Application of a machine learning algorithm for detection of atrial fibrillation in secondary care

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• Atrial fibrillation (AF) is a prevalent heart arrhythmia increasing stroke risk.
• Current AF detection methods, like electrocardiograms, can be inefficient for high-risk individuals.
• Risk-prediction algorithms and machine learning (ML) show promise for improving AF diagnosis.

Purpose of the Study:

• To evaluate an AF-risk prediction algorithm's performance using linked primary and secondary care data.
• To identify patients with undiagnosed AF in secondary care who were missed in primary care.
• To assess the impact of newly identified AF cases on the algorithm's predictive accuracy.

Main Methods:

• Applied an ML-based AF-risk prediction algorithm to the DISCOVER database, linking primary and secondary care data.
• Analyzed algorithm performance in identifying patients with AF diagnoses solely in secondary care.
• Calculated algorithm sensitivity and specificity at a 7.4% risk threshold for patients aged 30 years and older.

Main Results:

• An additional 5,444 patients with AF diagnoses exclusively in secondary care were identified.
• The algorithm accepted 49.5% of these patients, correctly assigning 97.8% to the AF cohort.
• At a 7.4% risk threshold, algorithm sensitivity was 38% and specificity was 95%.

Conclusions:

• The ML algorithm effectively identifies previously undiagnosed atrial fibrillation in secondary care settings.
• Integrating secondary care data enhances AF detection, revealing a significant number of missed diagnoses.
• This approach offers a valuable tool for proactive AF identification and stroke risk mitigation.