Assigning cancer stage to patients using administrative claims data: an example for colon and rectal cancers.

[OBJECTIVES] This study sought to design predictive machine learning models to assign stage at diagnosis to colorectal cancer patients using only administrative claims data. The models were subsequently applied to Medicare fee for service (FFS) and commercially insured populations to assess the relationship between healthcare costs and stage at diagnosis.

[METHODS] FFS patients diagnosed with colon or rectal cancer tumors in 2016 or 2017 were identified in the Surveillance, Epidemiology and End Results (SEER)-Medicare linked dataset. Patients were assigned a cancer stage via multinomial logistic regression using treatment and demographic variables derived from claims data. The models were applied to data from colorectal cancer patients observed in a commercially insured population, and cumulative allowed costs for one to three years post-diagnosis were summarized by stage and compared across insurance coverages (FFS and commercial).

[RESULTS] The predictive cancer staging models achieved an overall staging accuracy of 83% (colon) and 69% (rectal) for newly diagnosed patients. When used to examine healthcare costs by stage, costs for patients diagnosed at stages IIC/III/IV were 2-4x higher than those of patients diagnosed at stages 0/I/IIA/IIB. Average costs for commercially insured patients were 30-40% higher than Medicare patients' costs.

[CONCLUSIONS] Novel multinomial logistic regression models successfully predicted incident colorectal cancer patients' stages at diagnosis using administrative claims data alone. When applied to FFS and commercial populations, use of model-predicted staging confirmed that colorectal cancer patients' costs increase substantially as stage at diagnosis increases. This suggests earlier diagnosis of colorectal cancer could reduce treatment costs.