P1938Comparing atrial fibrillation detection algorithms in smart devices on validated mobile ECG data

Abstract

Background: Previous trials show encouraging results that detection of Atrial Fibrillation (AF) by fingertip single lead ECGs (iECGs) and photoplethysmography (PPG) with smartphone cameras or smartwatches is feasible with high accuracy. The European Heart Rhythm Association added PPG-Apps and iECGs to their AF screening recommendations. Purpose: Improving automated categorization into Sinus Rhythm (SR) or AF by detection algorithms in smart devices could result in increasing detection rates, allow broader monitoring and therefore might reduce AF caused stroke risk. In a previously conducted trial 125 iECGs out of 662 (18.9%) could not be classified by the primarily used algorithm (A) but were of sufficient quality to be analysed by cardiologists. We hypothesized that another algorithm, originally designed to detect AF in PPG signals (B) may be as well applicable on the iECG-traces. A detects p-waves. B is based on non-linear beat-to-beat variability and discriminates between SR and absolute arrhythmia. Methods: We conducted a prospective, blinded analysis from a single center population of a multinational, multicenter prospective trial and tested the performance of two algorithms applied on iECG data compared to a cardiologist-based diagnosis. To make the iECG-files suitable for analysis by B we transformed a PDF-file containing the ECG-traces of the original iECG into digital coordinates using a semi-automatic extraction tool. An automated R-spike detector was applied. B classifies ECG data into AF or SR only and without regard to possible interference during the transformation process. Results: 371 iECGs were transformed and tested for its performance. B reached an overall sensitivity of 94% (95% CI: 89.3-97.3%), specificity of 87.9% (95% CI: 82.8-91.9%), accuracy and correct classification rate (CCR) of 90.6%. Excluding the iECGs not classifiable by A, Algorithm B reached a sensitivity of 96.5% (95% CI: 92.1-98.9%), specificity of 93.3% (95% CI: 88.2-96.6%), CCR and accuracy of 94.7%. Out of those not classifiable by A, Algorithm B showed a sensitivity of 69.2% (95% CI: 38.6-90.9%), specificity of 70.6% (95% CI: 56.2-82.5%), accuracy and CCR of 70%. Algorithm A did not show a diagnosis in 17.2%. In the remaining, sensitivity and specificity was 100%, and 97.5% respectively (95% CI: 93.8%-99.3%). Accuracy was 98.7% and CCR 81.6%. Conclusion: An algorithm using beat-to-beat variabilities of pulse waves from PPG signals can be also applied for R-R interval variability analysis in iECGs with very strong performance and led to a higher number of correctly classified patients. If able to classify, Algorithm A using p-wave detection showed strong results. Combining advantages of algorithm A and B might result in an overall higher diagnostic yield for iECG users and will be further investigated. Decreasing possible interference by optimizing tools for the extraction of the ECG-trace from the iECG may increase diagnostic performance of B.