OBJECTIVES: Current literature suggests that up to 30% of persons with atrial fibrillation are undiagnosed, representing unacceptable risk for thrombotic stroke in the population over age 50. This study attempts to identify undiagnosed persons with atrial fibrillation through the use or artificial intelligence/machine learning techniques.

METHODS: A de-identified 2.5-million member dataset was used to train a Gradient Boosted Tree model, identifying patterns in structured data associated with the presence of ICD coding for atrial fibrillation. In collaboration with a 350,000 member suburban multi-specialty practice, the resulting model was then applied to a 1,100 persons dataset where 1,000 were known to have an atrial fibrillation diagnosis code, and 100 did not. The atrial fibrillation codes were stripped off before running this test set, and the 1,100 persons were ordered in decreasing probability of having atrial fibrillation. A probability of 50% was used as the threshold.

RESULTS: The model had a sensitivity (recall) of 70% (correctly identifying 70% of A-fib coded persons), and a positive predictive value (precision) of 88% (88% of positive results were correct). Further, despite using only a 50% probability as the imputation threshold, there are very few false positives (specificity = 99%).

CONCLUSIONS: Artificial intelligence techniques hold promise as a means to more effectively capture persons likely to be experiencing occult atrial fibrillation, and directing them to effective anti-coagulant therapy. Future study will include a population health management intervention including ambulatory rhythm testing and subsequent rates of anticoagulation for persons exceeding the 50% probability threshold.