Leading Predictors and Their Associations with Combination Pain Therapy in Older Adults with Cancer: Application of Machine Learning Approaches

OBJECTIVES: Opioid combination therapy is frequently prescribed in older adult cancer survivors despite negative outcomes. The objective of this study was to identify the leading predictors and their associations with opioid combination therapy prescribing after cancer diagnosis using interpretable machine learning approaches. METHODS: This is a retrospective longitudinal cohort of older (> 66 years old) cancer survivors (N = 2,673) diagnosed with primary and incident cancer in 2014 using the Surveillance, Epidemiology, and End Results (SEER) cancer registry linked with Medicare claims. Recursive feature elimination with random forest was used to extract the optimal number of predictors out of 119 likely ones for predictive modeling. eXtreme Gradient Boosting (XGBoost), SHapley Additive exPlanations (SHAP), and global feature importance were used to identify the leading predictors and their associations with opioid combination therapy. SAS 9.4 was used for data management and Python 3.9.7 was used for machine learning model calibration and tuning. RESULTS: Specificity (0.858), sensitivity (0.843), and area under the curve (AUC, 0.85) of our predictive model were high. Thirty-four features were included in the final predictive model. Baseline use of NSAIDs, opioids, benzodiazepines, and gabapentinoids, and chemotherapy, surgery, Complex relationships were observed between zip code percent of Hispanic and Native American residents living below poverty, care fragmentation (FCI), age at diagnosis, and opioid combination therapy. CONCLUSIONS: 1 in 3 older cancer survivors were prescribed opioid combination therapy. Patient-level baseline medication use, biological factors, cancer treatment, and zip code level social determinants were leading predictors of opioid combination therapy. Although observed relationships were complex, further analysis of predictors may help compute individual risk of patients on combination therapy, which in turn may help clinicians and policy makers utilize targeted interventions at the outset and prevent long-term effects of combination pain therapy such as prolonged and inappropriate use.