Scalable Approach to Consumer Wearable Postmarket Surveillance: Development and Validation Study

Area of Science:

• Medical Informatics
• Cardiology
• Digital Health

Background:

• Consumer wearables offer prediagnoses, potentially impacting healthcare.
• Electronic health record (EHR) systems lack standardized terms for wearable data, hindering postmarket surveillance.
• Atrial fibrillation (AF) prediagnoses from wearables are a key area of interest.

Purpose of the Study:

• To develop and validate an EHR-based surveillance system for consumer wearable AF prediagnoses.
• To demonstrate the feasibility and efficacy of using weak supervision for EHR data labeling.
• To compare patient characteristics and care patterns following wearable-detected AF prediagnoses.

Main Methods:

• Applied data programming and the Snorkel framework to create labeling functions for AF prediagnoses.
• Developed a labeler model to probabilistically label clinical notes for AF prediagnoses.
• Fine-tuned a Clinical-Longformer classifier on labeled notes to identify patients with AF prediagnoses.
• Conducted a retrospective cohort study comparing patients with and without AF prediagnoses.

Main Results:

• The labeler model achieved high accuracy (0.92) and F1-score (0.77).
• The Clinical-Longformer classifier accurately identified AF prediagnoses (0.95 accuracy, 0.83 F1-score).
• Patients with AF prediagnoses were older, male, White, and had higher CHA2DS2-VASc scores.
• Prediagnoses correlated with increased anticoagulant use (e.g., apixaban, rivaroxaban) and eventual AF diagnosis.

Conclusions:

• An EHR-based surveillance system for wearable AF prediagnoses is feasible and effective.
• The system can identify patient cohorts for further study and verify findings from large trials.
• Further research is needed to generalize findings across diverse patient populations and healthcare settings.