Optimal adherence thresholds for oral anticoagulants in patients with atrial fibrillation using machine learning and population administrative data.

BACKGROUND

Adherence to oral anticoagulants (OACs) for atrial fibrillation (AF) stroke prevention is traditionally defined as taking 80% of doses as prescribed, but this threshold has not been clinically validated.

OBJECTIVES

We identified OAC adherence thresholds maximally associated with risk of clinical events in patients with AF and compared them with the conventional 80% threshold.

METHODS

This was a cohort study using retrospective data of patients newly diagnosed with AF who were new OAC users from 1996 to 2019, based on population-based administrative data from British Columbia, Canada. We used Cox proportional hazards models with OAC proportion of days covered (PDC) as the main exposure captured during 90 days before outcome events or at the end of follow-up, then applied least absolute shrinkage and selection operator-estimated coefficient paths to identify candidate thresholds, and identified which were optimal in terms of outcomes.

RESULTS

A total of 44 172 patients were included. For all outcomes, vitamin K antagonist (VKA) optimal thresholds were between PDC 0.85 and 0.95, and the direct OAC (DOAC) optimal threshold was PDC 0.9. Above vs below threshold outcome hazard reduction was greater for DOACs than for VKAs (DOACs: 19%-52% stroke risk reduction; VKAs: not significant for most outcomes). The PDC 0.8 threshold had inferior model fit and outcome effect compared with the optimal thresholds identified.

CONCLUSION

Optimal adherence thresholds are higher than conventionally assumed and, especially for DOACs, are strongly associated with clinical outcomes. These results provide a basis for considering updating the definition of OAC nonadherence to PDC <0.9 for DOACs and to PDC <0.9 or <0.95 for VKAs, depending on the outcome.