The study population consisted of all HA patients who submitted a KM ECG from the start of the program in January 2017 until March 2018. The Hartwacht program is available to patients of CCN, a private outpatient cardiology clinic. Typically, patients presenting with paroxysmal AF, palpitations of unknown origin or near-collapse were selected by the cardiologists of this clinic to participate in the Hartwacht program, although indications for inclusion in the program were left at the discretion of the physician.

After inclusion in the HA program, participants received the KM device at home, downloaded the Kardia smartphone application and were instructed on its use by the Hartwacht team. Whenever participants experienced palpitations or related complaints, they were encouraged to record an ECG with the KM device, after which the ECG and its classification by the algorithm were automatically transferred to the patient’s electronic patient record. There was no limit to the number of ECGs that could be recorded. ECGs were assessed by the Hartwacht team, consisting of a supervising cardiologist (0.05 FTE), a specialised cardiology nurse (1.0 FTE) and a doctor’s assistant (0.02 FTE), working on weekdays from 08.00 hrs to 17.00 hrs. Furthermore, a cardiologist who could directly access all Hartwacht ECGs was available 24/7 for emergency purposes. Patients received feedback from the Hartwacht team within one working day by phone or email.

Anonymised data were obtained from Hartwacht for analysis. Variables included in the analyses were patient characteristics, number of ECGs per patient per month and time of day that ECGs were received, classification of the ECG by the KM algorithm and results of assessment by the Hartwacht team. The KM algorithm classifies each ECG as one of four categories: (a) normal sinus rhythm, (b) possible AF, (c) unclassified or (d) unreadable. Assessment by the Hartwacht team resulted in classification as (a) sinus rhythm, with or without premature atrial contractions (PACs) and/or premature ventricular contractions PVCs, (b) atrial fibrillation, (c) other arrhythmias (including wide and narrow complex tachycardias and complete atrioventricular block) and (d) uninterpretable. Continuous variables were presented as mean ± standard deviation, categorical variables as frequencies. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and accuracy were calculated from the KM interpretation, with the Hartwacht team interpretation as reference standard. KM ECGs categorised as unclassified or unreadable were included in the sensitivity/specificity calculation when the Hartwacht team was able to provide a diagnosis. For example: An ECG interpreted by the KM algorithm as unclassified and subsequently by the Hartwacht team as atrial fibrillation deemed false negative for AF.

Irregular rhythms other than AF pose a challenge to the algorithm, which is illustrated by the distribution of ECGs diagnosed as sinus rhythm with ectopy over the KM classes. KM ECGs diagnosed by the cardiologist as sinus rhythm with PACs (5%) were classified by the KM algorithm as sinus rhythm (47%), atrial fibrillation (26%) or unclassified (27%) (Fig. 3c). ECGs with sinus rhythm and PVCs were similarly interpreted (Fig. 3d). In addition to sinus rhythm and atrial fibrillation, a small portion of ECGs (2%) showed other arrhythmias. For details see Tab. 2. There were several remarkable ECGs, shown in Fig. 4.

Using the assessment of the Hartwacht team as reference standard, the reliability of the different KM classifications in terms of sensitivity, specificity, PPV, NPV and accuracy, was determined (Fig. 5). For diagnosing AF these were 0.92, 0.95, 0.80, 0.98 and 0.94 respectively (upper table); for normal sinus rhythm (without any PACs or PVCs) 0.85, 0.83, 0.90, 0.76 and 0.84, (middle table); and for any form of sinus rhythm (with or without PACs or PVCs) 0.80, 0.91, 0.96, 0.65 and 0.83, respectively (lower table).

The KM algorithm focusses on detection of AF. In previous studies validating the device, high sensitivity and specificity were obtained for detection of AF. Lau et al. compared the KM ECG to a 12-lead ECG in 109 cardiology patients, a third of whom were in AF, and found a sensitivity of 87% and specificity of 97% [1]. Using a similar study design, Haberman compared 381 athletes with healthy volunteers and cardiology clinic patients (sensitivity 94%, specificity 99%) [2]. On the other hand, screening of patients for AF in a geriatric ward resulted in a specificity of 98% but a disappointing sensitivity of 55%. The researchers partly attribute these findings to interference/artefacts resulting in poor P-wave recognition, which may be an issue in the elderly population.

In the present study, sensitivity and specificity for detection of AF were comparable to previously reported values (92% and 95%, respectively). PPV, however, was substantially lower (80%) and NPV was high (98%) compared to a previous study using the same algorithm [6]. Most likely, this is attributable to the relatively low prevalence of AF in our study cohort. Of note, comparison of outcomes between the present study and literature should be done with care since all ECGs categorised by the KM algorithm as unclassified or unreadable were often discarded in previous studies, whereas they were included in the present one. Although the majority of AF ECGs are classified correctly, 78 (7%) are erroneously allocated to the unclassified category and 11 AF (<1%) to the normal sinus rhythm category. In the present study, these ECGs were re-evaluated by the Hartwacht team, which significantly increased diagnostic yield with regard to AF and shows the added value of the HA program.

The KM device has also been studied as an AF screening instrument in asymptomatic patients aged 65 years or older [7‐11], which resulted in 0.8–1.5% cases of newly diagnosed atrial fibrillation. The present study shows that when using the KM device for screening a general population for AF, two issues need to be considered. Firstly, prevalence of AF will be even lower compared to the present study and consequently PPV will be low with a high number of false positives resulting in a negative impact on cost-effectiveness. Secondly, a substantial number of AF ECGs will be not classified as possible AF emphasising the need for manual assessment of all ECGs.

Sinus rhythm with ectopy (PACs and/or PVCs) provides a challenge to the algorithm. The algorithm is unable to categorise these arrhythmias properly, as is shown in Fig. 3c, d, where all three KM classes contain a substantial percentage of PACs and PVCs respectively. The importance of recognising these rhythms depends on the indication for participation in HA. Screening for AF would require these rhythms to be discarded whereas for analysis of palpitations of unknown origin, detection of SR with ectopy might provide a useful explanation. This might also apply for other rhythm disorders. Moreover, the algorithm classified 1,033 (17%) ECGs as unclassifiable. From 81% of those a diagnosis could be provided by the Hartwacht team, further supporting the need for manual evaluation of all ECGs in such a program.

Several limitations of this study need to be addressed. Firstly, this is a retrospective analysis of a patient population included in the HA program for various reasons and at the discretion of the physician. This may have introduced a substantial selection bias. Consequently, predictive values provided should be interpreted with caution as they vary with prevalence. Furthermore, in the present study the reference for analysis of the 30-second one-lead KM ECGs is the Hartwacht team. In a previous study, Bumgarner et al. reported a specificity of 97% for detection of AF when comparing the KM classification to interpretation by a cardiologist [6]. However, specificity dropped to 84% when comparing the KM classification to evaluation of a simultaneously obtained 12-lead ECG. As 12-lead ECGs were not available in the present study, specificity data might have been overestimated.