OBJECTIVES :

Atrial fibrillation and atrial flutter (AF/F) are common cardiac arrhythmias associated with morbidity, heart failure and stroke. Consequently, detecting undiagnosed AF/F is a valuable concept for ascertaining those who may benefit from preventive treatment. We assessed whether non-linear machine learning can improve predictive accuracy compared to conventional linear statistical methods.

METHODS :

A retrospective observational study of data from the UK Clinical Practice Research Datalink (CPRD) between 01-01-2006 and 31-12-2016 was undertaken. Patients were included if they had a complete set of key clinical variables within a 12-month window, aged > 30 years and without a history of AF/F. Four machine learning algorithms (LASSO, neural networks, random forests, and support vector machines) were compared to a Cox model and previously published statistical models to predict the probability of AF/F. Models were fitted, optimised and assessed on randomly selected, independent subsets. Accuracy was primarily assessed by area under the `receiver operating characteristic' curve (AUC).

RESULTS

Of 2,994,837 analysed patients, 95,607 developed AF/F during follow-up. The Cox model resulted in an AUC of 0.782, compared with 0.811, 0.811, 0.812 and 0.818 for LASSO, support vector machines, random forests and neural networks, respectively. Compared to Cox modelling, neural networks exhibited greater sensitivity (0.837 versus 0.739), resulting in an additional 3,107 AF/F patients (9.7%) correctly identified as ‘at high-risk’. Compared to the broader population, patients identified as being ‘at high-risk’ but not yet diagnosed with AF/F tended to be older, were more likely to be male, in receipt of antihypertensive medication and have a history of other cardiovascular comorbidities.

CONCLUSIONS

Machine learning improved the accuracy of AF/F prediction compared with conventional statistical methods. Machine learning could potentially lead to better targeting of screening strategies and thus preventive treatment, resulting in improved patient outcomes and more efficient use of finite resources.