`
`OPEN
`REVIEW ARTICLE
`Photoplethysmography based atrial fibrillation detection: a
`review
`Tania Pereira 1*, Nate Tran1, Kais Gadhoumi1, Michele M. Pelter1, Duc H. Do2, Randall J. Lee3, Rene Colorado4, Karl Meisel4 and
`Xiao Hu 1,5,6,7
`
`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.
`
`npj Digital Medicine (2020) 3:3 ; https://doi.org/10.1038/s41746-019-0207-9
`
`1234567890():,;
`
`INTRODUCTION
`Atrial fibrillation (AF) is an abnormal cardiac rhythm characterized
`by a disorganized atrial activity. AF is recognized in the
`electrocardiogram (ECG) as an irregularly irregular rhythm lasting
`more than 30 s, with no discernible P-waves preceding the QRS
`complex.1 AF prevalence is age, gender, and race dependent.2 It is
`particularly high in the elderly population, reaching 10–17% in
`subjects 80 years and older.3 In addition, AF is more prevalent in
`males and in the white population.3 AF is associated with
`significant morbidity and mortality. One in five strokes is
`associated with AF and one-third of cardiac arrhythmias
`hospitalizations are due to AF-related complications. AF has been
`associated with a twofold increase in the risk of death.4
`Additionally, the aging population in the US and worldwide is
`leading to a markedly increasing AF prevalence3,5.
`The high prevalence of asymptomatic AF has significant clinical
`implications on the diagnosis and management of AF.6 Inter-
`mittent ECG evaluation during clinical visits has a low likelihood of
`detecting paroxysmal AF. Continuous monitoring would increase
`the chances of AF detection, thereby allowing appropriate primary
`and secondary stroke prevention strategies to reduce the high
`morbidity and mortality of stroke.
`For patients with acute ischemic stroke or transient ischemic
`attack, approximately 10% will have new AF detected during their
`hospital admission.7–9 Continuous ECG monitoring for 30 days is
`recommended in case of an embolic stroke of undetermined
`cause (cryptogenic).9 Novel non-intrusive approaches for cardiac
`rhythm monitoring can potentially enable early and accurate
`detection of asymptomatic paroxysmal AF and create a shift in AF
`management.10,11 Especially for asymptomatic AF cases, new tools
`
`that allow the AF detection will help make the appropriate clinical
`decisions.10
`Photoplethysmography (PPG) has emerged as a low-cost and
`non-intrusive modality for continuous monitoring of heart rate. A
`variety of wearable devices offer PPG-based monitoring, including
`smartphones and smartwatches. A photoplethysmogram is a
`pulse pressure signal resulting from the propagation of blood
`pressure pulses along arterial blood vessels. Measured on the
`periphery,
`it carries rich information about the cardiac activity,
`cardiovascular condition, the interaction between parasympa-
`thetic and sympathetic nervous systems, and hemoglobin level.12
`Many physiological parameters can be derived from PPG,
`including oxygen saturation, heart
`rate, blood pressure, and
`cardiac output.13 These capacities of PPG open the door to
`develop new ambulatory diagnosis tools enabling early screening
`of heart conditions, including arrhythmia.14
`This review provides an account of the approaches used in PPG-
`based AF detection. A brief overview of the technology behind
`PPG is first presented, followed by a summary of methods and
`algorithms developed for PPG-based AF detection. Recognizing
`the importance of using PPG to detect AF at scale, the motivation
`of this review is to guide the future development of algorithms
`towards clinical-grade applications.
`
`PHOTOPLETHYSMOGRAPHY
`PPG signal
`PPG waveform is generated during a cardiac cycle and typically
`measured at a peripheral site. Therefore, it is essentially a pulse
`pressure waveform that originates from the heart contraction and
`propagates through the vascular tree. As blood flow is controlled
`
`1Department of Physiological Nursing, University of California, San Francisco, CA, USA. 2David Geffen School of Medicine, University of California, Los Angeles, CA, USA.
`3Cardiovascular Research Institute, Department of Medicine, Institute for Regeneration Medicine, University of California, San Francisco, CA, USA. 4Department of Neurology,
`School of Medicine, University of California, San Francisco, CA, USA. 5Department of Neurosurgery, School of Medicine, University of California, Los Angeles, CA, USA.
`6Department of Neurological Surgery, University of California, San Francisco, CA, USA. 7Institute of Computational Health Sciences, University of California, San Francisco, CA, USA.
`*email: taniapereira10@gmail.com
`
`Scripps Research Translational Institute
`
`AliveCor Ex. 2023 - Page 1
`
`
`
`T. Pereira et al.
`
`Normal Sinus Rhythm
`
`PPG signal
`
`PPG Measurement
`
`%SpO2
`98
`
`PRbpm
`67
`
`AFib
`
`Noise Artifact
`
`Fig. 1 PPG signal acquired using a wearable device and typical waveforms representing NSR, AF, and noise artifact.
`
`by neural, cardiac, and respiratory interactions, various physiolo-
`gical parameters could theoretically be extracted from analyzing a
`PPG signal.15 For this reason, the PPG signal has rich information
`about physiological conditions.13
`PPG waveforms have typical morphological components
`corresponding to landmark events in the cardiac cycle. During
`the contraction of the left ventricle, blood is ejected out of the
`heart and propagates along the arterial tree, this corresponds to
`the initial positive slope of a PPG pulse. The systolic peak marks
`the maximum of the waveform. A decrease in amplitude following
`the systolic peak is marked by a local minimum, or the dicrotic
`notch, which corresponds
`to the closing of aortic valves
`separating the systolic and diastolic phases.
`In some cases, a
`third peak following the dicrotic notch can be identified.
`It
`corresponds to a reflected component of the forward wave from
`various reflection sites including vessel bifurcations.16
`
`Clinical parameters
`One primary clinical application of PPG is arterial blood oxygen
`saturation (SpO2) estimation through pulse oximetry.17 SpO2 is
`defined as the percentage of oxygen saturation in the arterial
`blood, which can be measured by the ratio of oxygenated
`hemoglobin concentration to the total hemoglobin concentration,
`with a normal range between 97% and 98%.18 Recently, new
`applications of PPG have emerged for the continuous estimation
`of valuable cardiovascular parameters in ambulatory settings.
`Heart rate, blood pressure, and respiratory rate could be closely
`monitored for fitness or health assessment.19 Advanced diagnostic
`applications of PPG were also envisaged. Cardiac function, arterial
`stiffness, autonomic nervous system (ANS) responses, and apnea
`are among conditions that could potentially be detected or
`evaluated using PPG.
`Changes in blood volume are synchronous with the heart beats,
`such synchrony is manifested by the concordance of inter-beat
`intervals (RR intervals) measured in PPG and time-synchronized
`ECG.20 Heart rate variability (HRV) is an indirect measurement of
`ANS, and it has also been considered as a surrogate parameter of
`the interaction between the brain and cardiovascular system.21
`HRV metrics can be derived from analyzing RR intervals in time
`and/or frequency domain as well as using nonlinear dynamic
`analysis approaches.22 Respiratory rate is one of the fundamental
`vital signs and can be determined from the time–frequency
`representation of a PPG signal.23
`Some hemodynamic parameters such as augmentation index
`(AIx) and pulse wave velocity (PWV) are important biomarkers of
`arterial stiffness, which is a direct cause of hypertension and a
`major risk factor for cardiovascular events such as myocardial
`
`infarction and stroke. Both AIx and PWV could be derived from
`PPG,24,25 Subendocardial Viability Ratio (SEVR %) and Ejection
`Time Index (ETI) are two hemodynamic parameters used in the
`evaluation of cardiac workload that can be estimated with PPG
`analysis.25 Additionally, some studies claim that arterial blood
`pressure could be estimated using advanced analysis of PPG.17
`
`Modes of PPG measurement
`A PPG signal has two main components: a quasi-static direct
`current (DC) component, which represents light reflected/trans-
`mitted from static arterial blood, venous blood, skin and tissues;
`and pulsatile alternate current (AC) component which arises from
`modulation in light absorption due to changes in arterial blood
`volume. PPG measurement can be carried out using two modes:
`transmission and reflectance.
`In transmission mode, the light
`transmitted through the medium is detected by a photodetector
`(PD), which is positioned in the opposite site of the light source.
`The sensor must be located on the body at a site where
`transmitted light can be detected. The measurement site is limited
`to the extremities of the body, such as the fingertip or earlobe.
`The greatest disadvantage of
`the transmission mode is the
`location of
`the device that can interfere with daily routine
`movements.26 In reflectance mode, the PD detects light that is
`back scattered or reflected from tissues, bone, and/or blood
`vessels, which means the light source and PD are positioned on
`the same side. Unlike the transmission mode, the measurement
`sites are not restricted to any particular location, which facilitates a
`user-friendly monitoring approach. The wrist, forearm, ankle, and
`forehead are common measurement sites.27
`Since the basic form of PPG technology requires only a few
`optoelectronic components (a light source and a PD: to measure
`the variations on the light reflected/transmitted by the tissues), it
`can be easily and inexpensively incorporated in various digital
`devices such as watches, smartphones, or wearables.28 The
`ubiquitous availability of PPG in a wide range of wearable digital
`devices has motivated the search for new applications and the
`development of novel biomedical solutions.
`
`PPG-BASED AF DETECTION
`In a PPG signal, AF is manifested as varying pulse-to pulse intervals
`and pulse morphologies. On the other hand, a normal sinus
`rhythm (NSR) is recognizable through regularly spaced PPG pulses
`with similar morphologies between consecutive pulses. Recogniz-
`ing an arrhythmia in a PPG signal can sometimes be challenging in
`the presence of artifacts. Common sources of artifacts are motion
`and poor sensor contacts. Artifacts can be misinterpreted as
`
`npj Digital Medicine (2020) 3
`
`Scripps Research Translational Institute
`
`2
`
`1234567890():,;
`
`AliveCor Ex. 2023 - Page 2
`
`
`
`T. Pereira et al.
`
`3
`
`Smart Watch
`
`Smart Phone
`
`Fingertip
`
`Ear Clip
`
`%SpO2
`98
`
`PRbpm
`67
`
`RR Intervals
`
`Time series
`
`Image representation
`
`PPG Measurement
`Device
`
`PPG Representations
`
`PPG Length
`
`[10 sec: 10 min]
`
`Traditional Statistics
`
`Machine Learning
`
`Deep Learning
`
`Classification
`approach
`
`NSR
`
`AF
`
`PPG feature
`
`Number of patients
`
`Computational resources
`
`Increase the amount of available data
`
`Hundreds
`
`Hundreds
`
`Thousands
`
`Fig. 2 Overview of the main features extracted from PPG signals used in the studies reviewed (see Tables 1–3). SpO2 oxygen saturation,
`PRbpm pulse rate (beats per minute).
`
`physiological abnormalities. Motion artifacts can be identified
`using accelerometry data. Most modern wearable devices include
`accelerometry sensors that measure acceleration forces along
`different spatial directions. It is a common practice to discard PPG
`contaminated with an artifact. Figure 1 depicts samples of PPG
`with NSR, AF, and artifact.
`ECG remains the gold standard for the electrophysiological
`definition and recognition of arrhythmias,1 including AF diag-
`nosis.29 In a recent study, new deep learning approaches achieved
`cardiologist-level AF detection of 12 types of arrhythmia (F1 score =
`0.84 vs F1 score = 0.78) when 91,232 single-lead ECGs from 53,549
`patients were analyzed.30 Compared to ECG, PPG-based AF
`detection is more challenging but also rewarding in situations
`where longer monitoring time and lower cost beyond what ECG
`offers is needed, e.g., screening AF at scale.
`Recent advances in sensor technologies and wearable devices
`have increased the role that a PPG-based solution could play in
`the assessment of health status. Electronics capable of recording
`PPG signals with relatively high signal-to-noise ratio (SNR) may
`warrant reliable PPG monitoring and screening of arrhythmia.11,31
`In a typical AF detection algorithm, features (temporal, spectral,
`or morphological) are extracted from the acquired PPG signal and
`analyzed by the detection algorithm to inform if an AF rhythm is
`detected.
`In some approaches,
`image representation of
`the
`temporal waveform has been considered. The derived image
`would then be analyzed using conventional image processing or
`intelligence-based methods (Fig. 2).32–34 Traditionally,
`artificial
`prominent
`features were derived from the tachogram (RR
`intervals) since it is a reliable measure of heart beats.35 Realizing
`that PPG waveforms may carry physiological information beyond
`heart rate, new features beyond RR intervals were derived.36 The
`use of PPG time series and their images representation (e.g. raw
`
`plot of the signal, fast Fourier transform spectrum, or wavelet
`spectrogram—represented in the Fig. 2 in PPG representation
`part) were used with promising results in the detection of
`physiological events,32,37,38 Images for PPG representation in Fig. 2
`is a general depiction of the format types of information used by
`the different algorithms.
`In the following sections, we review studies of PPG-based AF
`detection. A body of white papers and peer-reviewed works
`indexed by PubMed, Scopus, IEEE Xplore, and Web of Science up
`to June 2019 was selected based on the following search
`(PPG “OR” Photoplethysmography)
`“AND”
`expression:
`(atrial
`fibrillation “OR” AF “OR” AFib) “AND” (detection “OR” recognition).
`Each study is reviewed with respect to the size, the number of
`patients, and recording settings of data analyzed, the PPG device
`and site of
`recording,
`the AF detection algorithm, and its
`performance. Figure 2 summarizes the main features examined
`in these studies, described with more details in Tables 1–3.
`
`Performance metrics
`AF detection algorithms can be evaluated using several perfor-
`mance metrics. It is common for many studies to report sensitivity,
`specificity, and accuracy. Sensitivity is defined as the probability to
`detect
`true AF events, while the specificity measures the
`proportion of actual Non-AF instances correctly identified as such.
`Accuracy is a balanced metric of sensitivity and specificity. The
`accuracy of an AF detection algorithm is its ability to differentiate
`between AF and Non-AF cases.39 Generally, accuracy is the most
`common reported metric, along with the area under the curve
`(AUC) of the receiver operating characteristic (ROC). A ROC for
`differentiating AF vs Non-AF is generated by plotting sensitivity vs
`(1-specificity) at different classification thresholds. AUC is a
`
`Scripps Research Translational Institute
`
`npj Digital Medicine (2020) 3
`
`AliveCor Ex. 2023 - Page 3
`
`
`
`4
`
`T. Pereira et al.
`
`Sp=0.9539
`Sen=0.9716,
`Acc=0.9645,
`
`forbestROC
`valuesoffeatures
`Derivedthreshold
`
`andShannonentropy
`coherencefunctions
`Timevarying
`RRtimesseriesfeatures
`
`detection
`forrhythms
`Performanceresults
`
`Methodology
`
`Inputdata
`
`Inpatient
`
`conditions
`Acquisition
`
`smartphone
`Videocameraof
`
`2min
`
`Measurementdevice
`
`segments
`LengthPPG
`
`population
`Ageof
`
`Datasetfeatures
`
`patients
`Numberof
`
`(year)[ref.]
`Author
`
`Studiesonphotoplethysmography-basedAFdetectionusingstatisticalanalysisapproaches.
`
`Table1.
`
`–
`
`–
`
`symptomatic
`15withAF31non-
`Database)
`Arrhythmia
`BIHNSR+MIT-BIH
`(MIT-BIHAF+MIT-
`Publicdatabases
`cardioversion+
`74priorandafter
`
`46
`
`(2016)52
`Nematietal.
`
`74
`
`(2013)91
`Leeetal.
`
`Sp=0.909
`Sen=0.909,
`Acc=0.952,
`
`forbestROC
`valuesoffeatures
`Derivedthreshold
`
`AUC=0.95
`Acc=0.918,
`Sp=0.976
`Sen=0.733,
`Acc=0.960,
`Sp=0.980.PVC:
`Sen=0.667,
`Acc=0.955,
`Sp=0.935.PAC:
`Sen=0.970,
`AF:Acc=0.951,
`
`logisticmodel
`Elasticnet
`
`forbestROC
`valuesoffeatures
`Derivedthreshold
`
`Sp=0.950
`Sen=0.950,
`AUC=0.931,
`
`forbestROC
`valuesoffeatures
`Derivedthreshold
`
`deviation,averageof
`variation,standard
`features.Coefficientof
`RRtimesseries
`entropyfeatures)
`maxofthesample
`deviation,minandthe
`versionofthestandard
`deviation,robust
`entropy,standard
`seriesfeatures(sample
`theCNN+RRtimes
`featuresobtainedfrom
`wavelettransform—
`representationof
`PPGimagespectral
`
`Entropy,Poincareplot
`RMSSD,Shannon
`RRtimesseriesfeatures
`thePoincareplot
`indexextractedfrom
`RMSSDandSD1/SD2
`RRtimesseriesfeatures
`
`Sp=0.99±0.03
`Sen=0.97±0.02,
`
`AUC=0.99
`Sp=0.94,
`Sen=0.97,
`Acc=0.95,
`
`Markovmodel
`
`RRtimesseriesfeaturesFirst-order11-state
`standarddeviation
`deviation,robust
`SampEn2),standard
`2(SampEn1and
`dimensionsm=1,and
`theembedding
`sampleentropywith
`RRtimesseriesfeatures
`
`logisticmodel
`ElasticNet
`
`Inpatient
`
`PPGsensor
`Singleearlobe
`
`5min
`
`68±11
`45±17,
`64±11,
`38±12,
`
`withAF
`healthysubjects,21
`33withAF,13
`44healthysubjects,
`
`77Test:34
`
`(2017)94
`T.Conroyetal.
`
`Inpatient
`
`SamsungSimband
`Wrist-worndevice
`
`30s
`
`–
`
`otherrhythms(ARR)
`45withAF,53with
`
`98
`
`etal.(2017)51
`Shashikumar
`
`Inpatient
`
`checkpoint
`outpatient
`In-and
`
`measurement
`continuous
`Outpatient—
`
`smartphone
`Videocameraof
`
`2min
`
`66
`
`PAC15withPVC
`98withAF15with
`
`121
`
`etal.(2016)93
`D.McManus
`
`smartphone
`Videocameraof
`MonitoringModule
`Motion
`PhilipsCardioand
`Wrist-wearablesensor—
`
`80±875±75min
`
`AF
`40withAF40Non-
`
`30s
`
`65.2±14.0
`
`flutter,11NSR
`4withAF,1atrial
`
`80
`
`16
`
`etal.(2016)92
`J.Eckstein
`
`(2016)50
`Bonomietal.
`
`Inpatient
`
`SamsungSimband
`Wrist-worndevice
`
`3.5to8.5min
`
`npj Digital Medicine (2020) 3
`
`Scripps Research Translational Institute
`
`AliveCor Ex. 2023 - Page 4
`
`
`
`T. Pereira et al.
`
`5
`
`Sp=0.980
`Sen=0.984,
`Acc=0.981,
`Sp=0.9913
`Sen=0.9845,
`
`Sp=0.9714
`Sen=0.9765,
`Acc=0.9667,
`
`Acc=0.9116
`
`AUC=0.973
`10-min:
`2-min:AUC=0.972,
`1-min:AUC=0.949,
`
`detection
`forrhythms
`Performanceresults
`
`prematureatrialcontraction,pNN35/pNN40/pNN70percentageofdifferencesofsuccessiveRRthatexceeded35or40or70msbythetotalnumberofRRintervals
`RRRtoRinterval,NSRnormalsinusrhythm,ARRotherarrhythmias,VAventriculararrhythmias,AUCareaunderthecurve,Accaccuracy,Sensensitivity,Spspecificity,PVCprematureventricularcontractions,PAC
`
`forbestROC
`valuesoffeatures
`Derivedthreshold
`
`entropy
`coefficientofsample
`sampleentropy,and
`RMSSD,pNN40,pNN70,
`RMSSD,normalized
`Shannonentropy,
`RRtimesseriesfeatures
`
`thresholds
`detection:
`thresholds.AF
`detection:
`Noise/movement
`
`vectormachines
`detection:Support
`thresholds.AF
`detection:
`Noise/movement
`
`RRtimesseriesfeaturesMarkovmodel
`ShannonEntropy(ShE)
`detection:RMSSDand
`kurtosischange.AF
`ratiochanges,and
`changes,turningpoint
`detection:Signalslope
`Noise/movement
`sampleentropy)
`Shannonentropyand
`seriesfeatures(RMSSD,
`detection:RRtimes
`demodulation.AF
`frequencycomplex
`detection:Variable
`Noise/movement
`pointratio
`entropy,andturning
`methods:Shannon
`HF.Nonlinearanalytical
`andtheratioofLFand
`frequencyrange(HF),
`(LF),powerinthehigh-
`low-frequencyrange
`Frequencydomain:
`deviation,andRMSSD.
`mean,standard
`features.Timedomain:
`RRtimesseries
`to-beat:pNN35
`thedifferenceinbeat-
`
`analysis
`Logisticregression
`
`measurement
`continuous
`Outpatient—
`
`Inpatient
`
`PhilipsCardio
`Wrist-worndevice
`PulseOnLtd.
`Wrist-worndevice
`
`120s
`consecutiveRR
`20
`
`67±13
`69±101,
`74.8±8.3
`67.5±10.7,
`
`8AF19non-AF
`
`15NSR,14withAF
`
`27
`
`29
`
`(2018)74
`Morreeetal.
`H.M.de
`(2018)49
`Tarniceriuetal.
`
`checkpoint
`Outpatient—
`
`smartphone
`Videocameraof
`
`2min
`
`–
`
`healthsubjects
`cardioversion11
`priorandafter
`88patientswithAF
`
`99
`
`(2018)65
`Chongetal.
`
`checkpoint
`Outpatient—
`
`smartphone
`Videocameraof
`
`30s
`
`45yearsold
`Olderthan
`
`–
`
`200
`
`(2018)64
`Basharetal.
`
`Inpatient
`
`Bedsidemonitor
`
`10min
`1min,2min,
`
`66.3±14.8
`74.5±12.8,
`
`Non-AF
`150withAF,516
`
`patients
`666stroke
`
`(2017)48
`Tangetal.
`
`Methodology
`
`Inputdata
`
`conditions
`Acquisition
`
`Measurementdevice
`
`segments
`LengthPPG
`
`population
`Ageof
`
`Datasetfeatures
`
`patients
`Numberof
`
`(year)[ref.]
`Author
`
`Table1continued
`
`Scripps Research Translational Institute
`
`npj Digital Medicine (2020) 3
`
`AliveCor Ex. 2023 - Page 5
`
`
`
`prematureatrialcontraction
`RRRtoRinterval,NSRnormalsinusrhythm,ARRotherarrhythmias,VAventriculararrhythmias,AUCareaunderthecurve,Accaccuracy,Sensensitivity,Spspecificity,PVCprematureventricularcontractions,PAC
`
`=0.962,Sp=0.928
`(NSR&VA):Acc=0.950,Sen
`0.981,Sp=0.887.AFvs
`VA:Acc=0.959,Sen=
`=0.997,Sp=0.924.AFvs
`AFvsNSR:Acc=0.981,Sen
`
`Decisiontrees
`
`T. Pereira et al.
`
`PerfectdetectionofAF
`Sen=0.758,Sp=0.768
`Sp=0.963.ARR:
`Sen=0.754,
`Sp=0.928.AF:
`NSR:Sen=0.773,
`
`machines
`Supportvector
`
`KNNclassifier
`
`Acc=0.9385
`
`machine
`Supportvector
`
`Acc=0.957
`Sen=0.942,
`AUC=0.971,
`
`machines
`Supportvector
`
`andspectralpurityindex).
`fractionalspectralradius,
`entropy,spectralentropy,
`difference,permutation
`oftheslopeofthephase
`organizationindex,variance
`features:adaptive
`RMSSD.PPG-waveform
`minimum,maximum,
`median,interquartile,
`mean,standarddeviation,
`RRtimesseriesfeatures:
`spectralpowers
`andquotientsofthese
`frequency,highfrequency
`verylowfrequency,low
`consecutivepulsesegments,
`cross-correlationof
`height,waveformwidth,
`time,peakriseheight,fall
`median,SDandMAD,crest
`waveformfeatures:mean,
`Shannonentropy.PPG-
`RMSSD;normalizedRMSSD;
`absolutedeviation(MAD);
`deviationandthemean
`Mean,median,standard
`RRtimesseriesfeatures
`features
`analysis,PPG-waveform
`variabilityandirregularity
`features:spectralanalysis,
`RRtimesseriesfeatures24
`rangeofRR
`andinterquartile
`mean,minimum,median,
`RRtimesseriesfeatures
`pointratio
`Shannonentropy,Turning
`multi-scaleentropy,
`powerinLFandHF(LF/HF),
`frequency(HF),ratioof
`lowfrequency(LF),high
`verylow-frequencyrange,
`deviation,RMSSD,powerin
`Mean,median,standard
`RRtimesseriesfeatures.
`
`Inpatient
`
`Wrist-worndevice
`
`10s
`
`57±13
`
`arrhythmias(VA)
`AFNSRventricular
`
`17
`
`(2019)60
`etal.
`S.Fallet
`
`Inpatient
`
`smartphone
`Videocameraof
`
`20s
`
`–
`
`NSR,12ofNoise
`20withAF,294of
`
`EmpaticaE4wristbandInpatient
`
`2min
`
`9,65±15
`40±17,76±
`
`ARR,31NSR
`30withAF,9
`
`Inpatient
`
`baseddevice
`PPGwrist-
`
`10s
`
`Inpatient
`
`Bedsidemonitor
`
`2min
`
`–
`
`–
`
`326
`
`70
`
`20
`
`(2017)95
`etal.
`T.Schack
`
`(2017)58
`etal.
`V.Corino
`
`(2016)57
`etal.
`M.Lemay
`
`–
`
`patients
`468stroke
`
`(2016)56
`Shanetal.
`
`rhythmsdetection
`Performanceresultsfor
`
`Methodology
`
`Inputdata
`
`conditions
`Acquisition
`
`Measurementdevice
`
`segments
`PPG
`Length
`
`population
`Ageof
`
`Datasetfeatures
`
`patients
`Numberof
`
`(year)[ref.]
`Author
`
`StudiesonphotoplethysmographybasedAFdetectionusingMLapproaches.
`
`Table2.
`
`6
`
`npj Digital Medicine (2020) 3
`
`Scripps Research Translational Institute
`
`AliveCor Ex. 2023 - Page 6
`
`
`
`7
`
`prematureatrialcontraction,AUCprareaundertheprecision–recallcurve
`RRRtoRinterval,NSRnormalsinusrhythm,ARRotherarrhythmias,VAventriculararrhythmias,AUCareaunderthecurve,Accaccuracy,Sensensitivity,Spspecificity,PVCprematureventricularcontractions,PAC
`
`T. Pereira et al.
`
`AUC=0.996
`0.998.Acc=0.9715,
`Acc=0.9758,AUC=
`
`Acc=0.95
`AUCpr=0.97,Sp=1.0,
`
`AUC=0.97,
`
`neuralnetwork
`network.Recurrent
`convolutionalneural
`1-Dimensional
`
`transferlearning
`Network(BRNN)with
`RecurrentNeural
`Bidirectional
`
`AUC=0.949
`
`1DResNet
`
`Sen=0.999,Sp=0.998
`
`AUC=0.999,
`
`network
`recurrentneural
`Convolutional-
`
`Inpatient
`
`checkpoint
`Outpatient—
`measurement
`continuous
`
`PPGsegment
`based-test
`image-
`training.PPG
`based–
`ECGimage-
`
`PPGsegmentOutpatient—
`measurement
`asleepcontinuous
`OutpatientNSR:
`Inpatient+
`
`PPGsegment
`
`PPGfingertip
`
`30s
`
`63±7.8
`
`SamsungSimband
`Wrist-worndevice
`
`SamsungSimband
`Wrist-worndevice
`
`yearsold
`18–89
`years.Test:
`
`30s
`
`30s
`
`Test:44AFand53ARRTrain:47±25
`20NSR
`AF/32NSR.Test:10AF/
`Train+validation:19
`
`—
`
`fitnesstrackerdevice
`Wrist-wornprototype
`
`—
`
`yearsold
`37–85
`
`Sp=0.965
`0.988.AF:Sen=0.976,
`Sen=0.722,Sp=
`0.991,Sp=0.982.ARR:
`Sp=1.NSR:Sen=
`Noise:Sen=0.970,
`OverallAcc=0.961.
`0.677,Sp=0.676
`AUC=0.72,Sen=
`0.902.Testset1:
`Sen=0.980,Sp=
`Testset1:AUC=0.97,
`
`AUC=0.9967
`Acc=0.9819,
`
`withsixdenseblocks
`CNNarchitecture
`
`layers
`Neuralnetworkof8-
`Model(CRNN)
`RecurrentHybrid
`Convolution-
`CNNAF.detection:
`Qualityclassification:
`
`PPGsegmentOutpatient
`
`Smartphone
`
`17s
`
`68.4±12.2
`Testgroup:
`
`checkpoint
`2:Outpatient
`InpatientTest.set
`Testset1:
`
`seriesfeatures
`RRtimes
`
`Applewatch
`Wrist-worndevice
`
`5s
`
`66.1±10.7
`Testset1:
`Train:42±12.
`
`8216noAF
`Train:347withAF,
`
`PPGsegment—
`
`Samsunggeardevice
`
`30s
`
`—
`
`AF4NSR
`NSR.Test:7with
`Train:29withAF13
`dataset)
`RateBenchmark
`TBMEPPGRespiratory
`volunteers+IEEE-
`datasetfromhealthy
`database+Vortal
`(MIMIC-IIIcriticalcare
`Train:Publicdatabases
`
`75
`
`(2019)96
`S.Kwonetal.
`
`Test:97
`Train:2850.
`
`etal.(2018)32
`S.Shashikumar
`
`81
`
`(2019)81
`M.Voisinetal.
`
`Test:11
`Train:42.
`
`etal.(2018)97
`I.Gotlibovych
`
`Test:1013
`Train:3373.
`
`(2018)73
`M.Pohetal.
`
`2:1617
`Testset
`Testset1:51.
`Train:9750.
`
`(2018)72
`Tisonetal.
`
`19
`
`Shen(2018)75
`Aliamiriand
`
`forrhythmsdetection
`Performanceresults
`
`Methodology
`
`conditions
`Acquisition
`
`Inputdata
`
`Measurementdevice
`
`segments
`PPG
`Length
`
`population
`Ageof
`
`Datasetfeatures
`
`patients
`Numberof
`
`(year)[Ref]
`Author
`
`StudiesonphotoplethysmographybasedAFdetectionusingDLapproaches.
`
`Table3.
`
`Scripps Research Translational Institute
`
`npj Digital Medicine (2020) 3
`
`AliveCor Ex. 2023 - Page 7
`
`
`
`8
`
`T. Pereira et al.
`
`measure of how well AF cases ranked higher than Non-AF cases.
`Since AF has a low prevalence it is generally required that PPG-
`based AF detectors show high precision (positive predictive value).
`Rather than reporting the AUC, the area under the precision-recall
`curve (AUPRC)
`is an alternative metric suitable for highly
`imbalanced data (i.e. low prevalence).40 In general, any reported
`performance metric should take into account the low prevalence
`of AF and be evaluated on an independent test dataset.
`
`AF detection studies
`Studies were split into three groups based on the approaches
`undertaken to build an AF detector: traditional statistical analysis,
`machine learning (ML), and deep learning (DL) methods.
`In
`traditional statistical analysis, statistical metrics are derived from
`PPG signals, and classification thresholds were estimated to
`distinguish between AF and Non-AF segments. ML techniques call
`for the extraction of pre-selected features, a process that can be
`quite manual,
`labor-intensive, and can usually benefit
`from
`incorporating complex physiological knowledge. An ML classifier
`is then built upon extracted features from training data samples.
`DL approaches require less manual
`feature engineering than
`conventional ML since DL incorporates automatic features
`representation process of
`input data. Recently,
`there was a
`significant focus on DL methods driven notably by technological
`advancement in computational power and the acclaimed success
`in computer vision applications.41,42
`
`Statistical analysis approaches
`Statistical models for AF detection are built using the thresholds
`for a set of features extracted from the RR-interval time series of
`well-annotated and publicly available ECG databases, such as MIT-
`BIH atrial fibrillation, MIT-BIH normal sinus rhythm, or MIT-BIH
`arrhythmia database.43–45 Specifically, features were first extracted
`from the RR-interval time series of pre-annotated ECG waveforms.
`The histograms of each feature were analyzed respectively with or
`without the presence of AF and other cardiac rhythms in order to
`define the threshold that best separates the rhythm classes. These
`thresholds were then applied to the same RR time series-based
`features extracted from PPG signals.46,47 Other arrhythmias (i.e.,
`premature ventricular contractions, and premature atrial contrac-
`tion) could also be detected similarly in a sequence of binary
`classifications.36
`Other statistical approaches can also be applied to classify
`between AF and Non-AF such as logistic regression.48 Logistic
`regression models use the logistic function, instead of a straight
`line or a hyperplane, to fit output the probability between 0 and 1
`(corresponding to Non-AF and AF). Markov model
`is another
`statistical tool that could be used for AF detection. RR-interval time
`series features are used in this model to define the distributions
`that best fit the data, and the probability for various rhythms can
`be drawn from these distributions,44,49,50 Elastic net
`is a
`regularization method for regression and classification models.
`Elastic net performs both variable selection and regularization in
`order to enhance the prediction accuracy and interpretability of
`the logistic regression model. Regularization approaches were
`successfully
`applied to improve the performance of AF
`detection.51,52
`Table 1 summarizes a selection of PPG-based AF detection
`studies which used statistical models. Different study aspects are
`shown to depict the patient population and datasets used, the
`features and methods, the context (inpatient vs outpatient), and
`the performance results.
`
`Machine learning approaches. ML has been used for AF detection
`with interesting results. ML techniques require extensive domain
`expertise to design features
`suitable for a comprehensive
`representation of PPG waveforms and the detection of class-
`
`differentiating patterns. Features commonly extracted from PPG
`time series are morphological descriptors, time domain statistics,
`frequency domain statistics, nonlinear measures, wavelet based
`measures, and cross-correlation measures.53–60
`There were generally three main ML approaches used in the
`reviewed studies: k-nearest neighbors (KNN), support vector
`machine (SVM), and decision trees (DT). KNN classification is a
`relatively simple clustering technique where a sample is classified
`by a plurality vote of its neighbors and assigned to the class based
`on the most common class among its k closest neighbors.61
`SVM finds a hyperplane that separates two classes with a high
`margin that maximizes the distances between nearest data points
`from each class. SVMs prove to be successful
`in nonlinear
`classification problems by mapping non-separable features into
`a higher dimensional space, a procedure known as the kernel trick
`which uses kernel functions such as Radial Basis Function (RBF) or
`polynomial.62
`is continuously split
`the training set
`In DT approaches,
`according to a chosen feature. A feature tree can be explained
`by two entities, namely decision nodes and leaves. The leaves are
`the decisions or final outcomes. And the decision nodes are where
`the data is split.63 The objective is to find in each decision node of
`the tree, the best attribute allowing to diminish as much as
`possible the overlapping of classes. The classification starts from
`the root, and it evaluates the relative attribute and it takes the
`branch corresponding to the outcome. This process is repeated
`until a leaf is encountered and a sample is assigned the class
`labeling the leaf.
`Some studies used a combination of threshold-based and ML
`approaches. For example, the thresholds of some features were
`first used to exclude poor pulses, then an ML model was built for
`the detection of AF in the clean pulses.64,65 Table 2 is a
`chronological summary of
`the selected ML studies and the
`reported performance results. All the studies reported in the Table
`2 were based in short length of PPG segment, with maxima of
`2 min.
`
`Deep learning approaches. DL has recently emerged as a
`powerful method for the detection of abnormalities in physiolo-
`gical signals, encouraging applications of arrhythmia detection
`from ECG and PPG,
`including AF detection. Unlike ML, deep
`learning models automatically learn feature representations,
`sparing the tedious task of feature crafting. DL uses a neural
`network, a set of interconnected layers of computational nodes.
`The most common DL approaches used for AF detection are
`based on Convolutional Neural Networks (CNN). CNN was applied
`in automatic feature extraction and in classification problems.
`Some studies used CNNs only for automatic feature extraction.51
`In one stu