Validating the use of machine-learning cancer staging algorithms for Medicare cost analyses

e13548 Background: Administrative claims provide valuable real-world insight into the care of cancer patients; however, claims data lacks cancer stage information. This limitation constrains research on the value of early diagnosis and treatment as well as on the costs and savings associated with increased cancer screening. In prior work, our team used the SEER-Medicare data to develop machine learning (ML) algorithms to stage non-small cell lung (NSCLC), colon (CC), and rectal (RC) cancer patients using clinical flags derived from claims data. These algorithms were 69% (RC), 78% (NSCLC), and 83% (CC) accurate at matching incident cancer patients with their SEER-recorded AJCC stage (SEER-stage) at diagnosis. This work sought to test whether these ML algorithms are sufficiently accurate for use in claim cost analyses. Methods: Incident NSCLC, CC, and RC patients were identified using 2016-2017 SEER-Medicare data and assigned a cancer stage using a claims-based predictive multinomial logistic regression model (R Statistical Software - v4.1.2 R Core Team 2021; nnet package - Venables and Ripley 2002). Patients’ cumulative medical and pharmacy costs were summarized for 12 months starting with patients’ index month. Patients’ Medicare index month was set equal to the month of their first claim with a cancer diagnosis. Patients’ SEER index month was set equal to the diagnosis month associated with their incident tumor record in SEER. Patients with each cancer type were then grouped two ways - by ML-stage and by SEER-stage. Median patient costs were compared between stage groups for each cancer type and differences tested for statistical significance using Wilcoxon Rank-Sum Testing. Results: For NSCLC and CC, raw differences in median 12-month claim costs between the ML- and SEER-stage cohorts were small (1%-4%). Cost differences for RC were higher (7%-17%). ML and SEER costs were not significantly different (p > 0.05) between later-stage cohorts (NSCLC stages 3 and 4, CC stages 2C/3 and 4, and RC stage 4); however, early-stage groups were always significantly different (p < 0.05). Conclusions: Although costs were not statistically equivalent across all stage groups, the similarity of ML and SEER costs across higher-stage cohorts and small raw differences in median costs for each NSCLC and CC group suggests that ML algorithms with higher accuracy may be used to develop costs from administrative data for stage shift modeling and cost tradeoff analyses. [Table: see text]