OBJECTIVES: To determine the leading predictors of healthcare expenditures (Medicare payments) among older adults diagnosed with incident Hodgkin's lymphoma (HL) across pre-diagnosis, treatment, and post- treatment periods.

METHODS: A retrospective research design utilizing the SEER-Medicare database was employed for older Medicare beneficiaries (age >66 years) diagnosed with incident HL between 2009 and 2017. Three phases of cancer care pre-diagnosis, treatment, and post-treatment were indexed around the cancer diagnosis date with 12 months of baseline and 12 months of follow-up period for all three phases. XGBoost, Random Forest, and Cross-Validated Linear regressions on log-transformed expenditures with interpretable SHapley Additive exPlanations (SHAP) methods, were used to identify the leading predictors (from baseline period) of Medicare payment (in the follow-up period) for each phase.

RESULTS: We analyzed 1,242 patients with HL in the pre-diagnosis phase, 902 during treatment, and 873 in the post-treatment phase. The mean annual Medicare payments were $22,623 (pre-diagnosis), $84,563 (treatment), and $36,913 (post-treatment), XGBoost outperformed other models with overall performance in predicting expenditures with R² (RMSE) values of 0.42 (1.39), 0.43 (0.56), and 0.45 (0.91) across the three phases of expenditures. Interpretable SHAP value showed baseline expenditures as a leading predictor of healthcare expenditures in the pre-diagnosis phase. Meanwhile, treatment types (immunotherapy, chemotherapy, surgery) were the leading predictors in the treatment and post-treatment phases. Furthermore, the number of prescription drugs and chronic conditions consistently ranked among top 10 predictors for Medicare expenditures.

CONCLUSIONS: As AI tools are increasingly used to develop prediction models in clinical care, policies, and population health management, AI applications should consider the disease phases in training, testing, and model implementation.