Predictors of Adherence to Stroke Prevention in the BALKAN-AF Study: A Machine-Learning Approach

Background Compared with usual care, guideline-adherent stroke prevention strategy, based on the ABC (Atrial fibrillation Better Care) pathway, is associated with better outcomes. Given that stroke prevention is central to atrial fibrillation (AF) management, improved efforts to determining predictors of adherence with ‘A’ (avoid stroke) component of the ABC pathway are needed.

Purpose We tested the hypothesis that more sophisticated methodology using machine learning (ML) algorithms could do this.

Methods In this post-hoc analysis of the BALKAN-AF dataset, ML algorithms and logistic regression were tested. The feature selection process identified a subset of variables that were most relevant for creating the model. Adherence with the ‘A’ criterion of the ABC pathway was defined as the use of oral anticoagulants (OAC) in patients with AF with a CHA2DS2-VASc score of 0 (male) or 1 (female).

Results Among 2,712 enrolled patients, complete data on ‘A’-adherent management were available in 2,671 individuals (mean age 66.0 ± 12.8; 44.5% female). Based on ML algorithms, independent predictors of ‘A-criterion adherent management’ were paroxysmal AF, center in capital city, and first-diagnosed AF. Hypertrophic cardiomyopathy, chronic kidney disease with chronic dialysis, and sleep apnea were independently associated with a lower likelihood of ‘A’-criterion adherent management.

ML evaluated predictors of adherence with the ‘A’ criterion of the ABC pathway derived an area under the receiver-operator curve of 0.710 (95%CI 0.67–0.75) for random forest with fine tuning.

Conclusions Machine learning identified paroxysmal AF, treatment center in the capital city, and first-diagnosed AF as predictors of adherence to the A pathway; and hypertrophic cardiomyopathy, chronic kidney disease with chronic dialysis, and sleep apnea as predictors of non adherence.