A new multi-step diagnostic pathway using machine learning ECG analysis and wearables to detect Atrial fibrillation in embolic stroke of unknown source

Background Atrial fibrillation (AF) is the most frequent reason for Embolic strokes of unknown source (ESUS) in elder patients. Detecting paroxysmal AF requires time and resources. Single holter ECGs have surprisingly low AF detection rates, requiring repeated measurements. The gold standard, an implantable loop recorder (ILR) with continuous ECG monitoring, is cumbersome to organize and reimburse leaving diagnostic gaps following ESUS. Purpose We investigate feasibility and AF detection rate of a multi-step diagnostic pathway interconnecting three different methods of a machine learning-based algorithm, a software smartphone application and a wearable device in ESUS following acute ischemic stroke to close diagnostic gaps in post stroke AF detection. Methods We enrolled 50 patients (78,6 yo, 68-94 yo, 31m, 77,8 yo/19w; 79,6 yo, p=ns) presenting with acute ESUS to the stroke unit of our university hospital. On dismissal, a digital holter ECG of 7 days was conducted. This was conventionally analysed for AF. In absence of AF, a machine learning-based algorithm (Apoplex medical SRA) analyzed the ECG to detect hints for previous AF in sinus rhythm and thus categorize the predictive risk for AF (low- and high-risk) Over a subsequent 100 day outpatient period, patients wore a smartwatch (Fitbit Sense) continuously to enable bodyplethysmographic detection of irregular pulse and to actively measure a 30 second ECG three times daily. These intermittent measurements were accompanied by a software app on the smartphone (Preventicus Heartbeats) to measure the pulse of the patients by photoplethysmography. Results Measurements showed robust feasibility resulting in 4879 ECGs. Direct detection of AF in the 7-day ECG occurred in 3 % of patients. The machine learning-based SRA algorithm revealed indicators of AF during sinus rhythm in additional 28 %. In remaining patients without predictors of AF in holter ECG, no AF was detected by other wearables or holter in 71 %. The Fitbit Sense smartwatch alerted arrythmias in continuous photoplethysmography in 9,7 % of patients. All of these detected AF in a subsequent ECG. In addition, the repeated patient-induced ECG measurements using the smartwatch revealed AF in 29 % of patients. All of these detections were later on confirmed by ECG analysis of a qualified cardiologist. Preventicus App showed similar results (18% detection of irregularities). Conclusion The proposed multi-step diagnostic pathway detected AF in a third of ESUS patients additional to conventional measurements. All methods proved robust feasibility and reliable diagnostic value detecting AF in elder patients following acute ESUS. The Fitbit Sense smartwatch provided all of its detections with ECG to proof of AF and lead to a change in anticoagulant therapy post ESUS. The machine learning algorithm identified the remaining patients in low- and high-risk, selecting a minimal number of patients requiring ILR implantation.