OBJECTIVES: We aimed to estimate the average treatment effect (ATE) of bisphosphonates (BPs) on the risk of recurrent osteoporotic fractures (OFs) following denosumab discontinuation using causal survival forest (CSF) and explore heterogeneity within the ATE.

METHODS: We conducted retrospectively analysis from January 1, 2015, to May 31, 2022, utilizing the Health Insurance Review and Assessment database, representative of the Korean. We focused on patients (≥50 years) diagnosed with OFs in 2015, discontinuing denosumab between January 1, 2015, and May 31, 2021. The study population comprised BP cohort, prescribed BP at least once within nine months from the last denosumab and non-BP cohort. BP cohort’s index date was the first BP prescription, while it was matched based on individual characteristics for non-BP cohort. Sixty covariates, including age, sex, medication history, and comorbidities, were considered. The outcome was time-to-recurrent OFs within one year after the index date. Using CSF, we estimated ATE, survival probability differences between the two cohorts. Heterogeneity within the ATE was assessed through linearity regression based on covariates.

RESULTS: The study encompassed 31,217 patients, comprising 7,138 in the BP cohort(22.9%) and 24,079 in the non-BP cohort(77.1%), with a mean age of 74.8 ± 8.7 years. The CSF estimated a significant difference in survival probability of 0.032 [95% CI: 0.022, 0.042; p < 0.001]. Heterogeneity in ATE was identified based on the number of prior denosumab administration, duration of prior BPs use, history of steroid use, and history of peptic ulcer disease.

CONCLUSIONS: We demonstrated a significant association between BPs and a reduced risk of recurrent OFs following the discontinuation of denosumab. Moreover, it highlights the potential of employing causal machine learning in health-related outcomes research, introducing a novel methodology for evaluating treatment effects in clinical practice. However, further research is essential to generalize these clinical and methodological findings.