Predictive modelling of medication adherence in post-myocardial infarction patients: a Bayesian approach using...

Predicting medication adherence in post-myocardial infarction (MI) patients has the potential to improve patient outcomes. Most adherence prediction models dichotomize adherence metrics and status. This study aims to develop medication adherence prediction models that avoid dichotomizing adherence metrics and to test whether a simplified model including only 90-days adherence data would perform similarly to a full multi-variable model.

Post-MI adult patients were followed for 1-year post the event. Data from pharmacy records were used to calculate proportion of days covered (PDC). We used Bayesian beta-regression to model PDC as a proportion, avoiding dichotomisation. For each medication group, statins, P2Y12 inhibitors and aspirin, two prediction models were developed, a full and a simplified model. 3692 patients were included for model development. The median (inter-quartile range) PDC at 1-year for statins, P2Y12 inhibitors and aspirin was 0.8 (0.33, 1.00), 0.79 (0.23, 0.99), and 0.79 (0.23, 0.99), respectively. All models showed good fit to the data by visual predictive checks. Bayesian R2 for statins, P2Y12 inhibitors and aspirin models were 61.4%, 71.2%, and 55.2%, respectively. The simplified models showed similar performance compared with full complex models as evaluated by cross validation.

We developed Bayesian multi-level models for statins, P2Y12 inhibitors and aspirin in post-MI patients that handled 1-year PDC as a proportion using the beta-distribution. In addition, simplified models, with 90-days adherence as single predictor, had similar performance compared with full complex models.

Predicting adherence to medications in patients after myocardial infarction may help focusing resources on patients with the highest need for medical attention. Medication adherence is usually calculated from prescription filling data. Most previously published prediction models categorized patients as 'adherent' or 'non-adherent' and then tried to predict to which category a certain patient would belong. We suggest here a method to avoid the need for such categorisation. This method can successfully predict the extent of prescription filling. Moreover, we found that simple prediction models, needing only information on the first 3 months prescription filling behaviour, was as good as complex models that required many predictors.