Comparative Utilization, Effectiveness, and Safety of Denosumab versus Zoledronic Acid in Patients with Metastatic Lung, Breast, and Prostate Cancer

Abstract

Background: Metastatic lung, breast, and prostate cancer patients (MLBPC) experience several skeletal related events (SREs), which significantly decrease survival and quality of life. Denosumab (DS) and Zoledronic acid (ZA) are recommended treatments for these patients to prevent such SREs. However, there is a lack of evidence regarding the patterns and predictors of initiation, comparative effectiveness, and safety of DS and ZA using real-world data from older MLBPC patients in the United States. Methods: We used Surveillance, Epidemiology, and End Results (SEER)-linked Medicare data from 2011-2019. We identified newly diagnosed MLBPC patients 66 years and older who initiated DS/ZA treatment within 12 months of cancer diagnosis. Demographic characteristics, healthcare utilization, comorbidities, and disease and treatment attributes of new users were evaluated during a washout period (12 months) prior to the first DS/ZA prescription. Trends in treatment initiation were assessed using Cochran-Armitage tests. To compare time to first SRE between DS and ZA new users, hazard ratios (HR) with 95% confidence intervals (CI) were obtained from both the unadjusted and adjusted (inverse probability of treatment weighting) Cox models. Competing risk regressions were performed with death as the competing risk outcome. Persistence to treatments and incidences of predefined adverse events were compared using Chi-square tests. Finally, SRE prediction machine learning models were trained and developed, including LASSO logistic regression (LR), random forest (RF), adaptive and extreme gradient boosting (AdaBoost & XGBoost), and support vector machine (SVM). Evaluation matrices including accuracy, sensitivity, specificity, area under the curve (AUC) were used. Baseline risk factors/features were analyzed using Shapley Additive Explanations (SHAP) values to rank variables important in predicting SRE. Results: In 2012-2017, DS initiation trends increased across all individual cancer cohorts as well as the overall MLBPC sample, while ZA initiation notably decreased in the metastatic breast cancer (MBC) and MLBPC cohorts (all p<0.0001). Patients more likely to initiate DS over ZA were older at diagnosis, Hispanic, single, eligible for low-income subsidies, urban residents, had multiple comorbidities, impaired renal function, and prior chemotherapy use. DS significantly delayed the time to first SRE in MBC (adjusted HR=0.81, 95% CI=0.76-0.90) and metastatic prostate cancer (MPC) (HR=0.82, 95% CI=0.70-0.96) compared to ZA, but not in metastatic lung cancer (MLC) (HR=0.90, 95% CI=0.81-1.04). The results remained similar in the competing risk models. 1-year persistence with DS was significantly higher compared to ZA (72.4% vs 66.3%; p<0.0001) in MBC, and similar in MPC and MLC. Renal failure incidences were higher in ZA users (5.45% vs 4.31%, p=0.047 in MLC; 7.42% vs 4.91%, p=0.005 in MBC; 9.81% vs 6.76%, p=0.015 in MPC), while hypocalcemia was higher in MLC (5.62% vs 2.87%, p<0.0001) and MPC patients (7.19% vs 3.76%, p=0.001) receiving DS. In predicting SRE events, AdaBoost performed the best in both the DS (AUC=0.89) and ZA (AUC=0.90) user MLC cohorts. RF was the best performing models in both MBC and MPC with AUC values around 0.91. LR and DT performed poorly compared to the ensemble models and SVM with AUC values ranging between 0.69-0.77 in MLC, 0.72-0.79 in MBC, 0.70-0.72 in MPC. Significant differences according to the DeLong’s test (all p<0.05) were observed between the best and worst performing models. SHAP analysis revealed differences in feature importance ranking in predicting the outcome between DS vs ZA users within the same disease cohort. Conclusions: This real-world study confirmed the increased trend in uptake and superior effectiveness in delaying SREs onset for DS compared to ZA in older MLBPC patients. The different safety profiles of these drugs may guide their uptake and warrant patient monitoring. The novel machine learning models performed well in predicting SRE and identified the most important predictors of SRE in older MLBPC patients receiving DS/ZA.