Model Derived from Electronic Health Records System Predicts Risk of Atrial Fibrillation

Recent studies have demonstrated that screening for atrial fibrillation (AF) is feasible, but because of concerns about misdiagnosis, inappropriate use of oral anticoagulation and bleeding, not all professional societies have endorsed it. Risk-based AF screening might minimize misdiagnosis, but no study has examined AF screening based on predicted risk.

Subscribe to the latest updates from Cardiovascular Advances in Motion

Thank you for subscribing!

Error: Please enter a valid email address.

Email Address Submit

Olivia L. Hulme, MD, clinical fellow, Shaan Khurshid, MD, research fellow, and Steven A. Lubitz, MD, MPH, of the Cardiovascular Research Center and Telemachus & Irene Demoulas Family Foundation Center for Cardiac Arrhythmias at Massachusetts General Hospital, and colleagues recently developed a computer model that uses information from electronic health records (EHR) to estimate AF risk in clinical practice. As the researchers report in JACC: Clinical Electrophysiology, the risk score for AF also stratified the risk of subsequent stroke.

Study Subjects

From the EHR database of Mass General and its six partner hospitals, the researchers identified individuals who had at least one outpatient visit in each of two consecutive years between January 2000 and December 2014, were 45 to 95 years old at the first eligible visit and were free of AF. The researchers collected follow-up data through April 20, 2017.

Creation of the AF Prediction Model

The sample was divided into a derivation cohort (206,042 individuals) and a validation cohort (206,043 patients). Data extracted from the EHR included diagnostic and procedural codes, medications, cardiology test reports, discharge summaries, clinic notes and vital status. The presence of AF or atrial flutter was ascertained using a validated algorithm that Dr. Lubitz's group published in The American Journal of Cardiology.

The researchers used the derivation cohort to identify the best predictors of incident AF within five years. The variables selected were male sex, age, race, smoking history, height, weight, diastolic blood pressure, hypertension, hyperlipidemia, heart failure, coronary heart disease, valvular disease, previous stroke and transient ischemic attack, peripheral arterial disease, chronic kidney disease and hypothyroidism.

An EHR-based score for risk of AF was then obtained for each individual in the validation set. The researchers tested associations between that score and the actual incidence of AF within five years. They also compared the EHR-based score with three established methods of predicting AF risk: CHARGE–AF, C2HEST and CHA2DS2–VASc.

Comparison of Scoring Systems

In the validation cohort, the EHR-derived score was significantly associated with the five-year risk of new-onset AF, and it performed favorably compared with the other risk scoring systems. This was judged from the C-statistic (the probability that an individual who developed AF had a higher risk score than a patient who did not) and the calibration slope (predicted vs. observed probability of AF). They found:

• EHR-derived score: C-statistic, 0.777; calibration, 0.99
• CHARGE–AF: C-statistic, 0.753; calibration, 0.72
• C2HEST: C-statistic, 0.754; calibration, 0.44
• CHA2DS2–VASc: C-statistic, 0.702; calibration, 0.37

Stroke Risk

The EHR-derived AF risk score predicted an increased risk of:

• Incident stroke (C-statistic, 0.684)
• Stroke within 90 days before an AF diagnosis, potentially the initial manifestation of AF (C-statistic, 0.789)

The five-year predicted risk of AF was classified as low (<2.5%), intermediate (2.5%–5%) or high (>5%). Relative to individuals with low predicted AF risk, they found:

• Individuals at intermediate risk had a 1.85-fold increased risk of incident stroke and a 2.22-fold increased risk of stroke within 90 days prior to an AF diagnosis
• Individuals at high risk had a 4.47-fold increased risk of incident stroke and a 10.5-fold increased risk of stroke within 90 days prior to an AF diagnosis

Looking Ahead

Automated AF risk prediction may someday be incorporated into EHR systems as a clinical decision support tool. Clinicians would be able to give patients individualized risk estimates at the point of care without having to perform complex calculations themselves.

At the population level, automated risk estimation could identify groups of patients most likely to benefit from large-scale interventions to prevent or diagnose AF, such as AF screening.

Locally derived risk prediction scoring may prove to be more precise and accurate than scoring systems that are developed in large, nationwide research cohorts, as CHARGE-AF was. It may also become worthwhile to derive risk models for specific populations, such as elderly patients and individuals with recent cardiovascular events.

view original journal article Subscription may be required

Learn more about the Telemachus & Irene Demoulas Family Foundation Center for Cardiac Arrhythmias

Refer a patient to the Corrigan Minehan Heart Center