Clinical validation of an artificial intelligence algorithm offering cross-platform detection of atrial fibrillation using smart device electrocardiograms

Abstract

Abstract Introduction Multiple smart devices capable of “screening” for atrial fibrillation (AF) based on single-lead electrocardiogram (SL ECG) are presently available. Manufacturers' algorithm capabilities and accuracy for the automated detection of AF vary. Reliable artificial intelligence (AI) algorithms would be valuable to assist physicians with managing the large amount of data. We aimed to assess the clinical value of applying a smart device agnostic AI-based algorithm for the detection of AF from five different smart devices (four smartwatches, one handheld device) and compared the results to the cardiologist-interpreted 12-lead ECG in a real world cohort of patients. Methods This is a prospective, observational study enrolling patients presenting to a cardiology service at a tertiary referral center. Patients were prescribed a 12-lead ECG, followed by five consecutive smart device recordings from five different manufacturers. SL ECGs were exported as PDF files from the devices and analyzed by a deep neural network (DNN) based platform which allows automated AI assisted cardiac rhythm interpretation. Results We prospectively enrolled 157 patients (32% female, median age 66 years). AF was present in 48 patients (31%) at time of recording, as documented by the 12-lead ECG. Accuracy for the detection of AF by the DNN-based algorithm was 96.6% for the Apple Watch 6, 95.2% for the AliveCor Kardia Mobile, 96.0% for the Fitbit Sense, 95.7% for the Samsung Galaxy Watch 3 and 93.8% for the Withings Scanwatch, respectively (Figure 1, left). While diagnostic accuracy of the DNN-based algorithm was similar compared to each manufacturer's individual algorithm, the proportion of SL ECGs with a conclusive diagnosis was significantly higher for all smart devices when using the DNN-based algorithm, p<0.001 (Figure 1, right). As complementary analysis, we assessed sensitivity and specificity detection capabilities in both algorithms (Figure 2). Conclusion In this clinical validation, a DNN-based algorithm reported significantly more conclusive diagnoses for each smart device compared to the manufacturers' algorithms, whilst showing similarly high accuracy in the detection of AF compared to the cardiologist-interpreted standard 12-lead ECG. Given further validation, SL ECG assisted rhythm interpretation through a cross-platform AI-algorithm presents a promising clinical value for AF detection and offers a possible solution for managing the data surge for smart device-acquired ECGs. Funding Acknowledgement Type of funding sources: None.