Evaluating Smartwatch Detection of Atrial Fibrillation

Abstract

Smartwatches, such as the Apple Watch, are increasingly used for monitoring and detecting serious health conditions like atrial fibrillation (AF), a common arrhythmia with significant health risks. However, to avoid burdening the healthcare sector with false alarms, their detection algorithms are designed to minimize false positives, a choice that may compromise their sensitivity to sparse or intermittent AF. While previous studies have analyzed the specificity of these alerts, the rate of false negatives remains underreported. This thesis aims to quantify this performance gap by evaluating the Apple Watch’s photoplethysmography-based irregular-pulse notification and its rate of false negatives. Using a large clinical ambulatory ECG dataset, we modeled and simulated long-term AF episode patterns via two distinct stochastic methods: a two-state discrete-time Markov chain and a continuous-time alternating bivariate Hawkes process. Analysis of the simulated data revealed that a significant proportion of individuals, particularly those with low-to-moderate AF burden (0.5%–5.5%), risk going undetected for extended periods. For example, for a burden of 0.5%–1.5%, over 20% of cases remained undetected after five years, and for a burden of 3.5%–4.5%, over 5% remained undetected after five years of monitoring. The findings, consistent across both modeling approaches, demonstrate that detection is highly dependent on the temporal pattern of AF episodes and not just the AF burden. They also suggest that the Apple Watch algorithm has reduced sensitivity to certain AF episode patterns, and users should be cautious not to interpret the absence of an alert as confirmation of a normal rhythm.