BRIDGE: a barrier-informed Bayesian risk prediction model for risk identification, trajectory grouping, and profiling of non-adherence to cardioprotective medicines in primary care
H Koh, C Trin, Z Ademi, E Zomer, D Berkovic, P C Miranda, B Gibson, S Bell, J Illomaki, D Liew, C Reid, S Lybrand, A Earnest, S TalicAbstract
Background
Medication non-adherence is a major driver of cardiovascular morbidity, yet adherence to cardioprotective therapies (specifically lipid lowering) is often assessed using static measures that obscure important early behavioural patterns. Distinct longitudinal trajectories, such as rapid or gradual decline, remain poorly predicted in primary care. No current models combine patient-reported barriers at treatment initiation with real-world prescribing data to forecast these trajectories, limiting opportunities for timely, targeted adherence interventions.
Methods
We developed BRIDGE, a Bayesian model that predicts adherence trajectories based on GBTM trajectories. BRIDGE is a barrier-informed Bayesian multiclass model that uniquely integrates risk factors and patient-reported barriers as electronic medical record proxies grouped by WHO adherence domains, incorporating these into hierarchical priors to regularise trajectory prediction and enable counterfactual intervention analysis. Performance was compared with machine-learning baselines. We conducted counterfactual analyses to quantify the impact of modifying specific barriers at initiation.
Results
Among 50,881 LLT initiators (mean age 63 years; 90% statins), trajectories were: early discontinuation (41%), rapid decline (29%), persistent adherence (20%), and gradual decline (10%), nearly 70% exhibited early non-adherence patterns. BRIDGE achieved AUROC 0.808 (95% CI 0.805–0.811), accuracy 58.4%, and Brier score 0.531, comparable to XGBoost with superior interpretability. Incorporating hierarchical priors improved accuracy by 3.4 percentage points and enhanced calibration. Barrier profiles differed by trajectory: early discontinuers showed lesser healthcare utilisation (lowest visit frequency, 0.44/month); gradual decliners reflected medication burden (highest polypharmacy, 13%); rapid decliners indicated therapy instability (highest switching, 4%); persistent adherers displayed condition-related motivation despite higher clinical risk. Counterfactual analysis identified trajectory-specific intervention levers: timely adherence-support touchpoints and medication burden reduction for gradual decliners; continuity-of-care and early refill support for early discontinuers; and treatment stability with regimen simplification for rapid decliners.
Conclusions
Embedding structured, risk factors and patient-reported barriers as hierarchical priors improves prediction of longitudinal adherence trajectories while preserving transparency. BRIDGE enables barrier-targeted interventions at initiation using routine primary care data. External validation and a prospective implementation study are underway.Visual Ab