Photoplethysmography-Based Machine Learning Approaches for Atrial Fibrillation Prediction

Background Current wearable devices enable the detection of atrial fibrillation (AF), but a machine learning (ML)–based approach may facilitate accurate prediction of AF onset. Objectives The present study aimed to develop, optimize, and validate an ML-based model for real-time prediction of AF onset in a population at high risk of incident AF. Methods A primary ML-based prediction model of AF onset (M1) was developed on the basis of the Huawei Heart Study, a general-population AF screening study using photoplethysmography (PPG)–based smart devices. After optimization in 554 individuals with 469,267 PPG data sets, the optimized ML-based model (M2) was further prospectively validated in 50 individuals with paroxysmal AF at high risk of AF onset, and compared with 72-hour Holter electrocardiographic (ECG) monitoring, a criterion standard, from September 1, 2019, to November 5, 2019. Results Among 50 patients with paroxysmal AF (mean age 67 ± 12 years, 40% women), there were 2,808 AF events from a total of 14,847,356 ECGs over 72 hours and 6,860 PPGs (45.83 ± 13.9 per subject per day). The best performance of M1 for AF onset prediction was achieved 4 hours before AF onset (area under the receiver operating characteristic curve: 0.94; 95% confidence interval: 0.93-0.94). M2 sensitivity, specificity, positive predictive value, negative predictive value, and accuracy (at 0 to 4 hours before AF onset) were 81.9%, 96.6%, 96.4%, 83.1%, and 88.9%, respectively, compared with 72-hour Holter ECG. Conclusions The PPG- based ML model demonstrated good ability for AF prediction in advance. (Mobile Health [mHealth] technology for improved screening, patient involvement and optimizing integrated care in atrial fibrillation; ChiCTR-OOC-17014138)

A trial fibrillation (AF) is the most common cardiac rhythm disorder; because of its association with increased risk of stroke, heart failure, dementia, and death (1), efforts have been directed toward improving the detection of and screening for AF (2).
Screening for AF can be systematic or opportunistic, with the latter being more cost-effective, especially considering that patients at high risk of incident AF (eg, those with previous myocardial infarction, heart failure, chronic chest disease, or stroke) would normally attend clinical follow-up with health professionals. Various common and validated risk factors have been used to propose clinical risk prediction models for incident AF, but most of these models only have modest predictive value (3). Recent studies have shown that the clinical scores predicting AF recurrence after ablation have limited predictive ability (4). Nonetheless, clinical risk prediction scores can be used to identify high-risk subjects (eg, after stroke) who should be targeted for more intense screening efforts (5).
A clinical approach to AF screening has recently been complemented by various "smart" options for improving AF detection, including smart devices, wearable patches, and wearable devices, such as smartwatches linked to smartphones (6). Current wearable devices enable the detection of AF, but a machine learning (ML)-based approach may facilitate even more accurate prediction of incident AF. In the Huawei Heart Study, we conducted a populationbased screening study of 187,912 individuals, where 0.23% received a "suspected AF" notification and 87.0% of those were confirmed as having AF, with a positive predictive value (PPV) of 91.6% (7). Thus, continuous home monitoring with smart devicebased photoplethysmography (PPG) technology could be a feasible approach for AF screening.
Nevertheless, it has not yet been investigated whether the prediction of AF onset can be improved with the use of the PPG signals from smart devices.
The objectives of this prespecified ancillary analysis from the Huawei Heart Study were to develop, optimize, and validate an ML-based model for predicting the onset of AF from normal sinus rhythm in patients at high risk of incident AF, eg, those with paroxysmal AF.    Figure 2). The "AF onset" events (AF onset lasting >30 seconds) predicted by M2,      demonstrated an improved predictive ability in terms of the accuracy, precision, F1 score, and AUC ( Table 1).

METHODS
The ROC and the precision-recall curve of M1 and M2 are presented in Figure 3. M2 was superior to M1 in predicting AF onset, with a difference between AUCs of 0.01-0.04, in 2 randomly split data sets (DeLong test, all P < 0.05) ( Table 1, Supplemental Figure 4).
Compared with M1, the false-positive rate of predicted AF onset with M2 was significantly reduced, by 1.19 AE 0.28 and 1.60 AE 0.44 in 2 randomly split data sets (all P < 0.05) (Supplemental Table 1).
The optimized model (M2) was real-time tested among 1,709 individuals with detected AF during 3 months. During the 3-month testing period, the sensitivity and specificity of M2 for AF prediction 0 to 4 hours before AF onset were 0.78 and 0.93, respectively (Supplemental Figure 5).
The false-positive rate of AF prediction at 0 to 4 hours before AF onset with the use of M2 (with a cutoff at 0.5) was 5.58 (95% CI: 5.00-6.22). Of these false positives, more than 85% of the "real" events were atrial bigeminy, trigeminy, and atrial flutter (Supplemental Table 3).

Sensitivity analyses of the predictive ability of M2
with the use of 3,403 AF events with or without prior PPG within 4 hours are presented in Supplemental  Table 4).

DISCUSSION
In a population-based screening cohort using PPG-    We thank LetPub (www.letpub.com) for its linguistic assistance during the preparation of this manuscript.

FUNDING SUPPORT AND AUTHOR DISCLOSURES
This research project was funded by the National Natural Science The artificial intelligence-based model with more frequent, dynamic monitoring data could achieve an AUC >0.90.
TRANSLATIONAL OUTLOOK: A data-driven "early warning" smart tool not only could identify/diagnose the AF events that did actually happen, but also could predict AF onset in advance, with much better predictive ability than traditional clinical factor-based tools. This would likely change AF prevention and management, but large prospective cohort studies and randomized trials are needed to address the future application.