Using Machine Learning to Understand Predictors of Frequent Surveillance Testing in Colorectal Cancer Patients

OBJECTIVES: Colorectal cancer (CRC) survivors completing curative treatment require post-treatment surveillance for potential disease recurrence. Clinical guidelines recommend carcinoembryonic antigen (CEA) testing every three to six months over a five-year period, with little guidance on who should return for testing within three months versus waiting longer. This study aims to identify key patient and physician characteristics that predict the likelihood of frequent CEA testing in current practice. Understanding real-world surveillance patterns may help enhance post-treatment care for CRC survivors.

METHODS: We analyzed EHR data on 2,489 adult patients diagnosed with AJCC stage I-III CRC between 2008 and 2013, treating multiple testing visits during surveillance as a recurrent event. At each visit, we defined the binary outcome of whether the next test occurred within three months and used baseline and updated patient variables, such as prior and current CEA values, as candidate predictors. We used a Lasso regression model for variable selection among 191 patient and physician characteristics.

RESULTS: The Lasso model showed moderate discriminative ability in predicting CEA testing within three months (AUC = 0.72). Patients were more likely to undergo testing within three months if they had a longer time since previous visit, high prior CEA value, high current CEA value, and self-reported Black ethnicity. Key physician predictors included provider specialty and medical service area, with patients who saw oncologists being more likely to be tested frequently than those who saw primary care providers.

CONCLUSIONS: Longer time from previous visit and high prior and current CEA values appropriately suggested higher likelihood of testing within three months; however, further research is needed to understand why Black patients and patients seeing oncologists underwent more frequent testing. This research sheds light on current surveillance patterns and may be helpful in designing interventions that deliver appropriate, value-based post-treatment care for CRC survivors.