Data-Driven Clinical Phenotyping of Denosumab Exposure in a Large United States Cohort

Denosumab is a therapeutic monoclonal antibody originally developed for alleviation of osteoporosis symptoms. Due to known anti-osteoclastic effect via RANKL inhibition, indication for denosumab was expanded to include treatment for various cancers. Phenotyping for denosumab exposure occurred within a large (~50 million individuals) US-based insurance claims cohort between 2007 and 2016. Denosumab exposed patients (n=70,610) were randomly assigned to 20% training (n=14,122) and 80% analysis (n=56,488) cohorts. CCS level 2 hierarchies annotated phenotype labels (n=146) for training, analysis, and control cohorts, including data from inpatient and outpatient diagnosis (ICD9&10) codes 12 months prior and post index date. To reduce the dimension of our phenotypic feature space of 144 CCS categories, we applied an efficient and powerful gradient boosting ML algorithm named XGBoost on the training cohort to identify individuals who have been exposed to denosumab more than once. Relevant phenotypic features (n=31) were identified with 0.988 cross-validation balanced accuracy. Within the analysis cohort, our model replicated with balanced accuracy of 0.846. Several therapeutic indications were predictive of denosumab exposure such as in treatment of osteoporosis and various cancers, hypertension, eye disorders, disorders of lipid metabolism, disease of urinary system, bone disease and spondylosis (Table 1). Removing the primary indication (osteoporosis) from the denosumab model only slightly modified feature importance scores (Table 1 - FI2). Independent statistical characterization of features corroborated all associations between denosumab exposure and phenotype prevalence as significant (all adjusted p-values < 0.001). Interdependent effect between primary or secondary indications of denosumab exposure and phenotype prevalence are needed to elucidate potential disease co-occurrence or opportunities for therapeutic repositioning.