Atrial fibrillation detection plays an important role in the avoidance of mortality and morbidity associated with AF. To this end, traditional machine learning methods have been used in clinical practice for decades and include, for example, automated ECG machine diagnoses. Novel approaches have included transferring this well-known technology to mobile devices such as smartphones and wearable technology. While smartphone-based ECG devices have been developed by multiple companies [
14], the AliveCor Kardia AF detection algorithm has been evaluated most frequently by scientific studies and was shown to have a high negative predictive value for the presence of AF [
15‐
22]. Alternatively, smartphone applications such as the Cardiio app utilize photoplethysmographic measurements obtained through a built-in smartphone camera and was shown to be both sensitive and specific for the detection of AF in a recent publication by Yan et al. [
23]. However, currently, only two large-scale prospective studies exist on the utility of smart-technology based AF screening [
24,
25]. The Apple Heart study was conducted as a prospective, siteless study including 419,000 participants with an Apple smartwatch residing in the United States. Study subjects monitored their heart rhythm with a photoplethysmographic sensor and if a recording was interpreted by the automated algorithm as probable AF, a 7 day ECG screening was conducted by a mailed ECG patch [
24]. Of the 2161 participants (0.52%) who received a notification of irregular heart rhythm, 450 participants (21%) returned their ECG patches for analysis and AF was present in 34% of returned recordings. Similarly, the Huawei heart study included almost 190,000 Chinese participants who monitored their heart rhythm with a Huawei smartwatch-based photoplethysmographic algorithm [
25]. While 424 subjects (0.23%) received an automated algorithm interpretation of suspected AF, 262 (62%) were effectively followed up by either 12-lead or Holter ECG. Of those, 227 (87%) were confirmed to have a diagnosis of AF. These studies show the potential, but also the limitations of large population-based smart-technology screening programs, with an acceptable positive predictive value but also a significant loss of participants to clinical follow-up. Additionally, because of the study designs, the rate of false-negatives (i.e., persons with unknown AF and no photoplethysmographic recordings suspected to be AF) could not be reported.
Various studies have evaluated the utility of deep neural networks in AF screening. At least two working groups independently evaluated a DNN in the interpretation of smartwatch plethysmographic data und were able to document a significantly increased sensitivity and specificity compared to previously presented studies utilizing traditional machine learning methodology [
26,
27]. Concerning diagnosis from a one-lead ECG recording, Hannun et al. showed a superior accuracy of a DNN compared with board-certified cardiologists [
28], which furthermore extended to other arrhythmias such as regular supraventricular tachycardia and atrioventricular block. Unifying applications across different diagnostic modalities, Ramesh et al. recently reported the development of a DNN able to detect AF with a high diagnostic accuracy in both ECG and photoplethysmographic recordings [
29]. Importantly, neural networks can also aid the screening for AF even when it is absent at the time of presentation. For example, different groups were able to improve previous classical risk stratification models by utilizing DNNs to estimate the likelihood of AF occurrence in high-risk populations such as patients with chronic kidney disease [
30] or patients with a history of ischemic stroke [
31]. In a landmark publication, Attia et al. recently described a DNN trained on almost 650,000 ECG of 180,000 patients capable of detecting patients with AF from a sinus rhythm ECG [
32]. The DNN was trained on ECG obtained at the Mayo Clinic Rochester between 1993 and 2017 and sinus rhythm recordings were deemed to show a patient with AF if it was first documented within 31 days of the sinus rhythm recording. The authors were able to show a sensitivity of 79% and a specificity of 79.5% of the DNN in the detection of AF-patients from the sinus rhythm ECG, which further increased to 82.3% and 83.4%, respectively, when multiple ECG from the same patient were analyzed. A similar recent publication was able to considerably increase the diagnostic window by evaluating a DNN capable of detecting AF onset within one year of the recording of an index sinus rhythm ECG [
33].