Automated Atrial Fibrillation Detection Algorithm Using ... · T D ACCEPTED MANUSCRIPT 1 Automated...
Transcript of Automated Atrial Fibrillation Detection Algorithm Using ... · T D ACCEPTED MANUSCRIPT 1 Automated...
Accepted Manuscript
Automated Atrial Fibrillation Detection Algorithm Using Smartwatch Technology
Joseph M. Bumgarner, MD, Cameron T. Lambert, MD, Ayman A. Hussein, MD,Daniel J. Cantillon, MD, Bryan Baranowski, MD, Kathy Wolski, MPH, Bruce D.Lindsay, MD, Oussama M. Wazni, MD MBA, Khaldoun G. Tarakji, MD MPH
PII: S0735-1097(18)33486-7
DOI: 10.1016/j.jacc.2018.03.003
Reference: JAC 24734
To appear in: Journal of the American College of Cardiology
Received Date: 14 February 2018
Revised Date: 1 March 2018
Accepted Date: 2 March 2018
Please cite this article as: Bumgarner JM, Lambert CT, Hussein AA, Cantillon DJ, Baranowski B,Wolski K, Lindsay BD, Wazni OM, Tarakji KG, Automated Atrial Fibrillation Detection AlgorithmUsing Smartwatch Technology, Journal of the American College of Cardiology (2018), doi: 10.1016/j.jacc.2018.03.003.
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service toour customers we are providing this early version of the manuscript. The manuscript will undergocopyediting, typesetting, and review of the resulting proof before it is published in its final form. Pleasenote that during the production process errors may be discovered which could affect the content, and alllegal disclaimers that apply to the journal pertain.
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
1
Automated Atrial Fibrillation Detection Algorithm U sing Smartwatch Technology Joseph M. Bumgarner MDa, Cameron T. Lambert MDa, Ayman A. Hussein MDa, Daniel J. Cantillon MDa, Bryan Baranowski MDa, Kathy Wolski MPHb, Bruce D. Lindsay MDa, Oussama M. Wazni MD MBAa, Khaldoun G. Tarakji MD MPHa From the aDepartment of Cardiovascular Medicine and bthe Cleveland Clinic Coordinating Center for Clinical Research (C5Research), Cleveland Clinic, Cleveland Ohio Funding: AliveCor (AliveCor, Mountain View, CA) provided the Kardia Band monitors which were connected to an Apple Watch and paired via Bluetooth to a smartphone device for utilization in the study. AliveCor was not involved in the design, implementation, data analysis, or manuscript preparation of the study. Brief Title: Smartwatch CV Study Disclosures: JMB: None CTL: None AAH: Consulting Abbott, Biosense Webster DJC: Consulting Abbott, Boston Scientific, Stryker Sustainability, LifeWatch BB: None KW: None BDL: None OMW: Speaker honorarium, Spectranetics KGT: Medical advisory board for Medtronic, AliveCor Corresponding Author: Khaldoun G. Tarakji, MD MPH Section of Cardiac Pacing and Electrophysiology Heart and Vascular Institute Cleveland Clinic 9500 Euclid Avenue, J2-2 Cleveland, Ohio 44195 Telephone: (216) 445-9225 Fax: (216) 445-6149 E-mail: [email protected] Twitter: @khaldountarakji
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
2
ABSTRACT Background: The Kardia Band (KB) is a novel technology that enables patients to record a rhythm strip using an Apple smartwatch. The band is paired with an app providing automated detection of atrial fibrillation (AF). Objectives: To examine whether the KB could accurately differentiate sinus rhythm (SR) from AF compared to physician-interpreted 12-lead ECGs and KB recordings. Methods: Consecutive patients with AF presenting for cardioversion (CV) were enrolled. Patients underwent pre-CV ECG along with a KB recording. If CV performed, a post-CV ECG was obtained along with a KB recording. The KB interpretations were compared to physician-reviewed ECGs. The KB recordings were reviewed by blinded electrophysiologists and compared to ECG interpretations. Sensitivity, specificity and K coefficient were measured. Results: One hundred patients were enrolled (Age 68 ± 11 years). Eight patients did not undergo CV. There were 169 simultaneous ECG and KB recordings. Fifty-seven were non-interpretable by the KB. Compared to ECG, the KB interpreted AF with 93% sensitivity, 84% specificity and K coefficient 0.77. Physician-interpretation of KB recordings demonstrated 99% sensitivity, 83% specificity and K coefficient 0.83. Of 57 non-interpretable KB recordings, interpreting electrophysiologists diagnosed AF with 100% sensitivity, 80% specificity and K coefficient 0.74. Among 113 cases where KB and physician readings of the same recording were interpretable, agreement was excellent (K coefficient 0.88). Conclusions: The KB algorithm for AF detection, supported by physician review can accurately differentiate AF from SR. This technology can help screen patients prior to elective CV and avoid unnecessary procedures. CONDENSED ABSTRACT: The KB is a novel technology recently cleared by the FDA which allows patients to record a rhythm strip instantaneously interpreted by an algorithm as AF or SR. Our study is the first investigation to assess the accuracy and clinical utility of this device in a clinical setting. In a blinded and prospective manner, we specifically compared automated KB algorithm recordings to both simultaneous ECG tracings and physician-interpreted KB recordings and found that the KB algorithm for AF detection, when supported by physician review, can accurately differentiate AF from SR. Key Words: Smartwatch, ECG monitoring, atrial fibrillation, digital health, cardioversion Abbreviations: KB – Kardia Band AF – Atrial Fibrillation SR – Sinus Rhythm ECG - Electrocardiogram CV – Cardioversion BPM – Beats per minute ILR – Implantable Loop Recorder
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
3
Introduction
Atrial fibrillation (AF) is the most commonly encountered arrhythmia in clinical practice
and population-based studies forecast over 6 million individuals living with this diagnosis by
2050 (1,2). It is a chronic condition whose prevalence increases with age, and represents a
growing economic burden for our healthcare system (3, 4). While the journey of AF begins with
an initial diagnosis, its management is long term, nuanced, and often involves hospital-based
interventions along the way including electrical cardioversion (CV).
Recently, commercially available handheld cardiac rhythm recorders have been
developed that can record a rhythm strip using smartphone technology (5). In November 2017
the Kardia Band (KB) (AliveCor, Mountain View, CA) was introduced as the first FDA cleared
Apple Watch accessory that allows a patient to record a rhythm strip equivalent to lead I for 30
seconds. The KB is coupled with an application that provides an instantaneous and automatic
rhythm adjudication algorithm for the diagnosis of AF. The application can inform the patient
when AF is detected and transmit these results to the patient’s caring physician instantaneously.
The primary objective of our study was to examine whether the KB and AF detection
algorithm could accurately and reliably differentiate sinus rhythm (SR) from AF when compared
to physician interpreted 12-lead ECGs and KB recordings in patients with known AF presenting
to a high volume hospital-based electrophysiology practice for scheduled electrical CV.
Methods
Study design
This was a prospective, non-randomized and adjudicator-blinded study completed at a
tertiary care hospital-based electrical CV lab designed to evaluate the accuracy of the KB
automated algorithm for the detection of AF. AliveCor provided the KB connected to an Apple
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
4
Watch which was paired via Bluetooth to a smartphone device (Apple Inc., Cupertino, CA) for
utilization in the study. AliveCor was not involved in the design, implementation, data analysis,
or manuscript preparation of the study. The study was approved by the Cleveland Clinic
Institutional Review Board.
Study participants
Consecutive patients with a diagnosis of AF who presented for scheduled elective CV
with or without a planned transesophageal echocardiogram were screened for enrollment.
Inclusion criteria included all adult patients age 18 to 90 years old who were able to provide
informed consent and willing to wear the KB before and after CV. We excluded all patients with
an implanted pacemaker or defibrillator.
Once enrolled, patients underwent a pre CV ECG followed immediately by KB
recording. These paired recordings were considered simultaneous. If the CV was performed a
post CV ECG was then obtained along with another KB recording. The KB tracing was
automatically analyzed using the KB algorithm. This algorithm measures rhythm irregularity and
P wave absence in real time to classify the rhythm strip as “possible AF”. If the criteria for AF is
not met the KB algorithm classifies regular rhythms with P waves as “normal” if the rate is
between 50-100 bpm or “unclassified” for those rhythms with rates below 50 bpm or over 100
bpm or if the recording is noisy or shorter than 30 seconds. The KB rhythm strips were
automatically transferred to the secure AliveCor server, downloaded, and printed for review.
All automated KB rhythm strips and ECGs were anonymized and distributed to two
blinded electrophysiologists who independently interpreted each tracing and assigned a diagnosis
of SR, AF or atrial flutter, or unclassified. If the two electrophysiologists disagreed on the
diagnosis a third electrophysiologist reviewed the tracing and assigned a final diagnosis. In order
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
5
to assess the accuracy of the KB algorithm at appropriately identifying AF, the automated KB
interpretations were compared to both the physician interpreted KB rhythm strips and the
physician reviewed simultaneous ECGs.
Statistical analysis
Sensitivity and specificity were calculated for KB automated interpretation compared to
physician interpreted 12-lead ECG, for physician interpreted KB rhythm strip compared to
physician interpreted 12-lead ECG, and for KB automated interpretation compared to physician
interpreted KB recordings. Kappa (κ) coefficients for inter-observer agreement were assessed. κ
coefficients of >0.8 were considered to represent excellent agreement. AF and atrial flutter were
considered as a single disease state for all interpretations.
Results
One hundred patients were enrolled in the study from March 2017 through June 2017.
Demographics and clinical characteristics are summarized in Table 1. CV was performed in
85% of study participants. Of the 15 patients who did not undergo CV, 8 were cancelled due to
presentation in SR. There were 169 simultaneous 12-lead ECG and KB recordings obtained from
study participants, and 57 KB recordings were determined as unclassified by the KB algorithm.
Of the 57 unclassified KB tracings, 16 (28%) were due to baseline artifact and low amplitude of
the recording, 12 (21%) were due to a recording of less than 30 seconds in duration, 6 (10%)
were due to a heart rate of less than 50 bpm, 5 (9%) were due to a heart rate of greater than 100
bpm, and the remaining 18 (32%) were unclassified due to an unclear reason. Electrophysiologist
interpreted 12-lead ECGs were all interpretable.
In order to test the ability of the KB algorithm to detect AF, automated KB rhythm
interpretations and electrophysiologist interpreted 12-lead ECGs were compared. Among the
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
6
recordings where the KB provided a diagnosis, it correctly diagnosed AF with 93% sensitivity,
84% specificity and a K coefficient of 0.77 (95% confidence interval 0.65-0.89) when compared
to the electrophysiologist interpreted 12-lead ECG (Table 2). Because our analysis used multiple
observations from the same individual, we evaluated for possible intra-individual correlations by
comparing only pre CV KB recordings to electrophysiologist interpreted 12-lead ECGs and
found the performance of the KB algorithm to be unchanged (Online Table 1).
In order to determine whether the automated KB recordings labeled as “unclassified” by
the algorithm were still clinically useful, these tracings were interpreted by our blinded
electrophysiologists and compared to the electrophysiologist interpreted 12-lead ECGs. Of the 57
automated unclassified KB recordings, the interpreting electrophysiologists were able to
correctly diagnose AF with 100% sensitivity, 80% specificity and a K coefficient of 0.74 (Table
3).
In order to assess the fidelity and overall quality of the KB tracings produced by the
smartwatch, electrophysiologist interpreted KB recordings were compared to corresponding 12-
lead ECG tracings. Twenty-two recordings were determined to be non-interpretable by the
reading electrophysiologist, and these were predominately due to baseline artifact. Of the
remaining 147 simultaneous recordings, the electrophysiologist interpreted 12-lead ECGs and
electrophysiologist interpreted KB recordings, physician interpretation of the KB tracings
demonstrated 99% sensitivity, 83% specificity and a K coefficient of 0.83 (Table 4).
Additionally, in order to measure the quality of the KB recordings, we compared the KB
automated algorithm interpretation to physician interpretation of the same recordings. Of the
cases where both methods were interpretable, the KB automated algorithm was 93% sensitive
and 97% specific in detecting AF with a K coefficient of 0.88 (Table 5).
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
7
Discussion
The era of mobile healthcare technology has proliferated over the past decade.
Consumers from the general public now have direct access to devices and applications which
offer real-time measurements of cardiovascular physiology, and some technologies extrapolate
this data in order to provide diagnostic information (6). It is estimated that by 2019 annual sales
of such devices will reach 50 billion dollars worldwide (7). However, the ability of some devices
to accurately measure biometric endpoints has been questioned, and some mobile health
technologies are available without verification through rigorous clinical studies (8).
Alongside the growth of mobile healthcare technology has been the desire of many
physicians and patients to accurately monitor disease-related metrics of chronic conditions in the
ambulatory setting. AF is a good example of a relapsing condition that requires frequent
monitoring of clinical endpoints in order to assess the efficacy of treatment choices and plan
future interventions. The KB is the first smartwatch accessory cleared by the FDA and available
to the general public without a prescription which claims to instantaneously detect AF and
transmit this information to a patient’s treating physician.
In this study, we aimed to assess whether the KB and AF detection algorithm could
accurately and reliably differentiate SR from AF in patients with known AF presenting for
scheduled electrical CV (Central Illustration ). We compared automated KB interpretations to
simultaneously recorded ECGs read by blinded electrophysiologists and found very good
agreement between the two. When able to provide an interpretation, the automated KB readings
correctly identified AF with 93% sensitivity and 84% specificity (Figure 2). Of the 169 total KB
recordings, 57 (33.7%) were interpreted as unclassified by the automated KB algorithm. Reasons
that these recordings were deemed non-interpretable included short recordings less than 30
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
8
seconds, low amplitude P waves, and baseline artifact. For those recordings where the automatic
KB tracing was non-interpretable, direct physician interpretation could be used to correctly
identify AF with 100% sensitivity and 80% specificity (Figure 3). In general, the KB recordings
when interpreted by the physician had excellent agreement with simultaneous 12 ECG
interpretation with 99% sensitivity and 83% specificity.
Prior to the development of the KB smartwatch algorithm, several algorithms used by
implantable loop recorders (ILRs) were validated for the detection of AF. Currently available
ILRs detect AF by sensing R waves and applying a variety of regularity algorithms to detect AF.
The Confirm DM2101 (Abbott, Chicago, Illinois) detects RR interval regularity and measures
suddenness of an irregular rhythm’s onset and offset to diagnose AF using two probabilistic
scoring models. The BioMonitor (Biotronik, Berlin, Germany) also measures R wave variability
and allows the clinician to adjust the number of cycle lengths used and the confirmation time
needed to detect AF. The most studied of the ILRs is the Reveal LINQ (Medtronic, Minneapolis,
MN) system whose algorithm for AF detection uses both R wave irregularity and a
programmable P wave evidence discrimination tool that can be modified based on the individual
needs of a given patient (9-11). The Reveal LINQ system was evaluated in the XPECT trial. In
this study the sensitivity and specificity for identifying patients with any AF was 96.1% and
85.4% respectively (12). In our study, the accuracy of the KB algorithm for the detection of AF
was comparable to these results.
Wearable devices like the KB require a safe and durable platform upon which recordings
can be reviewed and stored. A secure cloud-based platform has been developed to view and
download KB recordings. The applicability of this platform to the outpatient management of
patients with AF needs to be evaluated and studied in future trials. Our study also demonstrated
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
9
that a subset of patients (8%) who presented for CV was found to be in SR. For each of these
patients, the automated KB algorithm did not erroneously identify AF, and the physician
interpretation of the KB recording correctly confirmed SR in each case. While this study was not
powered to assess the financial consequences of cancelled CVs, it is reasonable to conclude that
a measurable number of resources were forfeited by both the patient and the healthcare system in
anticipation of a procedure that was ultimately deemed unnecessary once SR was confirmed. As
data from the KB can be reviewed remotely, the resources used in preparation of these patients’
cancelled CVs could have been saved. The KB system has been previously shown to be cost
effective for AF screening. Our study suggests potential use of these products to provide more
effective healthcare delivery (13).
Limitations
This was a single center study at a tertiary referral center with a small sample size. The
population represented in this study had a known history of AF and a sufficient burden of AF to
prompt electrical CV. The performance of the KB smartwatch algorithm may be more variable in
a population with a lower AF burden. We did not evaluate socioeconomic status in our study and
only 17% of our enrolled patients were female. Additionally, none of the patients who
participated in our study had previously used the KB. These facts may limit the generalizability
of our findings in the general public, and future studies should consider measuring these
variables. Patients with cardiac implantable electronic devices were excluded from this study,
and further evaluation of the KB algorithm is needed in this patient population. Participants were
instructed on how to use the KB wristband while seated in a hospital bed immediately prior to
obtaining each recording. Their ability to record each tracing was directly observed. As a result,
the performance of the KB algorithm and the clarity of the recorded tracings may be less
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
10
accurate in an outpatient or ambulatory setting. For the same reason, some of the unclassified
recordings could have been avoided with more patient practice on the proper use of the KB
device. Additionally, the KB prototype used in our study did not display a real-time ECG tracing
on the watch screen at the time of recording. Since FDA clearance, the KB app is now permitted
to display this information. We anticipate the real-time display of the ECG recording will
improve quality of the recordings obtained by users of the device.
Conclusions
The KB smartwatch automated algorithm for AF detection, supported by physician
review of these recordings, can reliably differentiate AF from SR. Avoiding scheduling
unnecessary electrical CVs is one example of a clinical application of the KB system. Many
other potential applications warrant further investigation and might transform our longitudinal
care of AF patients.
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
11
CLINICAL PERSPECTIVES:
Competency in Patient Care: Among patients with AF who are being considered for elective
CV, the KB smartwatch algorithm can be used with physician oversight to accurately
differentiate between SR and AF.
Translational Outlook: As the prevalence of AF continues to rise within the era of expanding
access to mobile healthcare technology, randomized controlled trials will be needed to further
validate the clinical and financial risks and benefits of mobile healthcare devices.
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
12
REFERENCES
1. Coyne KS, Paramore C, Grandy S, et al., Assessing the direct costs of treating
nonvalvular atrial fibrillation in the United States. Value Health, 2006;9(5):348-56.
2. January CT, Wann LS, Alpert JS, et al., 2014 AHA/ACC/HRS guideline for the
management of patients with atrial fibrillation: a report of the American College of
Cardiology/American Heart Association Task Force on Practice Guidelines and the Heart
Rhythm Society. J Am Coll Cardiol, 2014;64(21):e1-76.
3. Becker C, Cost-of-illness studies of atrial fibrillation: methodological considerations.
Expert Rev Pharmacoecon Outcomes Res, 2014;14(5):661-84.
4. Wodchis WP, Bhatia RS, Leblanc K, et al., A review of the cost of atrial fibrillation.
Value Health, 2012;15(2):240-8.
5. Tarakji KG, Wazni OM, Callahan T, et al., Using a novel wireless system for monitoring
patients after the atrial fibrillation ablation procedure: the iTransmit study. Heart Rhythm,
2015;12(3):554-559.
6. Freedman B, Screening for Atrial Fibrillation Using a Smartphone: Is There an App for
That? J Am Heart Assoc, 2016;5(7).
7. Piwek L, Ellis DA, Andrews S, et al., The Rise of Consumer Health Wearables: Promises
and Barriers. PLoS Med, 2016;13(2):e1001953.
8. Gillinov S, Etiwy M, Wang R, et al. Variable Accuracy of Wearable Heart Rate Monitors
during Aerobic Exercise. Med Sci Sports Exerc, 2017;49(8):1697-1703.
9. Lee R, Mittal S, Utility and limitations of long-term monitoring of atrial fibrillation using
an implantable loop recorder. Heart Rhythm. 2018;15(2):287-95.
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
13
10. Passman RS, Rogers JD, Sarkar S, et al., Development and validation of a dual sensing
scheme to improve accuracy of bradycardia and pause detection in an insertable cardiac
monitor. Heart Rhythm. 2017;14(7):1016-23.
11. Mittal S, Rogers J, Sarkar S, et al., Real-world performance of an enhanced atrial
fibrillation detection algorithm in an insertable cardiac monitor. Heart Rhythm.
2016;13(8):1624-30.
12. Hindricks G, Pokushalov E, Urban L, et al., Performance of a new leadless implantable
cardiac monitor in detecting and quantifying atrial fibrillation: Results of the XPECT
trial. Circ Arrhythm Electrophysiol, 2010;3(2):141-7.
13. Lowres N, Neubeck L, Salkeld G, et al., Feasibility and cost-effectiveness of stroke
prevention through community screening for atrial fibrillation using iPhone ECG in
pharmacies. The SEARCH-AF study. Thromb Haemost, 2014;111(6):1167-76.
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
14
FIGURE LEGENDS
Central Illustration: Assessment of the Accuracy of the Kardia Band Smartwatch
Algorithm for AF Detection Compared to 12-Lead ECG in Patients Undergoing
Cardioversion. Left Side of Figure: Automated KB recordings are compared to physician-
interpreted 12-lead ECGs and detect AF with 93% sensitivity, 84% specificity. Physician-
reviewed unclassified automated KB recordings are compared to physician-interpreted 12-lead
ECGs and detect AF with 100% sensitivity, 80% specificity. Right Side of Figure: Physician-
interpreted KB recordings are compared to physician-interpreted 12-lead ECGs and detect AF
with 99% sensitivity and 83% specificity. Twenty-two physician-interpreted KB recordings were
non-interpretable. AF: Atrial Fibrillation, ECG: Electrocardiogram, KB: Kardia Band.
Figure 1: The Kardia Band from AliveCor Paired with an Apple Smartwatch.
AliveCor (AliveCor, Mountain View, California) Kardia Band monitors were connected to an
Apple smartwatch and paired via Bluetooth to a smartphone device for utilization in the study.
Figure 2: Correct KB Interpretations Compared to Simultaneous ECG
A. Simultaneous recordings of SR using KB (left) and 12-lead ECG (right). The KB
automated algorithm identifies SR for this sample.
B. Simultaneous recordings of AF using KB (left) and 12-lead ECG (right). The KB
automated algorithm identifies AF for this sample.
Figure 3: Incorrect KB Interpretations Compared to Simultaneous ECG
A. Heart rhythm recording defined by KB (left) as unclassified with simultaneous 12-lead
ECG (right) interpreted as SR
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
15
B. Heart rhythm recording defined by KB (left) as unclassified with simultaneous 12-lead
ECG (right) interpreted as AF.
C. Heart rhythm recording defined by KB (left) as too short to analyze with simultaneous
12-lead ECG (right) interpreted as AF.
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
16
Table 1: Demographics, Clinical, and Procedural Characteristics of Enrolled Patients
PARAMETER All Patients
Total Number of Patients
(n) 100
Age (average ± SD) 68.2 ± 10.86
Gender (n, %)
Female 17 (17.0)
Anticoagulant (n, %)
Warfarin (Coumadin) 32 (32.0)
Dabigatran (Pradaxa) 2 (2.0)
Rivaroxaban (Xarelto) 19 (19.0)
Apixaban (Eliquis) 47 (47.0)
TEE Performed (n, %)
Yes, scheduled 21 (21.0)
Yes, added on 2 (2.0)
TEE Finding (n,%)
No Thrombus 20 (20.2)
Sludge 1 (1.0)
Thrombus 2 (2.0)
CV Performed
Yes 85 (85.0)
No 15 (15.0)
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
17
Reason if No CV
Performed
Sub-therapeutic INR 4 (26.7)
Found to be in NSR 8 (53.3)
Thrombus on TEE 2 (13.3)
Hypotension during TEE 1 (6.7)
CV Outcome
Successful 78 (91.7)
Transient 3 (3.5)
Failed 4 (4.7)
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
18
Table 2 – KB Algorithm Reading Compared to Electrophysiologist Interpreted 12-Lead
ECG
Electrophysiologist Interpreted 12-lead ECG
KB Algorithm
Interpretation AF/Flutter SR
Non-
interpretable Total
AF/Flutter 63 7 0 42
SR 5 37 0 70
Missing/Unclassified 23 34 0 57
Total 91 78 0 169
Sensitivity, specificity and κ coefficient are calculated only for the
simultaneous transmission with interpretation (in bold). Sensitivity 93%
(63/68), (95% confidence interval 0.86-0.99), specificity 84% (37/44), (95%
confidence interval 0.73-0.95).
κ coefficient of 0.77 (95% confidence interval 0.65-0.89) for numbers in bold.
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
19
Table 3 – Unclassified KB Readings when Read by Electrophysiologist Compared to
Electrophysiologist Interpreted 12-Lead ECG
Electrophysiologist Interpreted 12-lead ECG
Electrophysiologist
Interpreted KB Reading AF/Aflutter SR
Non-
interpretable Total
AF/Flutter 14 5 0 19
SR 0 20 0 20
Missing/Non-interpretable 9 9 0 18
Total 23 34 0 57
Sensitivity, specificity and κ coefficient are calculated only for the simultaneous
transmission with interpretation (in bold). Sensitivity 100% (14/14), (95% confidence
interval 0.77-1.0), specificity 80% (20/25), (95% confidence interval 0.64-0.96).
κ coefficient of 0.74 (95% confidence interval 0.54-0.95) for numbers in bold.
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
20
Table 4 – Electrophysiologist Interpreted KB Reading Compared to Electrophysiologist
Interpreted 12-Lead ECG
Electrophysiologist Interpreted 12-lead ECG
Electrophysiologist
Interpreted KB Reading AF/Aflutter SR
Non-
interpretable Total
AF/Flutter 80 11 0 91
SR 1 55 0 56
Missing/Non-interpretable 10 12 0 22
Total 91 78 0 169
Sensitivity, specificity and κ coefficient are calculated only for the simultaneous
transmission with interpretation (in bold). Sensitivity 99% (80/81), (95% confidence
interval 0.96-1.00), specificity 83% (55/66), (95% confidence interval 0.74-0.92).
κ coefficient of 0.83 (95% confidence interval 0.74-0.92) for numbers in bold.
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
21
Table 5: KB Automated Reading Compared to Electrophysiologist Interpreted KB
Recordings
Electrophysiologist Interpreted KB Recordings
Kardia Band Automatic
Reading AF/Aflutter SR
Missing/Non-
Interpretable Total
AF/Flutter 71 1 2 74
SR 5 36 2 43
Missing/Unclassified 20 21 18 59
Total 96 58 22 176
Sensitivity, specificity and κ coefficient are calculated only for the simultaneous
transmission with interpretation (in bold). Sensitivity 93% (71/76), (95% confidence
interval 0.88-0.99), specificity 97% (36/37), (95% confidence interval 0.92-1.00).
κ coefficient of 0.88 (95% confidence interval 0.79-0.97) for numbers in bold.
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
MANUSCRIP
T
ACCEPTED
ACCEPTED MANUSCRIPT
1
Online Appendix
Online Table 1 – KB Algorithm Reading Compared to Electrophysiologist Interpreted 12-
Lead ECG Prior to CV (excludes post CV and unclassified measurements)
KB Algorithm
Interpretation
Electrophysiologist Interpreted 12-lead ECG
AF/Flutter SR
Total
AF/Flutter 60 0 60
SR 5 5 10
Total 65 5 70
Sensitivity 92% (60/55), (95% confidence interval 0.86-0.99); Specificity 100% (5/5), (95%
confidence interval 0.48-1.00).
κ coefficient of 0.63 (95% confidence interval 0.34-0.92) for numbers in bold.