzed using K-means clustering technique; k-nearest neighbor (kNN) algorithm and survival analysis. Age, sex, and diseases developed for each individual during 9 years were used for cluster and survival analysis.

Results

The risk models for AF, CAD, and SA were identified with high accuracy and sensitivity (0.98). Cluster analysis showed that CAD and high blood pressure (HBP) are the most prevalent diseases in the AF group, HBP is the most prevalent disease in CAD; and HBP and CAD are the most prevalent diseases in the SA group. Survival analysis demonstrated that individuals with HBP, CAD, and SA had a 1.5-fold increased risk of developing AF [hazard ratio (HR): 1.49, 95% CI: 1.18-1.87, p = 0.0041; HR: 1.46, 95% CI: 1.09-1.96, p = 0.01; HR: 1.54, 95% CI: 1.22-1.94, p = 0.0039, respectively] and individuals with chronic kidney disease (CKD) developed AF approximately 50% earlier than patients without these comorbidities in a period of 7 years (HR: 3.36, 95% CI: 1.46-7.73, p = 0.0023). Comorbidities that contributed to develop AF earlier in females compared to males in the group of 50-64 years were HBP (HR: 3.75 95% CI: 1.08-13, p = 0.04) CAD and SA in the group of 60-75 years were (HR: 2.4 95% CI: 1.18-4.86, p = 0.02; HR: 2.51, 95% CI: 1.14-5.52, p = 0.02, respectively).

Conclusion

Machine learning based algorithms demonstrated that CAD, SA, HBP, and CKD are significant risk factors for developing AF in a Latin-American population.

Non-Contact Atrial Fibrillation Detection From Face Videos by Learning Systolic Peaks

OBJECTIVE

We propose a non-contact approach for atrial fibrillation (AF) detection from face videos.

METHODS

Face videos, electrocardiography (ECG), and contact photoplethysmography (PPG) from 100 healthy subjects and 100 AF patients are recorded. Data recordings from healthy subjects are all labeled as healthy. Two cardiologists evaluated ECG recordings of patients and labeled each recording as AF, sinus rhythm (SR), or atrial flutter (AFL). We use the 3D convolutional neural network for remote PPG monitoring and propose a novel loss function (Wasserstein distance) to use the timing of systolic peaks from contact PPG as the label for our model training. Then a set of heart rate variability (HRV) features are calculated from the inter-beat intervals, and a support vector machine (SVM) classifier is trained with HRV features.

RESULTS

Our proposed method can accurately extract systolic peaks from face videos for AF detection. The proposed method is trained with subject-independent 10-fold cross-validation with 30s video clips and tested on two tasks. 1) Classification of healthy versus AF: the accuracy, sensitivity, and specificity are 96.00%, 95.36%, and 96.12%. 2) Classification of SR versus AF: the accuracy, sensitivity, and specificity are 95.23%, 98.53%, and 91.12%. In addition, we also demonstrate the feasibility of non-contact AFL detection.

CONCLUSION

We achieve good performance of non-contact AF detection by learning systolic peaks.

SIGNIFICANCE

non-contact AF detection can be used for self-screening of AF symptoms for suspectable populations at home or self-monitoring of AF recurrence after treatment for chronic patients.

Visual Reassessment with Flux-Interval Plot Configuration after Automatic Classification for Accurate Atrial Fibrillation Detection by Photoplethysmography

Atrial fibrillation (AFib) is a common type of arrhythmia that is often clinically asymptomatic, which increases the risk of stroke significantly but can be prevented with anticoagulation. The photoplethysmogram (PPG) has recently attracted a lot of attention as a surrogate for electrocardiography (ECG) on atrial fibrillation (AFib) detection, with its out-of-hospital usability for rapid screening or long-term monitoring. Previous studies on AFib detection via PPG signals have achieved good results, but were short of intuitive criteria like ECG p-wave absence or not, especially while using interval randomness to detect AFib suffering from conjunction with premature contractions (PAC/PVC). In this study, we newly developed a PPG flux (pulse amplitude) and interval plots-based methodology, simply comprising an irregularity index threshold of 20 and regression error threshold of 0.06 for the precise automatic detection of AFib. The proposed method with automated detection on AFib shows a combined sensitivity, specificity, accuracy, and precision of 1, 0.995, 0.995, and 0.952 across the 460 samples. Furthermore, the flux-interval plot configuration also acts as a very intuitive tool for visual reassessment to confirm the automatic detection of AFib by its distinctive plot pattern compared to other cardiac rhythms. The study demonstrated that exclusive 2 false-positive cases could be corrected after the reassessment. With the methodology’s background theory well established, the detection process automated and visualized, and the PPG sensors already extensively used, this technology is very user-friendly and convincing for promoted to in-house AFib diagnostics.

Multiclass Arrhythmia Detection and Classification From Photoplethysmography Signals Using a Deep Convolutional Neural Network

Background

Studies have reported the use of photoplethysmography signals to detect atrial fibrillation; however, the use of photoplethysmography signals in classifying multiclass arrhythmias has rarely been reported. Our study investigated the feasibility of using photoplethysmography signals and a deep convolutional neural network to classify multiclass arrhythmia types.

Methods and Results

ECG and photoplethysmography signals were collected simultaneously from a group of patients who underwent radiofrequency ablation for arrhythmias. A deep convolutional neural network was developed to classify multiple rhythms based on 10‐second photoplethysmography waveforms. Classification performance was evaluated by calculating the area under the microaverage receiver operating characteristic curve, overall accuracy, sensitivity, specificity, and positive and negative predictive values against annotations on the rhythm of arrhythmias provided by 2 cardiologists consulting the ECG results. A total of 228 patients were included; 118 217 pairs of 10‐second photoplethysmography and ECG waveforms were used. When validated against an independent test data set (23 384 photoplethysmography waveforms from 45 patients), the DCNN achieved an overall accuracy of 85.0% for 6 rhythm types (sinus rhythm, premature ventricular contraction, premature atrial contraction, ventricular tachycardia, supraventricular tachycardia, and atrial fibrillation); the microaverage area under the microaverage receiver operating characteristic curve was 0.978; the average sensitivity, specificity, and positive and negative predictive values were 75.8%, 96.9%, 75.2%, and 97.0%, respectively.

Conclusions

This study demonstrated the feasibility of classifying multiclass arrhythmias from photoplethysmography signals using deep learning techniques. The approach is attractive for population‐based screening and may hold promise for the long‐term surveillance and management of arrhythmia.

Registration

URL: www.chictr.org.cn. Identifier: ChiCTR2000031170.

Photoplethysmogram Analysis and Applications: An Integrative Review

Beyond its use in a clinical environment, photoplethysmogram (PPG) is increasingly used for measuring the physiological state of an individual in daily life. This review aims to examine existing research on photoplethysmogram concerning its generation mechanisms, measurement principles, clinical applications, noise definition, pre-processing techniques, feature detection techniques, and post-processing techniques for photoplethysmogram processing, especially from an engineering point of view. We performed an extensive search with the PubMed, Google Scholar, Institute of Electrical and Electronics Engineers (IEEE), ScienceDirect, and Web of Science databases. Exclusion conditions did not include the year of publication, but articles not published in English were excluded. Based on 118 articles, we identified four main topics of enabling PPG: (A) PPG waveform, (B) PPG features and clinical applications including basic features based on the original PPG waveform, combined features of PPG, and derivative features of PPG, (C) PPG noise including motion artifact baseline wandering and hypoperfusion, and (D) PPG signal processing including PPG preprocessing, PPG peak detection, and signal quality index. The application field of photoplethysmogram has been extending from the clinical to the mobile environment. Although there is no standardized pre-processing pipeline for PPG signal processing, as PPG data are acquired and accumulated in various ways, the recently proposed machine learning-based method is expected to offer a promising solution.

2021 Asia Pacific Heart Rhythm Society (APHRS) practice guidance on atrial fibrillation screening

In this paper, the Asia Pacific Heart Rhythm Society (APHRS) sought to provide practice guidance on AF screening based on recent evidence, with specific considerations relevant to the Asia-Pacific region. A key recommendation is opportunistic screening for people aged ≥65 years (all countries), with systematic screening to be considered for people aged ≥75 years or who have additional risk factors (all countries).

A Real-Time PPG Peak Detection Method for Accurate Determination of Heart Rate during Sinus Rhythm and Cardiac Arrhythmia

OBJECTIVE

We have developed a peak detection algorithm for accurate determination of heart rate, using photoplethysmographic (PPG) signals from a smartwatch, even in the presence of various cardiac rhythms, including normal sinus rhythm (NSR), premature atrial contraction (PAC), premature ventricle contraction (PVC), and atrial fibrillation (AF). Given the clinical need for accurate heart rate estimation in patients with AF, we developed a novel approach that reduces heart rate estimation errors when compared to peak detection algorithms designed for NSR.

METHODS

Our peak detection method is composed of a sequential series of algorithms that are combined to discriminate the various arrhythmias described above. Moreover, a novel Poincaré plot scheme is used to discriminate between basal heart rate AF and rapid ventricular response (RVR) AF, and to differentiate PAC/PVC from NSR and AF. Training of the algorithm was performed only with Samsung Simband smartwatch data, whereas independent testing data which had more samples than did the training data were obtained from Samsung's Gear S3 and Galaxy Watch 3.

RESULTS

The new PPG peak detection algorithm provides significantly lower average heart rate and interbeat interval beat-to-beat estimation errors-30% and 66% lower-and mean heart rate and mean interbeat interval estimation errors-60% and 77% lower-when compared to the best of the seven other traditional peak detection algorithms that are known to be accurate for NSR. Our new PPG peak detection algorithm was the overall best performers for other arrhythmias.

CONCLUSION

The proposed method for PPG peak detection automatically detects and discriminates between various arrhythmias among different waveforms of PPG data, delivers significantly lower heart rate estimation errors for participants with AF, and reduces the number of false negative peaks.

SIGNIFICANCE

By enabling accurate determination of heart rate despite the presence of AF with rapid ventricular response or PAC/PVCs, we enable clinicians to make more accurate recommendations for heart rate control from PPG data.

Machine Learning for Cardiovascular Outcomes From Wearable Data: Systematic Review From a Technology Readiness Level Point of View

BACKGROUND

Wearable technology has the potential to improve cardiovascular health monitoring by using machine learning. Such technology enables remote health monitoring and allows for the diagnosis and prevention of cardiovascular diseases. In addition to the detection of cardiovascular disease, it can exclude this diagnosis in symptomatic patients, thereby preventing unnecessary hospital visits. In addition, early warning systems can aid cardiologists in timely treatment and prevention.

OBJECTIVE

This study aims to systematically assess the literature on detecting and predicting outcomes of patients with cardiovascular diseases by using machine learning with data obtained from wearables to gain insights into the current state, challenges, and limitations of this technology.

METHODS

We searched PubMed, Scopus, and IEEE Xplore on September 26, 2020, with no restrictions on the publication date and by using keywords such as "wearables," "machine learning," and "cardiovascular disease." Methodologies were categorized and analyzed according to machine learning-based technology readiness levels (TRLs), which score studies on their potential to be deployed in an operational setting from 1 to 9 (most ready).

RESULTS

After the removal of duplicates, application of exclusion criteria, and full-text screening, 55 eligible studies were included in the analysis, covering a variety of cardiovascular diseases. We assessed the quality of the included studies and found that none of the studies were integrated into a health care system (TRL<6), prospective phase 2 and phase 3 trials were absent (TRL<7 and 8), and group cross-validation was rarely used. These issues limited these studies' ability to demonstrate the effectiveness of their methodologies. Furthermore, there seemed to be no agreement on the sample size needed to train these studies' models, the size of the observation window used to make predictions, how long participants should be observed, and the type of machine learning model that is suitable for predicting cardiovascular outcomes.

CONCLUSIONS

Although current studies show the potential of wearables to monitor cardiovascular events, their deployment as a diagnostic or prognostic cardiovascular clinical tool is hampered by the lack of a realistic data set and proper systematic and prospective evaluation.

Recent Advances in Materials and Flexible Sensors for Arrhythmia Detection

Arrhythmias are one of the leading causes of death in the United States, and their early detection is essential for patient wellness. However, traditional arrhythmia diagnosis by expert evaluation from intermittent clinical examinations is time-consuming and often lacks quantitative data. Modern wearable sensors and machine learning algorithms have attempted to alleviate this problem by providing continuous monitoring and real-time arrhythmia detection. However, current devices are still largely limited by the fundamental mismatch between skin and sensor, giving way to motion artifacts. Additionally, the desirable qualities of flexibility, robustness, breathability, adhesiveness, stretchability, and durability cannot all be met at once. Flexible sensors have improved upon the current clinical arrhythmia detection methods by following the topography of skin and reducing the natural interface mismatch between cardiac monitoring sensors and human skin. Flexible bioelectric, optoelectronic, ultrasonic, and mechanoelectrical sensors have been demonstrated to provide essential information about heart-rate variability, which is crucial in detecting and classifying arrhythmias. In this review, we analyze the current trends in flexible wearable sensors for cardiac monitoring and the efficacy of these devices for arrhythmia detection.

Contactless facial video recording with deep learning models for the detection of atrial fibrillation

Atrial fibrillation (AF) is often asymptomatic and paroxysmal. Screening and monitoring are needed especially for people at high risk. This study sought to use camera-based remote photoplethysmography (rPPG) with a deep convolutional neural network (DCNN) learning model for AF detection. All participants were classified into groups of AF, normal sinus rhythm (NSR) and other abnormality based on 12-lead ECG. They then underwent facial video recording for 10 min with rPPG signals extracted and segmented into 30-s clips as inputs of the training of DCNN models. Using voting algorithm, the participant would be predicted as AF if > 50% of their rPPG segments were determined as AF rhythm by the model. Of the 453 participants (mean age, 69.3 ± 13.0 years, women, 46%), a total of 7320 segments (1969 AF, 1604 NSR & 3747others) were analyzed by DCNN models. The accuracy rate of rPPG with deep learning model for discriminating AF from NSR and other abnormalities was 90.0% and 97.1% in 30-s and 10-min recording, respectively. This contactless, camera-based rPPG technique with a deep-learning model achieved significantly high accuracy to discriminate AF from non-AF and may enable a feasible way for a large-scale screening or monitoring in the future.

Artificial Intelligence for Detection of Cardiovascular-Related Diseases from Wearable Devices: A Systematic Review and Meta-Analysis

PURPOSE

Several artificial intelligence (AI) models for the detection and prediction of cardiovascular-related diseases, including arrhythmias, diabetes, and sleep apnea, have been reported. This systematic review and meta-analysis aimed to identify AI models developed for or applicable to wearable and mobile devices for diverse cardiovascular-related diseases.

MATERIALS AND METHODS

The searched databases included Medline, Embase, and Cochrane Library. For AI models for atrial fibrillation (AF) detection, a meta-analysis of diagnostic accuracy was performed to summarize sensitivity and specificity.

RESULTS

A total of 102 studies were included in the qualitative review. There were AI models for the detection of arrythmia (n=62), followed by sleep apnea (n=11), peripheral vascular diseases (n=6), diabetes mellitus (n=5), hyper/hypotension (n=5), valvular heart disease (n=4), heart failure (n=3), myocardial infarction and cardiac arrest (n=2), and others (n=4). For quantitative analysis of 26 studies reporting AI models for AF detection, meta-analyzed sensitivity was 94.80% and specificity was 96.96%. Deep neural networks showed superior performance [meta-analyzed area under receiver operating characteristics curve (AUROC) of 0.981] compared to conventional machine learning algorithms (meta-analyzed AUROC of 0.961). However, AI models tested with proprietary dataset (meta-analyzed AUROC of 0.972) or data acquired from wearable devices (meta-analyzed AUROC of 0.977) showed inferior performance than those with public dataset (meta-analyzed AUROC of 0.986) or data from in-hospital devices (meta-analyzed AUROC of 0.983).

CONCLUSION

This review found that AI models for diverse cardiovascular-related diseases are being developed, and that they are gradually developing into a form that is suitable for wearable and mobile devices.

An application for classifying perceptions on my health bank in Taiwan using convolutional neural networks and web-based computerized adaptive testing: A development and usability study

BACKGROUND

The classification of a respondent's opinions online into positive and negative classes using a minimal number of questions is gradually changing and helps turn techniques into practices. A survey incorporating convolutional neural networks (CNNs) into web-based computerized adaptive testing (CAT) was used to collect perceptions on My Health Bank (MHB) from users in Taiwan. This study designed an online module to accurately and efficiently turn a respondent's perceptions into positive and negative classes using CNNs and web-based CAT.

METHODS

In all, 640 patients, family members, and caregivers with ages ranging from 20 to 70 years who were registered MHB users were invited to complete a 3-domain, 26-item, 5-category questionnaire asking about their perceptions on MHB (PMHB26) in 2019. The CNN algorithm and k-means clustering were used for dividing respondents into 2 classes of unsatisfied and satisfied classes and building a PMHB26 predictive model to estimate parameters. Exploratory factor analysis, the Rasch model, and descriptive statistics were used to examine the demographic characteristics and PMHB26 factors that were suitable for use in CNNs and Rasch multidimensional CAT (MCAT). An application was then designed to classify MHB perceptions.

RESULTS

We found that 3 construct factors were extracted from PMHB26. The reliability of PMHB26 for each subscale beyond 0.94 was evident based on internal consistency and stability in the data. We further found the following: the accuracy of PMHB26 with CNN yields a higher accuracy rate (0.98) with an area under the curve of 0.98 (95% confidence interval, 0.97-0.99) based on the 391 returned questionnaires; and for the efficiency, approximately one-third of the items were not necessary to answer in reducing the respondents' burdens using Rasch MCAT.

CONCLUSIONS

The PMHB26 CNN model, combined with the Rasch online MCAT, is recommended for improving the accuracy and efficiency of classifying patients' perceptions of MHB utility. An application developed for helping respondents self-assess the MHB cocreation of value can be applied to other surveys in the future.

Atrial Fibrillation Classification with Smart Wearables Using Short-Term Heart Rate Variability and Deep Convolutional Neural Networks

Atrial fibrillation (AF) is a type of cardiac arrhythmia affecting millions of people every year. This disease increases the likelihood of strokes, heart failure, and even death. While dedicated medical-grade electrocardiogram (ECG) devices can enable gold-standard analysis, these devices are expensive and require clinical settings. Recent advances in the capabilities of general-purpose smartphones and wearable technology equipped with photoplethysmography (PPG) sensors increase diagnostic accessibility for most populations. This work aims to develop a single model that can generalize AF classification across the modalities of ECG and PPG with a unified knowledge representation. This is enabled by approximating the transformation of signals obtained from low-cost wearable PPG sensors in terms of Pulse Rate Variability (PRV) to temporal Heart Rate Variability (HRV) features extracted from medical-grade ECG. This paper proposes a one-dimensional deep convolutional neural network that uses HRV-derived features for classifying 30-s heart rhythms as normal sinus rhythm or atrial fibrillation from both ECG and PPG-based sensors. The model is trained with three MIT-BIH ECG databases and is assessed on a dataset of unseen PPG signals acquired from wrist-worn wearable devices through transfer learning. The model achieved the aggregate binary classification performance measures of accuracy: 95.50%, sensitivity: 94.50%, and specificity: 96.00% across a five-fold cross-validation strategy on the ECG datasets. It also achieved 95.10% accuracy, 94.60% sensitivity, 95.20% specificity on an unseen PPG dataset. The results show considerable promise towards seamless adaptation of gold-standard ECG trained models for non-ambulatory AF detection with consumer wearable devices through HRV-based knowledge transfer.

Considerations on Performance Evaluation of Atrial Fibrillation Detectors

OBJECTIVE

A large number of atrial fibrillation (AF) detectors have been published in recent years, signifying that the comparison of detector performance plays a central role, though not always consistent. The aim of this study is to shed needed light on aspects crucial to the evaluation of detection performance.

METHODS

Three types of AF detector, using either information on rhythm, rhythm and morphology, or segments of ECG samples, are implemented and studied on both real and simulated ECG signals. The properties of different performance measures are investigated, for example, in relation to dataset imbalance.

RESULTS

The results show that performance can differ considerably depending on the way detector output is compared to database annotations, i.e., beat-to-beat, segment-to-segment, or episode-to-episode comparison. Moreover, depending on the type of detector, the results substantiate that physiological and technical factors, e.g., changes in ECG morphology, rate of atrial premature beats, and noise level, can have a considerable influence on performance.

CONCLUSION

The present study demonstrates overall strengths and weaknesses of different types of detector, highlights challenges in AF detection, and proposes five recommendations on how to handle data and characterize performance.

Wearables in cardiology: Here to stay

The adoption of wearables in medicine has rapidly expanded worldwide. New generations of wearables are emerging, driven by consumers' demand to monitor their own health. With the ongoing development of new features capable of assessing real-time biometric data, the impact of wearables on cardiovascular management has become inevitable. Smartwatches, among other wearable devices, offer a user-friendly noninvasive approach to continuously monitor for health parameters. With advancements in artificial intelligence, the photoplethysmography-generated pulse waveform has the potential to accurately detect episodes of atrial fibrillation and one day could replace conventional diagnostic and long-term monitoring methods. Clinical benefits that could arise from the use of such devices include refining stroke prevention strategies, personalizing AF management, and optimizing the patient-physician relationship. Wearables are changing not only the way clinicians conduct research but also the future of cardiovascular preventive and therapeutic care. As such, wearables are here to stay.

Explainable Artificial Intelligence Model for Diagnosis of Atrial Fibrillation Using Holter Electrocardiogram Waveforms

Atrial fibrillation is a clinically important arrhythmia. There are some reports on machine learning models for AF diagnosis using electrocardiogram data. However, few reports have proposed an eXplainable Artificial Intelligence (XAI) model to enable physicians to easily understand the machine learning model's diagnosis results.We developed and validated an XAI-enabled atrial fibrillation diagnosis model based on a convolutional neural network (CNN) algorithm. We used Holter electrocardiogram monitoring data and the gradient-weighted class activation mapping (Grad-CAM) method.Electrocardiogram data recorded from patients between January 4, 2016, and October 31, 2019, totaling 57,273 electrocardiogram waveform slots of 30 seconds each with diagnostic information annotated by cardiologists, were used for training our proposed model. Performance metrics of our AI model for AF diagnosis are as follows: sensitivity, 97.1% (95% CI: 0.969-0.972); specificity, 94.5% (95% CI: 0.943-0.946); accuracy, 95.3% (95% CI: 0.952-0.955); positive predictive value, 89.3% (95% CI: 0.892-0.897); and F-value, 93.1% (95% CI: 0.929-0.933). The area under the receiver operating characteristic curve for AF detection using our model was 0.988 (95% CI: 0.987-0.988). Furthermore, using the XAI method, 94.5 ± 3.5% of the areas identified as regions of interest using our machine learning model were identified as characteristic sites for AF diagnosis by cardiologists.AF was accurately diagnosed and favorably explained with Holter ECG waveforms using our proposed CNN-based XAI model. Our study presents another step toward realizing a viable XAI-based detection model for AF diagnoses for use by physicians.

The application of deep learning in electrocardiogram: Where we came from and where we should go?

Electrocardiogram (ECG) is a commonly-used, non-invasive examination recording cardiac voltage versus time traces over a period. Deep learning technology, a robust artificial intelligence algorithm, can imitate the data processing patterns of the human brain, and it has experienced remarkable success in disease screening, diagnosis, and prediction. Compared with traditional machine learning, deep learning algorithms possess more powerful learning capabilities and can automatically extract features without extensive data pre-processing or hand-crafted feature extraction, which makes it a suitable tool to analyze complex structures of high-dimensional data. With the advances in computing power and digitized data availability, deep learning provides us an opportunity to improve ECG data interpretation with higher efficacy and accuracy and, more importantly, expand the original functions of ECG. The application of deep learning has led us to stand at the edge of ECG innovation and will potentially change the current clinical monitoring and management strategies. In this review, we introduce deep learning technology and summarize its advantages compared with traditional machine learning algorithms. Moreover, we provide an overview on the current application of deep learning in ECGs, with a focus on arrhythmia (especially atrial fibrillation during normal sinus rhythm), cardiac dysfunction, electrolyte imbalance, and sleep apnea. Last but not least, we discuss the current challenges and prospect directions for the following studies.

How machine learning is impacting research in atrial fibrillation: implications for risk prediction and future management

There has been an exponential growth of artificial intelligence (AI) and machine learning (ML) publications aimed at advancing our understanding of atrial fibrillation (AF), which has been mainly driven by the confluence of two factors: the advances in deep neural networks (DeepNNs) and the availability of large, open access databases. It is observed that most of the attention has centred on applying ML for dvsetecting AF, particularly using electrocardiograms (ECGs) as the main data modality. Nearly a third of them used DeepNNs to minimize or eliminate the need for transforming the ECGs to extract features prior to ML modelling; however, we did not observe a significant advantage in following this approach. We also found a fraction of studies using other data modalities, and others centred in aims, such as risk prediction, AF management, and others. From the clinical perspective, AI/ML can help expand the utility of AF detection and risk prediction, especially for patients with additional comorbidities. The use of AI/ML for detection and risk prediction into applications and smart mobile health (mHealth) technology would enable ‘real time’ dynamic assessments. AI/ML could also adapt to treatment changes over time, as well as incident risk factors. Incorporation of a dynamic AI/ML model into mHealth technology would facilitate ‘real time’ assessment of stroke risk, facilitating mitigation of modifiable risk factors (e.g. blood pressure control). Overall, this would lead to an improvement in clinical care for patients with AF.

Wrist Band Photoplethysmography Autocorrelation Analysis Enables Detection of Atrial Fibrillation Without Pulse Detection

Atrial fibrillation is often asymptomatic and intermittent making its detection challenging. A photoplethysmography (PPG) provides a promising option for atrial fibrillation detection. However, the shapes of pulse waves vary in atrial fibrillation decreasing pulse and atrial fibrillation detection accuracy. This study evaluated ten robust photoplethysmography features for detection of atrial fibrillation. The study was a national multi-center clinical study in Finland and the data were combined from two broader research projects (NCT03721601, URL: https://clinicaltrials.gov/ct2/show/NCT03721601 and NCT03753139, URL: https://clinicaltrials.gov/ct2/show/NCT03753139). A photoplethysmography signal was recorded with a wrist band. Five pulse interval variability, four amplitude features and a novel autocorrelation-based morphology feature were calculated and evaluated independently as predictors of atrial fibrillation. A multivariate predictor model including only the most significant features was established. The models were 10-fold cross-validated. 359 patients were included in the study (atrial fibrillation n = 169, sinus rhythm n = 190). The autocorrelation univariate predictor model detected atrial fibrillation with the highest area under receiver operating characteristic curve (AUC) value of 0.982 (sensitivity 95.1%, specificity 93.7%). Autocorrelation was also the most significant individual feature (p < 0.00001) in the multivariate predictor model, detecting atrial fibrillation with AUC of 0.993 (sensitivity 96.4%, specificity 96.3%). Our results demonstrated that the autocorrelation independently detects atrial fibrillation reliably without the need of pulse detection. Combining pulse wave morphology-based features such as autocorrelation with information from pulse-interval variability it is possible to detect atrial fibrillation with high accuracy with a commercial wrist band. Photoplethysmography wrist bands accompanied with atrial fibrillation detection algorithms utilizing autocorrelation could provide a computationally very effective and reliable wearable monitoring method in screening of atrial fibrillation.

Predicting Active NBA Players Most Likely to Be Inducted into the Basketball Hall of Famers Using Artificial Neural Networks in Microsoft Excel: Development and Usability Study

The prediction of whether active NBA players can be inducted into the Hall of Fame (HOF) is interesting and important. However, no such research have been published in the literature, particularly using the artificial neural network (ANN) technique. The aim of this study is to build an ANN model with an app for automatic prediction and classification of HOF for NBA players. We downloaded 4728 NBA players’ data of career stats and accolades from the website at basketball-reference.com. The training sample was collected from 85 HOF members and 113 retired Non-HOF players based on completed data and a longer career length (≥15 years). Featured variables were taken from the higher correlation coefficients (<0.1) with HOF and significant deviations apart from the two HOF/Non-HOF groups using logistical regression. Two models (i.e., ANN and convolutional neural network, CNN) were compared in model accuracy (e.g., sensitivity, specificity, area under the receiver operating characteristic curve, AUC). An app predicting HOF was then developed involving the model’s parameters. We observed that (1) 20 feature variables in the ANN model yielded a higher AUC of 0.93 (95% CI 0.93–0.97) based on the 198-case training sample, (2) the ANN performed better than CNN on the accuracy of AUC (= 0.91, 95% CI 0.87–0.95), and (3) an ready and available app for predicting HOF was successfully developed. The 20-variable ANN model with the 53 parameters estimated by the ANN for improving the accuracy of HOF has been developed. The app can help NBA fans to predict their players likely to be inducted into the HOF and is not just limited to the active NBA players.

Smart Wearables for Cardiac Monitoring-Real-World Use beyond Atrial Fibrillation

The possibilities and implementation of wearable cardiac monitoring beyond atrial fibrillation are increasing continuously. This review focuses on the real-world use and evolution of these devices for other arrhythmias, cardiovascular diseases and some of their risk factors beyond atrial fibrillation. The management of nonatrial fibrillation arrhythmias represents a broad field of wearable technologies in cardiology using Holter, event recorder, electrocardiogram (ECG) patches, wristbands and textiles. Implementation in other patient cohorts, such as ST-elevation myocardial infarction (STEMI), heart failure or sleep apnea, is feasible and expanding. In addition to appropriate accuracy, clinical studies must address the validation of clinical pathways including the appropriate device and clinical decisions resulting from the surrogate assessed.

Explainability Metrics of Deep Convolutional Networks for Photoplethysmography Quality Assessment

Photoplethysmography (PPG) is a noninvasive way to monitor various aspects of the circulatory system, and is becoming more and more widespread in biomedical processing. Recently, deep learning methods for analyzing PPG have also become prevalent, achieving state of the art results on heart rate estimation, atrial fibrillation detection, and motion artifact identification. Consequently, a need for interpretable deep learning has arisen within the field of biomedical signal processing. In this paper, we pioneer novel explanatory metrics which leverage domain-expert knowledge to validate a deep learning model. We visualize model attention over a whole testset using saliency methods and compare it to human expert annotations. Congruence, our first metric, measures the proportion of model attention within expert-annotated regions. Our second metric, Annotation Classification, measures how much of the expert annotations our deep learning model pays attention to. Finally, we apply our metrics to compare between a signal based model and an image based model for PPG signal quality classification. Both models are deep convolutional networks based on the ResNet architectures. We show that our signal-based one dimensional model acts in a more explainable manner than our image based model; on average 50.78% of the one dimensional model's attention are within expert annotations, whereas 36.03% of the two dimensional model's attention are within expert annotations. Similarly, when thresholding the one dimensional model attention, one can more accurately predict if each pixel of the PPG is annotated as artifactual by an expert. Through this testcase, we demonstrate how our metrics can provide a quantitative and dataset-wide analysis of how explainable the model is.

Prediction of vascular aging based on smartphone acquired PPG signals

Photoplethysmography (PPG) measured by smartphone has the potential for a large scale, non-invasive, and easy-to-use screening tool. Vascular aging is linked to increased arterial stiffness, which can be measured by PPG. We investigate the feasibility of using PPG to predict healthy vascular aging (HVA) based on two approaches: machine learning (ML) and deep learning (DL). We performed data preprocessing, including detrending, demodulating, and denoising on the raw PPG signals. For ML, ridge penalized regression has been applied to 38 features extracted from PPG, whereas for DL several convolutional neural networks (CNNs) have been applied to the whole PPG signals as input. The analysis has been conducted using the crowd-sourced Heart for Heart data. The prediction performance of ML using two features (AUC of 94.7%) – the a wave of the second derivative PPG and tpr, including four covariates, sex, height, weight, and smoking – was similar to that of the best performing CNN, 12-layer ResNet (AUC of 95.3%). Without having the heavy computational cost of DL, ML might be advantageous in finding potential biomarkers for HVA prediction. The whole workflow of the procedure is clearly described, and open software has been made available to facilitate replication of the results.

Digital Image Processing Features of Smartwatch Photoplethysmography for Cardiac Arrhythmia Detection

The aim of our work is to design an algorithm to detect premature atrial contraction (PAC), premature ventricular contraction (PVC), and atrial fibrillation (AF) among normal sinus rhythm (NSR) using smartwatch photoplethysmographic (PPG) data. Novel image processing features and two machine learning methods are used to enhance the PAC/PVC detection results of the Poincaré plot method. Compared with support vector machine (SVM) methods, the Random Forests (RF) method performs better. It yields a 10-fold cross validation (CV) averaged sensitivity, specificity, positive predicted value (PPV), negative predicted value (NPV), and accuracy for PAC/PVC labels of 63%, 98%, 83%, 94%, and 93%, respectively, and a 10-fold CV averaged sensitivity, specificity, PPV, NPV, and accuracy for AF subjects of 92%, 96%, 85%, 98%, and 95%, respectively. This is one of the first studies to derive image processing features from Poincaré plots to further enhance the accuracy of PAC/PVC detection using PPG recordings from a smartwatch.

Premature Atrial and Ventricular Contraction Detection using Photoplethysmographic Data from a Smartwatch

We developed an algorithm to detect premature atrial contraction (PAC) and premature ventricular contraction (PVC) using photoplethysmographic (PPG) data acquired from a smartwatch. Our PAC/PVC detection algorithm is composed of a sequence of algorithms that are combined to discriminate various arrhythmias. A novel vector resemblance method is used to enhance the PAC/PVC detection results of the Poincaré plot method. The new PAC/PVC detection algorithm with our automated motion and noise artifact detection approach yielded a sensitivity of 86% for atrial fibrillation (AF) subjects while the overall sensitivity was 67% when normal sinus rhythm (NSR) subjects were also included. The specificity, positive predictive value (PPV), negative predictive value (NPV) and accuracy values for the combined data consisting of both NSR and AF subjects were 97%, 81%, 94% and 92%, respectively, for PAC/PVC detection combined with our automated motion and noise artifact detection approach. Moreover, when AF detection was compared with and without PAC/PVC, the sensitivity and specificity increased from 94.55% to 98.18% and from 95.75% to 97.90%, respectively. For additional independent testing data, we used two datasets: a smartwatch PPG dataset that was collected in our ongoing clinical study, and a pulse oximetry PPG dataset from the Medical Information Mart for Intensive Care III database. The PAC/PVC classification results of the independent testing on these two other datasets are all above 92% for sensitivity, specificity, PPV, NPV, and accuracy. The proposed combined approach to detect PAC and PVC can ultimately lead to better accuracy in AF detection. This is one of the first studies involving detection of PAC and PVC using PPG recordings from a smartwatch. The proposed method can potentially be of clinical importance as this enhanced capability can lead to fewer false positive detections of AF, especially for those NSR subjects with frequent episodes of PAC/PVC.

Diagnostic accuracy of smart gadgets/wearable devices in detecting atrial fibrillation: A systematic review and meta-analysis

BACKGROUND

Recently, smart devices have been used for medical purposes, particularly to screen for atrial fibrillation. However, current data on the diagnostic performance of these devices are scarce.

AIMS

We performed a systemic review and meta-analysis to assess the accuracy of atrial fibrillation diagnosis by smart gadgets/wearable devices.

METHODS

We comprehensively searched the MEDLINE, EMBASE and Cochrane databases for all works since the inception of each database until January 2020. Included in this review were published observational studies of the diagnostic accuracy of smartphones or smartwatches in detecting atrial fibrillation. We calculated the area under the summary receiver operating characteristic curves and pooled sensitivities and specificities.

RESULTS

Participants in our study were from the general population or were patients with underlying atrial fibrillation. In the overall analyses, the areas under the summary receiver operating characteristic curves were 0.96 and 0.94 for smartphones and smartwatches, respectively. Smartphones had a sensitivity of 94% and a specificity of 96%, and smartwatches showed similar diagnostic accuracy, with a specificity of 94% and a sensitivity of 93%. In subgroup analyses, we found no difference in diagnostic accuracy between photoplethysmography and single-lead electrocardiography.

CONCLUSIONS

This study suggests that smart devices have similar diagnostic accuracies. Regarding atrial fibrillation detection methods, there was also no difference between photoplethysmography and single-lead electrocardiography. However, further studies are warranted to determine their clinical implications in atrial fibrillation management.

Detecting undiagnosed atrial fibrillation in UK primary care: Validation of a machine learning prediction algorithm in a retrospective cohort study

AIMS

To evaluate the ability of a machine learning algorithm to identify patients at high risk of atrial fibrillation in primary care.

METHODS

A retrospective cohort study was undertaken using the DISCOVER registry to validate an algorithm developed using a Clinical Practice Research Datalink (CPRD) dataset. The validation dataset included primary care patients in London, England aged ≥30 years from 1 January 2006 to 31 December 2013, without a diagnosis of atrial fibrillation in the prior 5 years. Algorithm performance metrics were sensitivity, specificity, positive predictive value, negative predictive value (NPV) and number needed to screen (NNS). Subgroup analysis of patients aged ≥65 years was also performed.

RESULTS

Of 2,542,732 patients in DISCOVER, the algorithm identified 604,135 patients suitable for risk assessment. Of these, 3.0% (17,880 patients) had a diagnosis of atrial fibrillation recorded before study end. The area under the curve of the receiver operating characteristic was 0.87, compared with 0.83 in algorithm development. The NNS was nine patients, matching the CPRD cohort. In patients aged ≥30 years, the algorithm correctly identified 99.1% of patients who did not have atrial fibrillation (NPV) and 75.0% of true atrial fibrillation cases (sensitivity). Among patients aged ≥65 years (n = 117,965), the NPV was 96.7% with 91.8% sensitivity.

CONCLUSIONS

This atrial fibrillation risk prediction algorithm, based on machine learning methods, identified patients at highest risk of atrial fibrillation. It performed comparably in a large, real-world population-based cohort and the developmental registry cohort. If implemented in primary care, the algorithm could be an effective tool for narrowing the population who would benefit from atrial fibrillation screening in the United Kingdom.

Detection of Atrial Fibrillation Using a Ring-Type Wearable Device (CardioTracker) and Deep Learning Analysis of Photoplethysmography Signals: Prospective Observational Proof-of-Concept Study

Background

Continuous photoplethysmography (PPG) monitoring with a wearable device may aid the early detection of atrial fibrillation (AF).

Objective

We aimed to evaluate the diagnostic performance of a ring-type wearable device (CardioTracker, CART), which can detect AF using deep learning analysis of PPG signals.

Methods

Patients with persistent AF who underwent cardioversion were recruited prospectively. We recorded PPG signals at the finger with CART and a conventional pulse oximeter before and after cardioversion over a period of 15 min (each instrument). Cardiologists validated the PPG rhythms with simultaneous single-lead electrocardiography. The PPG data were transmitted to a smartphone wirelessly and analyzed with a deep learning algorithm. We also validated the deep learning algorithm in 20 healthy subjects with sinus rhythm (SR).

Results

In 100 study participants, CART generated a total of 13,038 30-s PPG samples (5850 for SR and 7188 for AF). Using the deep learning algorithm, the diagnostic accuracy, sensitivity, specificity, positive-predictive value, and negative-predictive value were 96.9%, 99.0%, 94.3%, 95.6%, and 98.7%, respectively. Although the diagnostic accuracy decreased with shorter sample lengths, the accuracy was maintained at 94.7% with 10-s measurements. For SR, the specificity decreased with higher variability of peak-to-peak intervals. However, for AF, CART maintained consistent sensitivity regardless of variability. Pulse rates had a lower impact on sensitivity than on specificity. The performance of CART was comparable to that of the conventional device when using a proper threshold. External validation showed that 94.99% (16,529/17,400) of the PPG samples from the control group were correctly identified with SR.

Conclusions

A ring-type wearable device with deep learning analysis of PPG signals could accurately diagnose AF without relying on electrocardiography. With this device, continuous monitoring for AF may be promising in high-risk populations.

Trial Registration

ClinicalTrials.gov NCT04023188; https://clinicaltrials.gov/ct2/show/NCT04023188

An App for Detecting Bullying of Nurses Using Convolutional Neural Networks and Web-Based Computerized Adaptive Testing: Development and Usability Study

Background

Workplace bullying has been measured in many studies to investigate its effects on mental health issues. However, none have used web-based computerized adaptive testing (CAT) with bully classifications and convolutional neural networks (CNN) for reporting the extent of individual bullying in the workplace.

Objective

This study aims to build a model using CNN to develop an app for automatic detection and classification of nurse bullying-levels, incorporated with online Rasch computerized adaptive testing, to help assess nurse bullying at an earlier stage.

Methods

We recruited 960 nurses working in a Taiwan Ch-Mei hospital group to fill out the 22-item Negative Acts Questionnaire-Revised (NAQ-R) in August 2012. The k-mean and the CNN were used as unsupervised and supervised learnings, respectively, for: (1) dividing nurses into three classes (n=918, 29, and 13 with suspicious mild, moderate, and severe extent of being bullied, respectively); and (2) building a bully prediction model to estimate 69 different parameters. Finally, data were separated into training and testing sets in a proportion of 70:30, where the former was used to predict the latter. We calculated the sensitivity, specificity, and receiver operating characteristic curve (area under the curve [AUC]), along with the accuracy across studies for comparison. An app predicting the respondent bullying-level was developed, involving the model’s 69 estimated parameters and the online Rasch CAT module as a website assessment.

Results

We observed that: (1) the 22-item model yields higher accuracy rates for three categories, with an accuracy of 94% for the total 960 cases, and accuracies of 99% (AUC 0.99; 95% CI 0.99-1.00) and 83% (AUC 0.94; 95% CI 0.82-0.99) for the lower and upper groups (cutoff points at 49 and 66 points) based on the 947 cases and 42 cases, respectively; and (2) the 700-case training set, with 95% accuracy, predicts the 260-case testing set reaching an accuracy of 97. Thus, a NAQ-R app for nurses that predicts bullying-level was successfully developed and demonstrated in this study.

Conclusions

The 22-item CNN model, combined with the Rasch online CAT, is recommended for improving the accuracy of the nurse NAQ-R assessment. An app developed for helping nurses self-assess workplace bullying at an early stage is required for application in the future.

An App Developed for Detecting Nurse Burnouts Using the Convolutional Neural Networks in Microsoft Excel: Population-Based Questionnaire Study

Background

Burnout (BO), a critical syndrome particularly for nurses in health care settings, substantially affects their physical and psychological status, the institute’s well-being, and indirectly, patient outcomes. However, objectively classifying BO levels has not been defined and noticed in the literature.

Objective

The aim of this study is to build a model using the convolutional neural network (CNN) to develop an app for automatic detection and classification of nurse BO using the Maslach Burnout Inventory–Human Services Survey (MBI-HSS) to help assess nurse BO at an earlier stage.

Methods

We recruited 1002 nurses working in a medical center in Taiwan to complete the Chinese version of the 20-item MBI-HSS in August 2016. The k-mean and CNN were used as unsupervised and supervised learnings for dividing nurses into two classes (n=531 and n=471 of suspicious BO+ and BO−, respectively) and building a BO predictive model to estimate 38 parameters. Data were separated into training and testing sets in a proportion 70%:30%, and the former was used to predict the latter. We calculated the sensitivity, specificity, and receiver operating characteristic curve (area under the curve) across studies for comparison. An app predicting respondent BO was developed involving the model’s 38 estimated parameters for a website assessment.

Results

We observed that (1) the 20-item model yields a higher accuracy rate (0.95) with an area under the curve of 0.97 (95% CI 0.94-0.95) based on the 1002 cases, (2) the scheme named matching personal response to adapt for the correct classification in model drives the prior model’s predictive accuracy at 100%, (3) the 700-case training set with 0.96 accuracy predicts the 302-case testing set reaching an accuracy of 0.91, and (4) an available MBI-HSS app for nurses predicting BO was successfully developed and demonstrated in this study.

Conclusions

The 20-item model with the 38 parameters estimated by using CNN for improving the accuracy of nurse BO has been particularly demonstrated in Excel (Microsoft Corp). An app developed for helping nurses to self-assess job BO at an early stage is required for application in the future.

A Lightweight Deep Learning Model for Fast Electrocardiographic Beats Classification With a Wearable Cardiac Monitor: Development and Validation Study

Background

Electrocardiographic (ECG) monitors have been widely used for diagnosing cardiac arrhythmias for decades. However, accurate analysis of ECG signals is difficult and time-consuming work because large amounts of beats need to be inspected. In order to enhance ECG beat classification, machine learning and deep learning methods have been studied. However, existing studies have limitations in model rigidity, model complexity, and inference speed.

Objective

To classify ECG beats effectively and efficiently, we propose a baseline model with recurrent neural networks (RNNs). Furthermore, we also propose a lightweight model with fused RNN for speeding up the prediction time on central processing units (CPUs).

Methods

We used 48 ECGs from the MIT-BIH (Massachusetts Institute of Technology-Beth Israel Hospital) Arrhythmia Database, and 76 ECGs were collected with S-Patch devices developed by Samsung SDS. We developed both baseline and lightweight models on the MXNet framework. We trained both models on graphics processing units and measured both models’ inference times on CPUs.

Results

Our models achieved overall beat classification accuracies of 99.72% for the baseline model with RNN and 99.80% for the lightweight model with fused RNN. Moreover, our lightweight model reduced the inference time on CPUs without any loss of accuracy. The inference time for the lightweight model for 24-hour ECGs was 3 minutes, which is 5 times faster than the baseline model.

Conclusions

Both our baseline and lightweight models achieved cardiologist-level accuracies. Furthermore, our lightweight model is competitive on CPU-based wearable hardware.

Photoplethysmography based atrial fibrillation detection: a review

Atrial fibrillation (AF) is a cardiac rhythm disorder associated with increased morbidity and mortality. It is the leading risk factor for cardioembolic stroke and its early detection is crucial in both primary and secondary stroke prevention. Continuous monitoring of cardiac rhythm is today possible thanks to consumer-grade wearable devices, enabling transformative diagnostic and patient management tools. Such monitoring is possible using low-cost easy-to-implement optical sensors that today equip the majority of wearables. These sensors record blood volume variations-a technology known as photoplethysmography (PPG)-from which the heart rate and other physiological parameters can be extracted to inform about user activity, fitness, sleep, and health. Recently, new wearable devices were introduced as being capable of AF detection, evidenced by large prospective trials in some cases. Such devices would allow for early screening of AF and initiation of therapy to prevent stroke. This review is a summary of a body of work on AF detection using PPG. A thorough account of the signal processing, machine learning, and deep learning approaches used in these studies is presented, followed by a discussion of their limitations and challenges towards clinical applications.