Graduate Research Assistant Auburn University Auburn, United States
Background: Bone modifying agents, such as Zoledronic Acid (ZA), are recommended to prevent skeletal-related events (SREs) in metastatic breast cancer (MBC) patients. However, the baseline risk factors of serious adverse events (SAEs) in MBC patients receiving ZA are poorly understood. There are no SAE risk-prediction tools available to assist clinicians in ZA treatment decision-making.
Objectives: To build and compare prediction models of SAEs and to evaluate importance of baseline risk factors of SAEs in MBC patients receiving ZA.
Methods: This secondary analysis used the patient-level comparator arm (ZA) data from a phase III clinical trial (NCT00321464) comparing Denosumab and ZA in MBC (obtained from Project Data Sphere). The primary outcome was a composite of SAEs, defined as an adverse event characterized by any of fatal, life threatening, in-patient hospitalization, persistent or significant disability/incapacity, congenital anomaly/birth defect, or other significant medical hazards. The SAEs were recorded from the initiation through 30 days after the last dose of ZA or end of the study, whichever is longer. Prediction models were built with different approaches, including logistic regression (LR), random forest (RF), adaptive and extreme gradient boosting (AdaBoost & XGBoost), and support vector machine (SVM). The feature set comprised of patients’ demographic, disease, and treatment related baseline variables. The area under the curve (AUC) was used to evaluate the model performance and significant risk factors were identified using feature importance evaluation.
Results: Among 683 patients treated with ZA, 343 (50%) patients developed at least one SAE. Patients’ mean (SD) age at enrollment was 60.3 (±8.9) years. The AdaBoost model performed the best among all machine learning models (AUC= 0.885), followed by the LR (AUC= 0.865), RF (AUC= 0.863), SVC (AUC= 0.861), and XGBoost (AUC= 0.855) in predicting SAEs. Among the 25 baseline variables included in the feature set, most important risk factors in predicting the SAEs were age at enrollment, BMI, ECOG performance status, comorbidity, presence of liver metastasis, as well as estrogen, progesterone, and hormone receptor status.
Conclusions: The novel machine learning models performed well in predicting SAEs and can be used as tools to identify MBC patients who are receiving ZA and at risk of SAEs in clinical settings.
