Effects of bisphosphonates after denosumab discontinuation and treatment effect heterogeneity using causal machine learning.

Discontinuation of denosumab is associated with a rebound increase in osteoporotic fracture (OF) risk, and bisphosphonates (BPs) are commonly recommended as sequential therapy to mitigate this risk. However, their real-world effectiveness-and whether certain patients benefit more than others-remains unclear. To address this question, we employed a causal machine learning approach capable of estimating heterogeneous treatment effects. We conducted a nationwide retrospective cohort study using Korean health insurance claims data and applied the causal survival forest model to evaluate the effect of BPs on recurrent OFs and explore treatment effect heterogeneity. Patients were followed for up to one year after the index date to ascertain fracture outcomes. BPs significantly reduced the risk of recurrent OFs (fracture-free probability difference: 0.032; 95% CI 0.022-0.042). Treatment effects varied by age, prior BP use, diabetes without complications, peptic ulcer disease, steroid use, and type of treatment institution. The model incorporated doubly robust adjustment for treatment assignment and time-to-event outcomes, while correcting for censoring in survival time estimation, resulting in conservative but robust effect estimates. Sensitivity analyses using traditional statistical methods supported the findings. This study provides causally informed, real-world evidence for the effectiveness of BPs after denosumab discontinuation and highlights the potential of causal machine learning to support personalized fracture prevention by identifying patient subgroups most likely to benefit.