This study shows that AF screening in a high-risk population using the Kardia device is feasible, although expert interpretation of the 15.1% unclassified and AF outcomes is required. The Kardia algorithm’s performance in our study is influenced by unclassified outcomes and supported by the literature. Lau et al. showed the Kardia device’s sensitivity and specificity for detecting AF to be 87% and 97%, respectively, with 12-lead ECG interpretation as the gold standard [10]. Desteghe et al. compared the Kardia algorithm with a 12-lead ECG in 265 patients. The algorithm’s sensitivity and specificity were 54.5% and 97.5%, respectively, with a NPV of 96% and PPV of 66.7% [12]. Unfortunately, neither study reported the percentage of unclassified outcomes, impeding a reliable comparison of sensitivity with that in our study. Other authors reported NPVs of 98.4% in two studies comparing the algorithm with a 12-lead ECG, and 98% when comparing the algorithm with expert interpretation [14,15,16].

Although the NPV of the Kardia device is high in our study, enabling sufficiently reliable SR classification, the 35.0% false-positive outcomes are a limitation. This finding is in accordance with the previously mentioned studies, with PPVs ranging between 54.8% and 80% [12, 15, 16]. This might have been caused by premature complexes, as confirmed by the Hartwacht study, where experts interpreted 8% of the cases with an algorithm outcome of AF as SR with ectopic beats [16].

Not only the cases classified by the algorithm as AF should be reassessed. The majority of the unclassified sECGs could be interpreted by experts and therefore should always be reviewed. High proportions of unclassified sECGs and the expert’s ability to assess these sECGs were reported previously. The Hartwacht team showed 17% unclassified algorithm outcomes and 8% uninterpretable expert interpretations [16]. Other studies reported even higher percentages of unclassified outcomes: 19.5–27.5% [15, 17].

This study showed an overall prevalence of newly detected AF of 2.5%, with a higher prevalence at the GP. Although we assumed this difference could have been caused by selection bias from possibly screening symptomatic participants, the proportion of participants experiencing palpitations was not higher than in the other groups. Another explanation could be that most participants were included during cardiovascular risk management consultations, including patients at higher risk of AF. This is also supported by the higher incidence of hypertension and diabetes mellitus compared with other screening locations besides the nursing home. The relatively low rate of AF cases in nursing homes could have been a result of a higher prevalence of already detected AF.

Several screening studies have been performed with varying AF prevalences. In 2007, Fitzmaurice et al. studied 14,802 participants aged ≥ 65 years, randomised to (1) 12-lead ECG single screening, (2) 12-lead ECG single screening in the case of an irregular pulse, or (3) no screening. Detection rates of new AF cases were 1.62%, 1.64% and 1.04%, respectively [18]. The STROKESTOP study in 2015 included 7173 participants aged 75–76 years who recorded a sECG twice a day for 2 weeks using the Zenicor device (Zenicor Medical Systems, Stockholm, Sweden), resulting in an AF prevalence of 3.0% [19]. The REHEARSE-AF study in 2017 included 1001 patients aged ≥ 65 years who were randomised to a 12-month screening programme using the Kardia device twice a week or routine care and resulted in 3.8% and 1.0% AF, respectively [20]. Both the STROKESTOP and the REHEARSE-AF study showed the diagnostic benefits of periodic screening. However, the appropriate and cost-effective screening frequency remains to be defined [9]. The study by Kaasenbrood et al., resembling part of our study with AF screening during seasonal influenza vaccination using the MyDiagnostick device (MyDiagnostick Medical, Maastricht, The Netherlands) revealed 1.1% newly detected AF. Although all age groups were screened, no new cases were detected in participants aged below 60 years, which could be an explanation for the higher percentage of AF cases in our study [7].

A recently published EHRA position paper on searching for AF underlines the current challenges in AF detection and suggests the use of clinical risk scores (MR-DASH or C2HEST) to better refine target populations [21, 22]. Additionally, a monitoring time of 2 weeks or longer is preferred to maximise the possibility of identifying subjects with AF [23].

One major limitation of this study is that we did not collect data on the use of anticoagulation or an extensive medical history. Participants could not adequately recall their medication and in the setting of this population screening programme we did not include a search in medical records. Moreover, it remains unclear whether implementation of anticoagulation after opportunistic screening provides similar protection against stroke to when AF is diagnosed clinically.

Secondly, expert interpretation was considered the gold standard, which could still be wrongly interpreted. However, this is in line with the ESC guidelines, with a class 1B recommendation that definite diagnosis of AF in screen-positive cases is established after the physician reviews the sECG recording of ≥ 30 s [9].

Moreover, to facilitate rapid screening we only asked if participants experienced palpitations before screening for AF symptoms. Although AF can account for a broad spectrum of symptoms, we considered palpitations to be the most common symptom. Additionally, there is a wide variety in the number of participants included at different screening locations, requiring caution in the interpretation of differences in participant characteristics and screening outcomes. The yield of AF detection depends on the a priori chance of having AF, which is higher in our population of persons aged 65 years and above compared with the general population. A selection bias, selecting patients with higher comorbidity, could not be excluded. Therefore, extrapolation to the general population should be done carefully.

In our study, no cost-effectiveness analysis was performed. The previously mentioned influenza vaccination screening study showed that screening in primary care during seasonal influenza vaccination in a population aged ≥ 65 years would have an estimated probability of 99.8% for being cost-effective at a conservative willingness to pay of € 20,000/QALY [6]. The cost-effectiveness of screening is also implied by the STROKESTOP study [19].

In addition to the benefit of AF screening as described in the ESC guidelines, screening with the Kardia device is feasible when reviewing unclassified and AF outcomes [9]. Although the cost-effectiveness needs to be examined, our recommendation would be to implement screening at vaccination programmes. Ideally, persons at higher risk of new AF should be included and could be identified using risk scores [21, 22]. Preferably, persons with a high CHA2DS2-VASc score, thus requiring initiation of oral anticoagulation, should be selected. Persons already on oral anticoagulants might be excluded due to the lack of clinical consequences. Future research should examine the implementation of repeated or intensified screening programmes.

When persons are informed prior to the screening event, 2–3 min per person is required. It is recommended that cardiologists and GPs cooperate closely to facilitate quick reviewing: 15.1% of sECGs in the current study.