Individualized prediction of atrial fibrillation onset risk based on lifelogs

Abstract

Background and Objective

The Apple Watch alerts users to irregular heart rhythms and potential atrial fibrillation (AF), but delays in obtaining electrocardiograms (ECGs) after notifications can impede accurate disease diagnosis. We aimed to predict personalized AF risk using continuous Apple Watch lifelog data to facilitate timely ECG acquisition. We conducted two analyses: Keio and national. In the Keio analysis, AF patients underwent continuous 2-week Holter ECG monitoring, and a machine-learning model combining gradient-boosting decision trees and deep learning was developed. The national analysis recruited Apple Watch users across Japan to assess the model; data and survey responses were collected for seven days via a dedicated iPhone app.

Results

A total of 100 subjects (age: 63.9 ± 12.4 years, AF burden: 37.7 %) participated in the Keio analysis, while 8,935 subjects participated in the national analysis. Significant differences in Apple Watch data, including pulse rate (p < 0.001) and step count (p < 0.001), were observed between days with and without AF onset. Healthcare data measured by the Apple Watch, including sleep patterns, were significantly correlated with subjective survey responses (p < 0.001) and incorporated into the model. The model achieved an F-value of 90.7 % compared to diagnosis based on a 2-week Holter ECG. The model showed an additive benefit to Apple Watch irregular-rhythm notifications for AF detection (irregular-rhythm notification vs. model: 68.8 % vs. 88.2 % for paroxysmal AF and 84.4 % vs. 100.0 % for persistent AF).

Conclusions

Apple Watch-derived lifelogs enabled individualized AF onset risk assessment and the development of a machine-learning model for optimizing ECG timing for early AF detection.