European Journal of Internal Medicine 73 (2020) 72–75

Contents lists available at ScienceDirect

European Journal of Internal Medicine

journal homepage: www.elsevier.com/locate/ejim

Original article

Prospective blinded Evaluation of the smartphone-based AliveCor Kardia T

ECG monitor for Atrial Fibrillation detection: The PEAK-AF study
Felix K. Wegnera,⁎, Simon Kochhäusera,b, Christian Ellermanna, Philipp S. Langea,
Gerrit Frommeyera, Patrick Leitza, Lars Eckardta, Dirk G. Decheringa
a
Department of Electrophysiology, University Hospital Muenster, Muenster, Germany
b
Current address: Department of Cardiology and Vascular Medicine, West German Heart and Vascular Center Essen, University Duisburg-Essen, Essen, Germany

ARTICLE INFO ABSTRACT

Keywords: Introduction: The AliveCor Kardia ECG monitor (ACK) offers a smartphone-based one-lead ECG recording for the
Atrial fibrillation detection of atrial fibrillation. We compared ACK lead I recordings with the 12-lead ECG and introduce a novel
Stroke, ECG parasternal lead (NPL).
EHealth Methods: Consecutive cardiac inpatients were recruited. In all patients a 12-lead ECG, ACK lead I and NPL were
Prevention
obtained. Two experienced electrophysiologists were blinded and separately evaluated all ECG. We calculated
sensitivity, specificity, and predictive values of the ACK ECG compared to the 12-lead ECG.
Results: 296 ECG from 99 patients (38 female, age 64 ± 15 years, BMI 27.8 ± 5.1 kg/m2) were analyzed. 20% of
ACK lead I recordings contained a critical amount of artifact. The electrophysiologists’ interpretation of the ACK
recordings yielded a sensitivity of 100% and specificity of 94% for atrial fibrillation or flutter in lead I (κ = 0.90)
and a sensitivity of 96% and specificity of 97% in the NPL (κ = 0.92). The ACK diagnostic algorithm displayed a
significantly lower sensitivity (55–70%), specificity (60–69%), and accuracy (κ = 0.4–0.53) but a high negative
predictive value (100%). Patients with atrial flutter (n = 5) and with ventricular stimulation (n = 12) had a high
likelihood of being misclassified by the algorithm.
Conclusion: The AliveCor Kardia ECG monitor allows a highly accurate detection of atrial fibrillation by an
interpreting electrophysiologist both in the standard lead I and a novel parasternal lead. The diagnostic algo-
rithm offered by the system may be useful in screening recordings for further review. Diagnostic challenges
present in atrial flutter and ventricular pacemaker stimulation.

1. Introduction scale AF screening, especially outside of traditional health care settings.

The app has been reported to use a Random Forest machine learning
Stroke is the second leading cause of death worldwide and causes algorithm to differentiate sinus rhythm from AF and other rhythm
significant morbidity, especially if the source is a cardiogenic embolism disturbances [8].
[1,2]. Atrial fibrillation (AF) substantially increases an individual pa- The manufacturer-recommended lead I ACK recording may contain
tient's risk for stroke while oftentimes remaining asymptomatic [3]. A a critical amount of artifact, making an unequivocal rhythm diagnosis
diagnosis of AF can be reached using an ECG recording as short as 30 s difficult. We therefore introduce an easily reproducible novel para-
and can reduce the burden of stroke by establishing oral antic- sternal lead (NPL) by placing the ACK in a left parasternal position on a
oagulation. With the emergence of smartphones and smart watches supine patient (Fig. 1). No data have been published systematically
being increasingly used worldwide, technicians and policy makers alike comparing different ACK recording vectors to a gold standard 12-lead
see the opportunity to increase the likelihood of early AF diagnosis and ECG.
by this to reduce its sequelae [4–6]. A recent study by William et al. [8] excluded patients with an
The AliveCor Kardia (ACK) mobile ECG monitor was recently FDA- “unclassified” diagnosis by the ACK app algorithm, which constituted
certified as a one-lead ECG recorder [7]. It is a portable device with two about one in four patients, from a sensitivity and specificity analysis.
electrodes being able to write a single bipolar ECG lead utilizing a Since “unclassified” recordings contain all heart rates above 100/min
smartphone app. Its ease of use allows the ACK to be used for large- and all patients with ambiguous results, this may exclude the most

⁎
Corresponding author.
E-mail address: [email protected] (F.K. Wegner).

https://doi.org/10.1016/j.ejim.2019.11.018
Received 11 July 2019; Received in revised form 24 November 2019; Accepted 26 November 2019
Available online 02 December 2019
0953-6205/ © 2019 European Federation of Internal Medicine. Published by Elsevier B.V. All rights reserved.

Downloaded for Jan Svendsen ([email protected]) at Southwest Jutland Hospital Esbjerg from ClinicalKey.com by Elsevier on February 28, 2020.
For personal use only. No other uses without permission. Copyright ©2020. Elsevier Inc. All rights reserved.
F.K. Wegner, et al. European Journal of Internal Medicine 73 (2020) 72–75

Fig.. 1. Device placement and representative ECG recordings in the manufacturer-recommended lead I and the novel parasternal lead (NPL) proposed by this study.

relevant portion of patients from AF detection. We therefore designed 2.3. Statistical analysis
the present study to compare the diagnostic value of the ACK in a
prospective patient cohort with no predefined exclusion criteria to elicit For database management and statistical analysis, we used IBM
its use in everyday atrial fibrillation screening. SPSS Version 25 (IBM Corporation, Somers, NY, USA). Departure from
the mean was tested with Shapiro–Wilk's statistic. For a comparison of
2. Methods means, the student's t-test was used. Due to sample size, statistical
analysis of binary variables was conducted using the Fisher exact test.
For the present investigator-initiated, unsponsored study we pro- Where applicable, statistical significance was defined as a two-sided p-
spectively included 99 consecutive inpatients without further inclusion value of <0.05. κ values were calculated using the Spearman rank-
or exclusion criteria from our electrophysiology ward in a large tertiary- order correlation. A κ > 0.80 was defined as excellent agreement [9].
care university hospital. In each patient, a 12-lead ECG, an ACK lead I
recording and a novel parasternal lead were recorded sequentially. The 3. Results
local ethics committee approved the study protocol and all patients
gave informed consent. 38 of the 99 included patients were female. Mean age was 64 ± 15
years and mean BMI 27.8 ± 5.1 kg/m2. The NPL was able to be recorded
2.1. ECG acquisition in all patients. Seven patients (7%) were excluded from the sensitivity
and specificity analysis: In 6 patients, the 12-lead ECG was ambiguous (in
Patients were instructed to record a lead I ACK ECG by placing two all: continuous ventricular pacemaker stimulation with possible AF as
fingers of each hand on the electrodes of the ACK device. If a recording atrial rhythm). In one further patient, there was disagreement as to the
could not be obtained, different finger positions were allowed. The NPL 12-lead ECG rhythm diagnosis (electrophysiologist 1: AF, electro-
was recorded by asking the patient to lie supine with an exposed chest physiologist 2: sinus rhythm). The remaining 275 ECG of 92 patients
and breathe normally. The ACK device was then placed in a left para- were used for analysis. 65 patients exhibited sinus rhythm according to
sternal position so that a rhythm recording could be obtained (Fig. 1). the 12-lead ECG. 22 patients had AF and five had atrial flutter, resulting
The recordings were downloaded from the app and stored in the hos- in an arrhythmia prevalence in the study population of 29.4%.
pital IT system.
3.1. Electrophysiologists’ diagnosis
2.2. ECG analysis
Analyzing the ACK lead I, the blinded electrophysiologists char-
All ECGs were blinded and evaluated separately by two experienced acterized 74 ECG as being of adequate quality and 18 ECG (20%) as
electrophysiologists. The electrophysiologists were asked to either being of inadequate quality (Fig. 2). Using the NPL, 20 ECG (22%) were
classify the atrial rhythm as a) sinus rhythm or b) AF / atrial flutter. described as being of inadequate quality. Of note, inadequate quality of
Additionally, the diagnoses of the Kardia smartphone app were re- an ACK lead I recording was not associated with inadequate quality of
corded. These were a) sinus rhythm, b) possible AF, c) unclassified and the corresponding NPL recording and vice-versa (κ = 0.23).
d) no analysis. The diagnoses of the electrophysiologists using the A cross-tabulation of the electrophysiologists’ diagnoses based on
standard 12-lead ECG were defined as the gold standard. For intrinsic the respective ACK recordings with the gold-standard 12-lead ECG di-
validation, we first opted for interobserver agreement between the 12- agnoses is shown in Table 1. This resulted in a sensitivity of the elec-
lead ECG diagnoses of the two electrophysiologists. After that, we trophysiologists for the ACK lead I of 100% (27 of 27 patients with AF/
compared the individual smartphone diagnoses of the electro- atrial flutter detected) and a specificity of 94% (61 of 65 patients with
physiologists and of the app algorithm to the 12-lead ECG for sensitivity sinus rhythm correctly identified) with a κ of 0.90. Four patients were
and specificity. Additionally, the electrophysiologists were asked to rate mistakenly identified as being in AF on ACK lead I analysis while dis-
the ACK recording as either adequate or inadequate in terms of the playing sinus rhythm in the 12-lead ECG. Of these patients, two were in
overall quality and interpretability (subjective signal-to-noise ratio sinus rhythm with continuous ventricular stimulation and two patients
etc.). had a left bundle branch block (LBBB).

73

Downloaded for Jan Svendsen ([email protected]) at Southwest Jutland Hospital Esbjerg from ClinicalKey.com by Elsevier on February 28, 2020.
For personal use only. No other uses without permission. Copyright ©2020. Elsevier Inc. All rights reserved.
F.K. Wegner, et al. European Journal of Internal Medicine 73 (2020) 72–75

Table 2
App algorithm diagnoses of the ACK ECG recordings compared to the gold-
standard 12-lead ECG diagnosis. AF= atrial fibrillation, Afl= atrial flutter.
Table 2a 12 lead ECG
sinus rhythm AF/Afl
(n = 65) (n = 27)

App algorithm diagnosis of Sinus rhythm 45 0

lead I Possible AF 8 19
unclassified 11 4
no analysis 1 4

Table 2b 12-lead ECG

sinus rhythm AF/Afl
(n = 65) (n = 27)

App algorithm diagnosis of sinus rhythm 39 0

novel parasternal lead Possible AF 3 15
unclassified 21 7
Fig.. 2. ECGrecordings of a patient with a critical amount of artifact in the no analysis 2 5
AliveCor lead I showing no artifact in the novel parasternal lead (NPL).

Table 1 lead I and the NPL.

Electrophysiologists‘ diagnoses of the ACK ECG recordings compared to the
gold-standard 12-lead ECG diagnosis. EP= electrophysiologists, AF= atrial 4.1. Atrial flutter
fibrillation, Afl= atrial flutter.
Table 1a 12-lead ECG Five patients exhibited atrial flutter at the time of ECG recording.
sinus rhythm AF/Afl The electrophysiologists displayed an accuracy of 100% (5 of 5 ECGs
(n = 65) (n = 27)
correctly interpreted) in detecting atrial flutter on an ACK lead I re-
EP diagnosis of lead I sinus 61 0 cording. In the NPL recording, the electrophysiologists showed a 90%
rhythm accuracy: 4 of 5 ECGs were correctly interpreted, while there was dis-
AF/Afl 4 27 agreement in 1 recording (electrophysiologist 1: atrial flutter, electro-
physiologist 2: sinus rhythm). The app algorithm diagnosed 4 of 10 ECG
Sensitivity 100%
recordings with “possible AF” while diagnosing a further 4 of 10 ECGs
Specificity 94% with “unclassified”. 2 ECGs received a connotation of “no analysis”. The
app algorithm therefore displayed a sensitivity of 40% in the diagnosis
Table 1b 12-lead ECG of atrial flutter.
sinus rhythm AF/Afl
(n = 65) (n = 27)
4.2. Pacemaker stimulation
EP diagnosis of novel sinus 63 1
parasternal lead rhythm
Out of 99 patients included in the present study, 12 patients had
AF/Afl 2 26
continuous ventricular pacemaker stimulation at the time of ECG re-
Sensitivity 96% cording. Six of these 12 patients were previously excluded from the
statistical analysis because of the difficulty of diagnosing AF when reg-
Specificity 97% ular ventricular pacemaker stimulation is present. In these 6 patients
with AF and ventricular stimulation on 12-lead ECG, the electro-
physiologists and the app algorithm misidentified all ACK ECG record-
In the NPL, the electrophysiologists displayed a sensitivity of 96%, a ings as sinus rhythm. In the remaining 6 patients with sinus rhythm and
specificity of 97% and a κ of 0.92. One patient with atrial flutter was ventricular stimulation, the electrophysiologists had an accuracy of 75%
misidentified as being in sinus rhythm. Two patients in sinus rhythm in the lead I recordings and of 83% in the NPL. The app algorithm had an
were described as being in AF. Of these, one was in sinus rhythm with accuracy of 50% in the lead I and an accuracy of 17% in the NPL.
continuous ventricular stimulation and one patient displayed a LBBB.
4.3. The novel parasternal lead as a secondary recording vector
4. App algorithm diagnosis
To evaluate a potential benefit of the addition of the NPL to the
A cross-tabulation of the app algorithm diagnoses of the respective recording of a lead I ACK ECG, we decided to separately analyze the
ACK recordings with the gold-standard 12-lead ECG diagnoses is shown NPL recordings of those patients whose lead I ACK recordings were
in Table 2. As mentioned above, a goal of the present study was to ex- interpreted by the app algorithm as either “unclassified” or “no ana-
clude as few patients as possible from a sensitivity and specificity ana- lysis”. Table 3 displays the electrophysiologists’ and the app algorithm
lysis of the algorithm. We therefore calculated sensitivity and specificity diagnosis of the NPL recordings in these 20 patients.
for the app algorithm without excluding the groups of patients receiving The electrophysiologists’ sensitivity in these NPL recordings was
a diagnosis of “unclassified” or “no analysis” by the algorithm. 100% and the specificity 92% with an accuracy of κ = 0.90. The app
The app algorithm thus had a sensitivity of 70% and a specificity of algorithm diagnoses displayed a sensitivity of 38% and a specificity of
69% in the lead I recordings with a κ of 0.53. In the NPL, the algorithm 17% with a κ of 0.20.
had a sensitivity of 55% and a specificity of 60% with a κ of 0.40. All of
the patients with ECG recordings that the app algorithm defined as 4.4. Predictors of correct diagnoses
“normal” were indeed in sinus rhythm, thus leading to a negative
predictive value of 100% in the present patient population for both the To analyze for possible predictors of correct rhythm diagnoses, we

74

Downloaded for Jan Svendsen ([email protected]) at Southwest Jutland Hospital Esbjerg from ClinicalKey.com by Elsevier on February 28, 2020.
For personal use only. No other uses without permission. Copyright ©2020. Elsevier Inc. All rights reserved.
F.K. Wegner, et al. European Journal of Internal Medicine 73 (2020) 72–75

Table 3 should be sought.

Electrophysiologists‘ and app algorithm diagnoses of the novel parasternal lead The analysis by electrophysiologists had a very high sensitivity
recordings of 20 patients in which the app algorithm diagnosis of the lead I (96–100%) and specificity (94–97%) for the presence of AF or atrial
recording was either “unclassified” or “no diagnosis”. EP= electrophysiologists, flutter with an excellent agreement with the respective 12-lead ECG
AF= atrial fibrillation, Afl= atrial flutter.
diagnoses. In contrast to the ACK app algorithm, the electro-
Table 3a 12-lead ECG physiologists were able to detect atrial flutter and ventricular stimula-
sinus rhythm AF/Afl(n = 8) tion with sinus rhythm with a high accuracy. No patient-specific pre-
(n = 12)
dictors of inaccurate diagnoses or a large amount of artifact could be
EP diagnosis of novel Sinus rhythm 11 0 found.
parasternal lead Possible AF 1 8
6. Limitations
Table 3b 12-lead ECG
Sinus rhythm AF/Afl
(n = 12) (n = 8)
The present study was conducted as a single center study with its
inherent limitations. Although patients were recruited from an in-
App algorithm diagnosis Sinus rhythm 2 0 patient unit at a large tertiary center, the prevalence of AF recordings
of novel parasternal Possible AF 0 3 remained relatively low. However, we believe that a prevalence of ar-
lead Unclassified 9 1
rhythmias of 29% represent a sufficient number to draw conclusions
No analysis 1 4
from with regard to diagnostic accuracy. Apart from that, our patient
group is the largest with a systematic and blinded ECG analysis of ACK
recordings compared to the gold-standard 12-lead ECG.
compared the mean age, gender and other demographic variables.
There were no statistically significant predictors identifiable (all 7. Conclusion
p > 0.05).
The AliveCor Kardia is a smartphone-based one-lead ECG recorder
5. Discussion allowing for an accurate rhythm diagnosis by an interpreting cardiol-
ogist both in the manufacturer-recommended lead I and a novel para-
The present study is the first investigator-initiated, unsponsored sternal lead. The diagnostic app algorithm has a very high negative
report to analyze the diagnostic value of multiple recording vectors of predictive value. In cases where the app does not show an interpreta-
the AliveCor Kardia ECG recorder in a large “all-comer” inpatient co- tion of “normal rhythm”, the underlying recordings should be inter-
hort. We were able to demonstrate that the ACK is able to record high preted by a cardiologist.
quality one-lead ECG whose interpretation by experienced electro-
physiologists can reliably determine underlying rhythm in both the Decleration of Competing Interest
manufacturer-recommended lead I and a novel parasternal lead.
We introduced the NPL by placing the ACK in a left parasternal
None.
position, as one in five of all lead I ACK recordings contained a critical
amount of artifact. The novel parasternal lead was universally applic-
References
able, as a recording could be obtained in all included patients. Our
results show that the NPL yields comparable results to the standard ACK [1] Lin HJ, Wolf PA, Kelly-Hayes M, et al. Stroke severity in atrial fibrillation. The
vector. Furthermore, the NPL maintains its diagnostic accuracy when Framingham Study. Stroke 1996;27(10):1760–4. PubMed PMID: 8841325.
interpreted by an electrophysiologist in cases where the app algorithm [2] Lozano R, Naghavi M, Foreman K, et al. Global and regional mortality from 235 causes of
death for 20 age groups in 1990 and 2010: a systematic analysis for the global burden of
diagnosis of the lead I is ambiguous (see Table 3a). A possible everyday disease study 2010. Lancet 2012;380(9859):2095–128. https://doi.org/10.1016/S0140-
clinical approach might thus include instructing a patient to 1) always 6736(12)61728-0. PubMed PMID:23245604.
[3] Kirchhof P, Benussi S, Kotecha D, et al. ESC Guidelines for the management of atrial
record a lead I ACK ECG during symptomatic episodes, 2) record a NPL fibrillation developed in collaboration with EACTS. Eur Heart J 2016;37(38):2893–962.
ECG when the immediately available app algorithm diagnosis of the https://doi.org/10.1093/eurheartj/ehw210. PubMed PMID:27567408.
[4] Turakhia MP, Desai M, Hedlin H, et al. Rationale and design of a large-scale, app-based
lead I ECG recorded in step 1 shows a diagnosis of “unclassified” or “no study to identify cardiac arrhythmias using a smartwatch: the Apple Heart Study. Am
analysis” and 3) making both recordings available to the treating Heart J 2019;207:66–75. https://doi.org/10.1016/j.ahj.2018.09.002. PubMed
physician for interpretation. PMID:30392584.
[5] Freedman B, Camm J, Calkins H, et al. Screening for atrial fibrillation: a report of the AF-
We detected a significantly lower sensitivity (55–70%) and speci- screen international collaboration. Circulation 2017;135(19):1851–67. https://doi.org/
ficity (60–69%) of the ACK app algorithm for the diagnosis of AF or 10.1161/CIRCULATIONAHA.116.026693. PubMed PMID:28483832.
[6] Mairesse GH, Moran P, Van Gelder IC, et al. Screening for atrial fibrillation: a European
flutter than previous working groups [8,10,11]. This can probably be Heart Rhythm Association (EHRA) consensus document endorsed by the Heart Rhythm
largely attributed to the absence of exclusion criteria and our decision Society (HRS), Asia Pacific Heart Rhythm Society (APHRS), and Sociedad
Latinoamericana De Estimulacion Cardiaca y Electrofisiologia (SOLAECE). Europace
to include patients with “unclassified” and “no analysis” diagnoses by
2017;19(10):1589–623. https://doi.org/10.1093/europace/eux177. PubMed
the algorithm. While William et al. reported a sensitivity of 97% and a PMID:29048522.
specificity of 94% for the app algorithm in 52 patients, this was mainly [7] US Food and Drug Administration. 510(k) Premarket Notification 2014 [updated 29.04.
201906.05.2019]. Available from: https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/
due to the authors’ decision to exclude 28% of all ACK recordings, cfpmn/pmn.cfm?ID=K140933.
which were annotated as “unclassified” [8]. In our opinion, excluding a [8] William AD, Kanbour M, Callahan T, et al. Assessing the accuracy of an automated atrial
fibrillation detection algorithm using smartphone technology: the iREAD study. Heart
large amount of patients who use a population-based screening tool Rhythm 2018;15(10):1561–5. https://doi.org/10.1016/j.hrthm.2018.06.037. PubMed
misses its target. In addition, other previous studies applying the ACK to PMID:30143448.
larger patient populations [10,11] did not systematically compare the [9] Viera AJ, Garrett JM. Understanding interobserver agreement: the kappa statistic. Fam
Med 2005;37(5):360–3. PubMed PMID:15883903.
recordings to the gold-standard 12-lead ECG, thereby possibly dis- [10] Lowres N, Neubeck L, Salkeld G, et al. Feasibility and cost-effectiveness of stroke pre-
torting their analyses. vention through community screening for atrial fibrillation using iPhone ECG in phar-
macies. The SEARCH-AF study. Thromb Haemost 2014;111(6):1167–76. https://doi.org/
In contrast to previous publications, our data systematically com- 10.1160/TH14-03-0231. PubMed PMID:24687081.
paring the ACK to the 12-lead ECG indicate that the main utility of the [11] Chan PH, Wong CK, Pun L, et al. Head-to-Head comparison of the AliveCor heart monitor
and microlife watchbp office AFIB for atrial fibrillation screening in a primary care set-
app algorithm itself is its high negative predictive value. In addition,
ting. Circulation 2017;135(1):110–2. https://doi.org/10.1161/CIRCULATIONAHA.116.
when the app algorithm displays a diagnosis of “unclassified” or “no 024439. PubMed PMID:28028066.
analysis”, further analysis by an experienced medical professional

75

Downloaded for Jan Svendsen ([email protected]) at Southwest Jutland Hospital Esbjerg from ClinicalKey.com by Elsevier on February 28, 2020.
For personal use only. No other uses without permission. Copyright ©2020. Elsevier Inc. All rights reserved.