The Accuracy of Smartphone Camera Apps to Detect Atrial Fibrillation: A Systematic Review, Meta-Analysis, Meta-Regression and Modeling Study

Background: Atrial Fibrillation (AF) affects more than 6 million people in the United States (US) and greatly increases one's risk of stroke. Despite its prevalence, much AF remains undiagnosed. Smartphone applications (apps) have been proposed to aid the detection of subclinical AF and to monitor chronic AF. However, the accuracy of these apps remains unclear. We set out to determine the accuracy of smartphone camera apps to diagnose AF and compare the accuracy of apps. Methods: We searched MEDLINE and EMBASE databases and grey literature for studies that assessed the accuracy of any smartphone camera app to diagnose AF. To synthesize data, we constructed bivariate random-effects meta-analyses to determine the meta-analyzed sensitivity and specificity. Furthermore, we modeled the positive predictive value (PPV) and negative predictive value (NPV) for different population groups (age ≥65 and age ≥65 with hypertension). Lastly, we explored the effect of methodological limitations with sensitivity analyses and meta-regressions. Findings: We analyzed data from 3852 participants across 10 primary diagnostic accuracy studies, which evaluated four different apps. The apps analyzed the pulsewave signal (PPG) for an average of 2 minutes (range: 1 to 5 mins). The meta-analyzed sensitivity and specificity for all apps combined was 94·2% (92·2% to 95·7%) and 95·8% (92·4% to 97·7%) respectively. We found the PPV and NPV for the detection of AF was between 19·3% to 37·5% and 99·8% to 99·9%, respectively, in an asymptomatic population aged ≥65. The PPV increased for people aged ≥65 year old and with hypertension (PPV: 20·5% to 39·2%, NPV: 99·8% to 99·9%)). We found methodological limitations in a number of studies that did not appear to impact diagnostic performance, but an effect could not be definitively excluded given data sparsity. Interpretation: All smartphone camera apps have a relatively high sensitivity and specificity. The modeled NPV was high for all analyses, but the PPV was modest, suggesting that using these apps in an asymptomatic population may generate more false, rather than true, positive results. Future research could address the accuracy of these apps to screen further highrisk population groups, and to monitor chronic AF. Funding Statement: The lead author (JOS) is supported by an unrestricted NIH Fellowship ( T32 post-doctoral Fellowship), however there was no specific funding for this study. Declaration of Interests: The authors stated: "None declared." Ethics Approval Statement: The protocol for this systematic review was developed and registered a priori (PROSPERO: CRD42019125253). This systematic review and meta-analysis were conducted in line with the Preferred Reporting Items for a Systematic Review and Meta-analysis of Diagnostic Test Accuracy Studies (PRISMA-DTA).