a personalized monitoring model for electrocardiogram …

JMIR Preprints Kotorov et al

A Personalized Monitoring Model forElectrocardiogram (ECG) Signals: Diagnostic Accuracy

Study

Rado Kotorov, Lianhua Chi, Min Shen

Submitted to: JMIR Biomedical Engineeringon: September 17, 2020

Disclaimer: © The authors. All rights reserved. This is a privileged document currently under peer-review/communityreview. Authors have provided JMIR Publications with an exclusive license to publish this preprint on it's website forreview purposes only. While the final peer-reviewed paper may be licensed under a CC BY license on publication, at thisstage authors and publisher expressively prohibit redistribution of this draft paper other than for review purposes.

https://preprints.jmir.org/preprint/24388 [unpublished, peer-reviewed preprint]


Table of Contents

Original Manuscript ....................................................................................................................................................................... 5Supplementary Files ..................................................................................................................................................................... 17

Figures ......................................................................................................................................................................................... 18Figure 2 ...................................................................................................................................................................................... 19Figure 3 ...................................................................................................................................................................................... 20Figure 14 .................................................................................................................................................................................... 21Figure 15 .................................................................................................................................................................................... 22Figure 16 .................................................................................................................................................................................... 23Figure 13 .................................................................................................................................................................................... 24Figure 1 ...................................................................................................................................................................................... 25Figure 4 ...................................................................................................................................................................................... 26Figure 5 ...................................................................................................................................................................................... 27Figure 6 ...................................................................................................................................................................................... 28Figure 7 ...................................................................................................................................................................................... 29Figure 8 ...................................................................................................................................................................................... 30Figure 9 ...................................................................................................................................................................................... 31Figure 10 .................................................................................................................................................................................... 32Figure 11 .................................................................................................................................................................................... 33Figure 12 .................................................................................................................................................................................... 34



A Personalized Monitoring Model for Electrocardiogram (ECG) Signals:Diagnostic Accuracy Study

Rado Kotorov1 PhD; Lianhua Chi2 PhD; Min Shen1 MSc

1Trendalyze Inc. 1 Washington Park Newak US2La Trobe University Bundoora AU

Corresponding Author:Lianhua Chi PhDLa Trobe UniversityThomas Cherry Bldg, 3rd FlLa Trobe UniversityBundooraAU

Abstract

Background: Lately, the demand for remote ECG monitoring has increased drastically because of the COVID-19 pandemic. Toprevent the spread of the virus and keep individuals with less severe cases out of hospitals, more patients are having heart diseasediagnosis and monitoring remotely at home. The efficiency and accuracy of the ECG signal classifier are becoming moreimportant because false alarms can overwhelm the system. Therefore, how to classify the ECG signals accurately and send alertsto healthcare professionals in a timely fashion is an urgent problem to be addressed.

Objective: The primary aim of this research is to create a robust and easy-to-configure solution for monitoring ECG signal inreal-world settings. We developed a technique for building personalized prediction models to address the issues of generalizedmodels because of the uniqueness of heartbeats [19]. In most cases, doctors and nurses do not have data science background andthe existing Machine Learning models might be hard to configure. Hence a new technique is required if Remote PatientMonitoring will take off on a grand scale as is needed due to COVID-19. The main goal is to develop a technique that allowsdoctors, nurses, and other medical practitioners to easily configure a personalized model for remote patient monitoring. Theproposed model can be easily understood and configured by medical practitioners since it requires less training data and fewerparameters to configure.

Methods: In this paper, we propose a Personalized Monitoring Model (PMM) for ECG signal based on time series motifdiscovery to address this challenge. The main strategy here is to individually extract personalized motifs for each individualpatient and then use motifs to predict the rest of readings of that patient by an artificial logical network.

Results: In 32 study patients, each patient contains 30 mins of ECG signals/readings. Using our proposed PersonalizedMonitoring Model (PMM), the best diagnostic accuracy reached 100%. Overall, the average accuracy of PMM was alwaysmaintained above 90% with different parameter settings. For Generalized Monitoring Models (GMM1 and GMM2), the averageaccuracies were only around 80% with much more running time than PMM. Regardless of parameter settings, it normally took3-4 mins for PMM to generate the training model. However, for GMM1 and GMM2, it took around 1 hour and even more withthe increase of training data. The proposed model substantially speeds up the ECG diagnostics and effectively improve theaccuracy of ECG diagnostics.

Conclusions: Our proposed PMM almost eliminates much training and small sample issues and is completely understandableand configurable by a doctor or a nurse.

(JMIR Preprints 17/09/2020:24388)DOI: https://doi.org/10.2196/preprints.24388

Preprint Settings

1) Would you like to publish your submitted manuscript as preprint?Please make my preprint PDF available to anyone at any time (recommended).



Please make my preprint PDF available only to logged-in users; I understand that my title and abstract will remain visible to all users.Only make the preprint title and abstract visible.No, I do not wish to publish my submitted manuscript as a preprint.

2) If accepted for publication in a JMIR journal, would you like the PDF to be visible to the public?Yes, please make my accepted manuscript PDF available to anyone at any time (Recommended). Yes, but please make my accepted manuscript PDF available only to logged-in users; I understand that the title and abstract will remain visible to all users (see Important note, above). I also understand that if I later pay to participate in <a href="https://jmir.zendesk.com/hc/en-us/articles/360008899632-What-is-the-PubMed-Now-ahead-of-print-option-when-I-pay-the-APF-" target="_blank">JMIR’s PubMed Now! service</a> service, my accepted manuscript PDF will automatically be made openly available.Yes, but only make the title and abstract visible (see Important note, above). I understand that if I later pay to participate in <a href="https://jmir.zendesk.com/hc/en-us/articles/360008899632-What-is-the-PubMed-Now-ahead-of-print-option-when-I-pay-the-APF-" target="_blank">JMIR’s PubMed Now! service</a> service, my accepted manuscript PDF will automatically be made openly available.



Original Manuscript



Original Paper

A Personalized Monitoring Model for Electrocardiogram (ECG)Signals: Diagnostic Accuracy Study

Abstract

Background: Due to the COVID-19, pandemic demand for remote ECG monitoring has increaseddrastically to prevent the spread of the virus and keep vulnerable individuals with less severe casesout of hospitals. How to set up remote patient ECG monitoring easily by clinicians and how toclassify the ECG signals accurately to send only relevant alerts in timely fashion is an urgentproblem to be addressed in order for Remote Patient Monitoring (RPM) to be adopted on a widescale. Hence a new technique is required if RPM will take off on a grand scale as is needed due toCOVID-19.Objective: The primary aim of this research is to create a robust and easy-to-use solution forpersonalized ECG monitoring in real-world settings that is precise, easily configurable, andunderstandable by clinicians. Methods: In this paper, we propose a Personalized Monitoring Model (PMM) for ECG data basedon motif discovery. Motif discovery finds meaningful or frequently recurring patterns in patient ECGreadings. The main strategy is to use motif discovery to extract a small sample of personalized motifsfor each individual patient and then use these motifs to predict abnormalities during the real timereadings of that patient using an artificial logical network configured by a physician.Results: Our approach was tested on 30 minutes ECG readings from 32 patients. The averagediagnostic accuracy of PMM was always above 90% and under some parameters 100%, compared to80% accuracy for Generalized Monitoring Model (GMM). Regardless of parameter settings, PMMtraining models were generated within 3-4 mins, compared to 1 hour and even more with theincrease of training data for GMM.Conclusions: Our proposed PMM almost eliminates much training and small sample issues andaddresses accuracy and computational cost issues of the Generalized Monitoring Models (GMM)caused by the uniqueness of heartbeats and training issues. It also addresses the fact that doctors andnurses do not have data science training and skills to configure, understand and even trust existingblack-box Machine Learning models.

Keywords: COVID-19; personalized monitoring model; ECG; time series; motif discovery;

Introduction

Background

An electrocardiogram (ECG) is a medical test that records the electrical activities of heartbeats. It iswidely used by medical practitioners for diagnosing cardiac conditions by detecting irregular heartrhythm and abnormalities [1]. In some cases, the arrhythmic heartbeats can be lethal and the risks ofsudden death without Remote Patient Monitoring (RPM) are significant [2]. Therefore, it is highlydesirable for patients to have an efficient ECG remote monitoring system which can identify life-threatening situations and send alerts to their healthcare providers [3]. Lately, the demand for remoteECG monitoring has increased drastically because of the COVID-19 pandemic. To prevent thespread of the virus and keep individuals with less severe cases out of hospitals, more patients arehaving heart disease diagnosis and monitoring remotely at home. The accuracy of the ECG signalclassifier is becoming more important because false alarms can overwhelm the system. Therefore,



how to classify the ECG signals accurately and send alerts to healthcare professionals in a timelyfashion is an urgent problem to be addressed.

The classification of ECG signals is an extremely challenging problem since there is no definedoptimal classification rules. Many researchers focused on developing machine learning models suchas Bayesian framework [4], Random Forest [5], Gradient Boosting [6], Ensemble Boosting [36], andSupport Vector Machine [7], etc. and achieved relatively high accuracy. To extract different featuresfor the model, various techniques were proposed such as Principal Component Analysis (PCA) [12-15], Wavelet Transform (WT) [16], and filter banks [17] etc. Deep learning methods such as deepneural networks [29], convolutional neural networks [8-11, 25] and recurrent neural networks [26]are also applied extensively for classification problems. Deep learning can be a powerful tool tosolve cognitive problems [24], but generally it requires massive amounts of labeled data toaccurately train the model [27]. For clinical applications, due to limited patient contact, variation inmedical care and privacy issues, getting a massive amount of high-quality data can be verychallenging [18], and the efficacy of deep learning methods can be greatly affected by the lack oftraining data [31-35]. According to Chen et al.’s investigation of the training time of deep learningand machine learning methods [28], deep learning method has the longest training time comparing toconventional machine learning algorithms. Also, building and maintaining the computationalinfrastructure for deep learning can be too costly to small healthcare organizations to implement.Thus, a less computationally expensive method is needed to effectively resolve the issue.

Another limitation for deep learning methods is that the models are not able to capture theindividuality of the ECG features and patterns [19]. Most of the deep learning models are generalizedmodels and are not able to be built in individual levels due to data sparsity [32]. However, eachpatient has unique heartbeats and the waveforms can be totally different in individual levels. Hencethe accuracy might be an issue for these models when using real-time data. For traditional machinelearning methods, they usually require more efforts in data preprocessing and feature engineeringcomparing to deep learning models [35] and tend to be black boxes to medical practitioners withoutdata science background. The lack of interpretability can hinder the process of decision making andcommunications with patients for healthcare providers.

In this paper, we propose a Personalized Monitoring Model (PMM) for ECGmonitoring based on motif discovery to address the challenges. Motif discovery isa method for analyzing massive time series data. In healthcare domain, it hasbeen used for trend analysis and data summarization [20]. Motifs are defined asfrequently recurrent patterns in certain time series [21]. In a motif discoveryprocess, similarity search is conducted based on certain similarity threshold todetect and locate previously defined patterns. In a similarity search, the distancesbetween time series subsequences are calculated, which indicates how similartwo subsequences are.

Objectives

The primary aim of this research is to create a robust and easy-to-configure solution for monitoringECG signal in real-world settings. We developed a technique for building personalized predictionmodels to address the issues of generalized models because of the uniqueness of heartbeats [19]. Themain strategy of the model is to extract personalized motifs for each patient and use the motifs topredict the rest of the readings of that patient by artificial logical network. By performing systematicanalysis and evaluation, we investigate the hypothesis that the proposed Personalized Monitoring



Model (PMM) is more accurate and efficient than generalized models. In most cases, doctors andnurses do not have data science background and the existing Machine Learning models might behard to configure. Hence a new technique is required if Remote Patient Monitoring will take off on agrand scale as is needed due to COVID-19. The main goal is to develop a technique that allowsdoctors, nurses, and other medical practitioners to easily configure a personalized model for remotepatient monitoring. The proposed model can be easily understood and configured by medicalpractitioners, since it requires less training data and fewer parameters to configure.

Methods

In this section, we discuss the proposed Personalized Monitoring Model (PMM) for ECG data indetail. The whole process includes Time Series Sampling, Personalized Motif Discovery and Motifbased Prediction using Artificial Logical Network.

Time Series Sampling

We take each patient's electrocardiogram (ECG) measures as the individual research data, which is atime series. ECG is a medical test that detects heart problems by measuring the electrical activitygenerated by the heart as it contracts. An ECG complex is comprised of different components, orwaves, that represent the electrical activity in specific regions of the heart. ECG from healthy heartshave a characteristic shape. If the ECG shows a different shape, it could suggest a heart problem. During this stage, we have a very important parameter which represents the training ratio (0< <1)

that affects the sampling process. Before we start to discover the motifs, we need to sample eachindividual patient's ECG time series as an individual training data. Let's suppose the length of eachpatient's ECG time series is . Based on the value of training ratio , we use the equation (1) to

calculate the length of time series we should take from the whole ECG readings of each patient:

(1)

As calculated, we take the first length of time series from the ECG readings of each individual

patient as the individual training sample data to generate the personalized motifs. During sampling,we start from the first point of the patient's ECG time series and stop sampling until the length ofsample reaches the expected sample size S. Then, in this sampled ECG data, we divide it into Msubsequences based on equation (2) and each subsequence is regarded as a pattern unit. (2)

In the equation (2), the number 180 is the sampling rate of the ECG recordingdevice. It partitions the ECG into heartbeats with sufficient precision of intervalsfor HRV analysis [23] As in the Figure 1, it is an ECG sample with 1800 pointsfrom one of the patients.

Personalized Motif Discovery

After sampling, we discover personalized motifs from the subsequences sampled from an

individual patient. In this model design, we need to consider another two major parameters that couldaffect the performance of motif discovery: : time series similarity threshold

: number of motifs.



We calculate all Euclidean distances between each subsequence and generate motif candidates. Ineach motif circle/cluster, the distances from the central subsequence to other subsequences thatbelongs to the same motif circle must be less than and all motif circles cannot share the same

subsequence as shown in the Figure 1 as proposed by the paper [22]. In Figure 2, each black pointrepresents each subsequence and each red point represents the central subsequence in that motifcircle. We suppose = 3, as we can see 1-motif has the most subsequences, 2-motif has the second

most and 3-motif has the least subsequences. All motif circles have the same radius . The distance

between the red central subsequence of 1-motif and the red central subsequence of 2-motif is morethan , which naturally do not share the same subsequence. However, the distance between the

central subsequence of 1-motif and the central subsequence of 3-motif is less than but they do not

share the same subsequence which is allowable during the motif discovery. And in this example, wecan have the three red central subsequences as the extracted 3 motifs.

Based on this strategy of motif discovery, we can generate all motif circle candidates out of the

subsequences. Based on the parameter which is the number of motifs, we only keep the first

motif circles and use the central subsequences of all these motif circles as our extracted heartbeat

motifs. From this stage, we can get personalized motifs for each individual patient.

Motif Based Prediction Using Artificial Logical Network

In this section, we use the generated k personalized motifs to predict the rest of ECG readings ofeach corresponding patient. Suppose we have two types of heartbeats (N and V) and need to predictwhich type (N or V) the test subsequence belongs to. Following the personalized motif discovery, wecan generate k of N motifs and k of V motifs from the sampled subsequences of each individualpatient. The generated N and V motifs are organized into an artificial logical network where each Nand V motif is a dedicated evaluation node as shown in Figure 3. Then for the rest of subsequencesof that patient, we can test each subsequence by comparing its distance to all N motifs and thedistance to all V motifs in the artificial logical network. A simple logical rule is applied to select theclosest one as the predicted type of test subsequence. Then finally we can predict all labels for therest of subsequences of that patient. If the subsequences do not meet the matching criteria of any ofN and V nodes, the logical rule identifies them as new anomalies for future learning. Thecombination of dedicated motif comparison and logical rules allows us to easily build a predictionsystem.

Results

Benchmark Data

In this benchmark data, we use 32 patients' ECG measures, and each measure contains 30 mins ofECG readings of each patient. In the dataset, we have 5 different types of heartbeats (V, N, A, F andS). The V pathology is expected to be morphological different than normal N. For detecting A, weneed to monitor the frequency of the heartbeats and catch the heartbeats that appear faster than whatis expected. The F pathology also is expected to have different morphology than the normal N. The Spathology is related to heart rhythm abnormalities that may not drastically change the morphology,but its occurrence is going to be completely out of rhythm. Before training and testing the models,we remove the noisy heartbeats and only keep N heartbeats and V heartbeats as two labels for themodels to identify.



Baseline Models

In this section, we evaluate and compare three following models: Generalized Monitoring Model 1 (GMM1): Based on the training ratio , we take the first

percent samples from each of 32 patients and combine all those samples from 32 patientstogether to extract of N motifs and of V motifs as the V heartbeat motifs and N heartbeat

motif separately. During the testing stage, we apply extracted N motifs and V motifs to the rest ofECG readings for each patient and predict the label (N or V).

Generalized Monitoring Model 2 (GMM2): We first extract all N heartbeats and V heartbeatsfrom all of 32 patients. Then, based on the training ratio , we randomly extract percent samples

from all N heartbeats and percent samples from all V heartbeats. Next, we extract the of N

motifs from N heartbeat samples and of V motifs from V heartbeat samples. The testing is the

same as in GMM1. The main difference between GMM1 and GMM2 is how to sample thetraining data.

Our proposed Personalized Monitoring Model (PMM): In this personalized model, based on thetraining ratio , we only extract the first percent of N heartbeats and the first percent of V

heartbeats from an individual patient and then generate a set of personalized of N motifs and

of V motifs to test the rest of ECG readings of this patient. The main strategy here is toindividually extract personalized motifs for the current patient and use those extracted motifs topredict the rest of readings of this patient.

We have three major parameters that could affect the models' performance and will be compared indetail in this section: : time series similarity threshold

: number of motifs

: training ratio.

To compare the models and understand which model is the best one, in this paper, we evaluate basedon the below two factors: Accuracy: we need to get an estimate for how accurate each model is on unseen/test data. The

accuracy is normally calculated by Equation (3). In all tested heartbeats, we have thecorresponding ground truth information which is the original labels. By comparing the predictedlabel with the original label, we can calculate how many heartbeats are correctly predicted.

(3)

Running Time: In order to get the final prediction results faster, we also need to guarantee thechosen model meets the least time consumption, including training and testing time.

Effectiveness Evaluation

Considering the three main parameters , and , we design the effectiveness evaluation by

adjusting the values of these three parameters and see how does the performance of each modelchange. In the result figures, we use the corresponding capital letters , and of , and for

clear representation.



Firstly, we adjust the similarity threshold values from 0.8 to 1.6 increased by 0.2 and see how does

the average performance change on each model. The average accuracy is calculated based on thesum accuracy on all 32 patients. In the Figure 4, it shows three curves, each representing the averageaccuracy change of each model when is increasing from 0.8 to 1.6. The green line represents PMM

model which performs the best among these three models in terms of stability and accuracy.

Secondly, we adjust the values from 2 to 10 increased by 2 and see how does the average

performance change on each model. In the Figure 5, we can see PMM model still always performsthe best even though the varies a lot from 2 to 10.

Lastly and also very importantly, we adjust the training ratio from 0.1 to 0.25 increased by 0.05.

Generally, more training samples can result in more accurate prediction. However, from the Figure 6,this rule does not apply to GMM1 and GMM2, but does apply to PMM. From the curve comparison,we can see PMM significantly outperforms the GMM1 and GMM2.

Considering is the training ratio that could affect the performance even more than and in most

of cases. Hence, we list the detailed performance result on each patient of these three models basedon two sets of values of and : (1) R=0.8 and K= 2; (2) R=1 and K= 6;

From Figure 7 and Figure 8, we can see GMM2 fluctuates much more than GMM1 and PMM.However, if we compare GMM1 and PMM when is changing from 0.1 to 0.25, PMM gradually

performs better on almost all patients, but GMM1 is even getting worse somehow. From Figure 9and Figure 10, the two bar charts show the average accuracy of each model with different . As the

increases, GMM1's performance becomes even worse and GMM2's performance fluctuate the most.Only PMM always performs the best.

Efficiency Evaluation

In this section, we will evaluate the time efficiency of each model and see which model is runningfastest. During the whole process, we have two stages: training and testing. We will consider allcomputation time including the training and testing time in this evaluation. From the evaluated threeparameters in the previous, we can know the training ratio is the one that most affects the training

time. So in the section we mainly adjust the value of and see the corresponding running time of

each model.

From Figure 11, we can see a linear change of running time on each model. And Figure 12 is a barchart of the running time. From these two figures, we can clearly see the time consumption ofGMM1 and GMM2 increases almost exponentially with the increase of . However, the time

consumption of PMM is linear and the least among these three models.



Discussion

Principal Results

According to the empirical results in results section, the Personalized Monitoring Model (PMM)always performs the best in all cases in terms of predicting accuracy and time efficiency.

As in Figure 4, it shows the average classification accuracy (y-axis) with respect to the (x-axis) on

all three models by using different similarity threshold . From the green line, we can see PMM's

performance is getting better as the increases from 0.8 to 1.4 and then get stable when reaches

1.6. This green curve shows the prediction performance of PMM is relatively stable for similaritythreshold . If we compare GMM1 and GMM2, we can find GMM1 outperforms GMM2. However,

from the blue curve, we can see GMM1's performance is quite unstable with respect to , and from

the orange curve, GMM2's performance is getting better as increases even though the overall

performance of GMM2 is still worse than GMM1. The reason is: GMM2 extracts all N heartbeatsand V heartbeats from all of 32 patients for motif discovery and a bigger similarity threshold canallow GMM2 to aggregate more similar heartbeats for motif circles, which could help GMM2 findmore representative motifs for heartbeat prediction. However, if the similarity threshold is too big,

it may introduce different type of heartbeats into the motif circle, which may mislead prediction atthe end. For GMM1, it takes the first percent of samples from 32 patients and then extract motifs

based on the similarity distance threshold . The increase of may introduce more heartbeats from

different patients and the extracted motifs may result in a fluctuation in accuracy. However, forPMM, it only extracts the first percent of N samples and V samples from one patient and then

generate personalized motifs to test the rest of ECG readings of this patient. As the increase, more

similar N heartbeats and V heartbeats will be collected for motif discovery and that could helpenhance the prediction if is not too bigger to introduce another type of heartbeats.

As in the Figure 5, it shows the average classification accuracy (y-axis) with respect to the (x-axis)

on all three models by using different number of motifs .

We can see the green line is always on the top, but the blue and orange curves fluctuate much moreseriously as the changes. As introduced in the previous section, we know represents the number

of motifs. A certain number of motifs can be representative and guide prediction well. However, toomany motifs may introduce unrepresentative motifs, which could mislead the prediction task. This isthe reason why all models fluctuate as the changes.

As in the Figure 6, it shows the average classification accuracy (y-axis) with respect to the (x-axis)

on all three models by using different training ratios from 0.1 to 0.25. As the increases from 0.1

to 0.25 by 0.05, we can see the green curve is becoming higher and higher. However, the blue andorange curves fluctuate a lot and both reach to the lowest when the is 0.2. Generally, in machine

learning, more training data could help improve model's performance. However, if more interference/noisy data is introduced into the training data, then the performance will degrade. That could explainwhy GMM1 and GMM2 fluctuate a lot due to the designing nature of these two models (training



data from different patients and most heartbeats are unique on a patient level). As for PMM, a biggertraining ratio could help improve the performance because all training heartbeats are from the samepatient, which naturally avoids the introduction of interference/noisy data. Figures from 7 to 10further validate that PMM always outperform the GMM1 and GMM2 as the changes.

Based on the overall evaluation results, we can conclude that PMM can significantly outperformGMM1 and GMM2 in terms of the prediction accuracy.

Considering the time efficiency, as in the Figures 11 and 12, they show the average running time (y-axis) with respect to the training ratio (x-axis) on all three models by using different training ratios

from 0.1 to 0.25. Compared with GMM1 and GMM2 models, PMM requires significant less

running time. The average running time of the three models increases as the training ratio

increases. This is because a larger training ratio would result in more training data, which increasethe training time accordingly. For the two benchmark models, the overall average time of GMM1 isclose to GMM2 and both increase almost exponentially. Therefore, we can summarize that PMM hasmore stable and better efficiency as the training ratio increases.

Limitations

The limitation of the personalized model is that it might be hard to maintain the models for eachpatient. But this limitation is not significant given that retaining the model is fast. Also, the cost ofmaintenance and computation might be lower in the future as more industries are adoptingpersonalized models.

Conclusions

In this paper, we proposed a Personalized Monitoring Model (PMM) for ECG measures for tworeasons: (1) Covid-19 has accelerated the adoption of remote diagnosis and patient monitoring, and(2) personalized care promises better outcomes in general and in specific as it applies to digitalhealth. Digital healthcare allows for continuous 24/7 care in the home environment while minimizingthe risks of fatal accidents and re-admissions. For remote monitoring to gain traction at scale, severalrequirements must be met. First, the monitoring has to be sufficiently automated with fewer falsepositive alarms in order to minimize the number of health professionals involved in it. Second, it hasto be quickly and easily configurable by the healthcare professionals themselves. While traditionalmachine and deep learning approaches can be used in automation, they cannot be easily configurednor adjusted by the healthcare professionals themselves due to lack of modeling skills andinsufficient data to build a personalized model. To solve these challenges, we employed motifdiscovery algorithm to individually extract personalized motifs for each individual patient andcombined artificial logical network for the ECG signal prediction. We proposed a personalizedmodel for a faster ECG signal detection, which significantly improves the efficiency of ECGprediction to satisfy the dramatic demand especially in current COVID-19 situation. As a result, bycomparing with two generalized monitoring models, experiments on a real-world patient ECG dataset demonstrate that our proposed model PMM outperforms the generalized models in bothprediction accuracy and time efficiency.

Per our discussions with clinicians, this approach can easily be deployed for outpatient monitoring inthe following way which is the subject of a forthcoming clinical trial. A wearable 12-channel ECGmonitor is sent to a patient or configured during a hospital stay. Augmented reality app or videoconference is used to remotely guide the patient to accurately place the electrodes, whilesimultaneously testing the accuracy of the received signal. During the setup, personalized motifs are



automatically extracted, and the physician selects the center motifs to be used in the artificial logicalnetwork. The artificial logical network is a flexible structure that allows for learning when particularreference motifs are missing from the setup sample. For example, if samples of atrial fibrillationmotifs are not recorded during the setup, reference motifs from a general library can be used, or ageneral anomaly detection can be applied alerting a medical professional to review the anomalousoccurrences. If the physician determines that the anomaly is atrial fibrillation, she can instantly pushthe motif to the artificial logical network. These configurations are the object of the forthcomingclinical trial. By augmenting the expert’s knowledge with algorithmic computational power, hospitalstays can be significantly reduced, and care can be delivered in the comfort of the home.

Acknowledgements

We would like to thank Dr. Mazomenos from the University of Leeds for helping us prepare the dataset used in this research. He has annotated the heartbeats and pathologies and developed the 180-records heartbeat methodology.

Conflicts of Interest

None declared.

Abbreviations

ECG: ElectrocardiogramGMM: Generalized Monitoring ModelPCA: Principal Component AnalysisPMM: Personalized Monitoring ModelRPM: Remote Patient MonitoringWT: Wavelet Transform

References

1. Mann H. A method of analyzing the electrocardiogram. Archives of Internal Medicine. 1920Mar 1;25(3):283-94.

2. Nattel S. Sudden cardio arrest: when normal ECG variants turn lethal. Nature medicine. 2010Jun;16(6):646-7.

3. Guo SL, Han LN, Liu HW, Si QJ, Kong DF, Guo FS. The future of remote ECG monitoringsystems. Journal of geriatric cardiology: JGC. 2016 Sep;13(6):528.

4. Sayadi O, Shamsollahi MB, Clifford GD. Robust detection of premature ventricularcontractions using a wave-based Bayesian framework. IEEE Transactions on BiomedicalEngineering. 2009 Sep 15;57(2):353-62.

5. Kropf M, Hayn D, Schreier G. ECG classification based on time and frequency domainfeatures using random forests. In2017 Computing in Cardiology (CinC) 2017 Sep 24 (pp. 1-4). IEEE.

6. Alarsan FI, Younes M. Analysis and classification of heart diseases using heartbeat featuresand machine learning algorithms. Journal of Big Data. 2019 Dec;6(1):1-5.

7. Li Q, Rajagopalan C, Clifford GD. A machine learning approach to multi-level ECG signalquality classification. Computer methods and programs in biomedicine. 2014 Dec1;117(3):435-47.

8. Pyakillya B, Kazachenko N, Mikhailovsky N. Deep learning for ECG classification.InJournal of physics: conference series 2017 Oct (Vol. 913, No. 1).

9. Zhang J, Tian J, Cao Y, Yang Y, Xu X, Wen C. Fine-grained ECG classification based on deepCNN and online decision fusion. Comput. Res. Repos. 2019 Jan.

10. Jun TJ, Nguyen HM, Kang D, Kim D, Kim D, Kim YH. ECG arrhythmia classification using



a 2-D convolutional neural network. arXiv preprint arXiv:1804.06812. 2018 Apr 18.11. Rajpurkar P, Hannun AY, Haghpanahi M, Bourn C, Ng AY. Cardiologist-level arrhythmia

detection with convolutional neural networks. arXiv preprint arXiv:1707.01836. 2017 Jul 6.12. Castells F, Laguna P, Sörnmo L, Bollmann A, Roig JM. Principal component analysis in ECG

signal processing. EURASIP Journal on Advances in Signal Processing. 2007 Dec1;2007(1):074580.

13. Monasterio V, Laguna P, Martinez JP. Multilead analysis of T-wave alternans in the ECGusing principal component analysis. IEEE Transactions on Biomedical Engineering. 2009Mar 4;56(7):1880-90.

14. Martis RJ, Acharya UR, Mandana KM, Ray AK, Chakraborty C. Application of principalcomponent analysis to ECG signals for automated diagnosis of cardiac health. ExpertSystems with Applications. 2012 Oct 15;39(14):11792-800.

15. Kallas M, Francis C, Kanaan L, Merheb D, Honeine P, Amoud H. Multi-class SVMclassification combined with kernel PCA feature extraction of ECG signals. In 2012 19thInternational Conference on Telecommunications (ICT) 2012 Apr 23 (pp. 1-5). IEEE.

16. Kachuee M, Fazeli S, Sarrafzadeh M. Ecg heartbeat classification: A deep transferablerepresentation. In 2018 IEEE International Conference on Healthcare Informatics (ICHI)2018 Jun 4 (pp. 443-444). IEEE.

17. Afonso VX, Tompkins WJ, Nguyen TQ, Luo S. ECG beat detection using filter banks. IEEEtransactions on biomedical engineering. 1999 Feb;46(2):192-202.

18. Shyu LY, Wu YH, Hu W. Using wavelet transform and fuzzy neural network for VPCdetection from the Holter ECG. IEEE Transactions on Biomedical Engineering. 2004 Jun21;51(7):1269-73.

19. Jambukia SH, Dabhi VK, Prajapati HB. Classification of ECG signals using machinelearning techniques: A survey. In2015 International Conference on Advances in ComputerEngineering and Applications 2015 Mar 19 (pp. 714-721). IEEE.

20. Pathinarupothi RK, Rangan E. Discovering vital trends for personalized healthcare delivery.In Proceedings of the 2016 ACM International Joint Conference on Pervasive and UbiquitousComputing: Adjunct 2016 Sep 12 (pp. 1106-1109).

21. Mueen A, Chavoshi N. Enumeration of time series motifs of all lengths. Knowledge andInformation Systems. 2015 Oct 1;45(1):105-32.

22. Chi LH, Chi HH, Feng YC, Wang SL, Cao ZS. Comprehensive and efficient discovery oftime series motifs. Journal of Zhejiang University SCIENCE C. 2011 Dec 1;12(12):1000-9.

23. Kwon O, Jeong J, Kim HB, Kwon IH, Park SY, Kim JE, Choi Y. Electrocardiogram samplingfrequency range acceptable for heart rate variability analysis. Healthcare informaticsresearch. 2018 Jul 1;24(3):198-206.

24. Fawaz HI, Forestier G, Weber J, Idoumghar L, Muller PA. Deep learning for time seriesclassification: a review. Data Mining and Knowledge Discovery. 2019 Jul 1;33(4):917-63.

25. Tang W, Long G, Liu L, Zhou T, Jiang J, Blumenstein M. Rethinking 1D-CNN for TimeSeries Classification: A Stronger Baseline. arXiv preprint arXiv:2002.10061. 2020 Feb 24.

26. Chen W, Wang S, Long G, Yao L, Sheng QZ, Li X. Dynamic illness severity prediction viamulti-task rnns for intensive care unit. In2018 IEEE International Conference on Data Mining(ICDM) 2018 Nov 17 (pp. 917-922). IEEE.

27. Chartrand G, Cheng PM, Vorontsov E, Drozdzal M, Turcotte S, Pal CJ, Kadoury S, Tang A.Deep learning: a primer for radiologists. Radiographics. 2017 Nov;37(7):2113-31.

28. Chen D, Liu S, Kingsbury P, Sohn S, Storlie CB, Habermann EB, Naessens JM, Larson DW,Liu H. Deep learning and alternative learning strategies for retrospective real-world clinicaldata. NPJ digital medicine. 2019 May 30;2(1):1-5.

29. Ribeiro AH, Ribeiro MH, Paixão GM, Oliveira DM, Gomes PR, Canazart JA, Ferreira MP,Andersson CR, Macfarlane PW, Wagner Jr M, Schön TB. Automatic diagnosis of the 12-lead



ECG using a deep neural network. Nature communications. 2020 Apr 9;11(1):1-9.30. Montavon G, Samek W, Müller KR. Methods for interpreting and understanding deep neural

networks. Digital Signal Processing. 2018 Feb 1;73:1-5.31. Wang F, Casalino LP, Khullar D. Deep learning in medicine—promise, progress, and

challenges. JAMA internal medicine. 2019 Mar 1;179(3):293-4. 32. Xiao C, Choi E, Sun J. Opportunities and challenges in developing deep learning models

using electronic health records data: a systematic review. Journal of the American MedicalInformatics Association. 2018 Oct;25(10):1419-28.

33. Shameer K, Johnson KW, Glicksberg BS, Dudley JT, Sengupta PP. Machine learning incardiovascular medicine: are we there yet? Heart. 2018 Jul 1;104(14):1156-64.

34. Wiens J, Shenoy ES. Machine learning for healthcare: on the verge of a major shift inhealthcare epidemiology. Clinical Infectious Diseases. 2018 Jan 1;66(1):149-53.

35. Miotto R, Wang F, Wang S, Jiang X, Dudley JT. Deep learning for healthcare: review,opportunities and challenges. Briefings in bioinformatics. 2018 Nov;19(6):1236-46.

36. Wang H, Wu J. Boosting for Real-Time Multivariate Time Series Classification. In AAAI2017 Feb 4 (pp. 4999-5000).



Supplementary Files



Figures



Subsequence motifs (k=3).



Motif based Prediction using Artificial Logical Network.



R=1 and K=6: Average performance comparison with different T.



Linear comparison: Time efficiency.



Bar comparison: Time efficiency.



R=0.8 and K=2: Average performance comparison with different T.



ECG sample.



Performance comparison with different R.



Performance comparison with different K.



Performance comparison with different T.



GMM1 R=0.8 and K=2: Performance comparison on each patient with different T.



GMM2 R=0.8 and K=2: Performance comparison on each patient with different T.



PMM R=0.8 and K=2: Performance comparison on each patient with different T.



GMM1 R=1 and K=6: Performance comparison on each patient with different T.



GMM2 R=1 and K=6: Performance comparison on each patient with different T.



PMM R=1 and K=6: Performance comparison on each patient with different T.

Powered by TCPDF (www.tcpdf.org)


http://www.tcpdf.org

a personalized monitoring model for electrocardiogram …

Documents