RISK PREDICTION FOR COLORECTAL CANCER USING QUANTITATIVE FECAL IMMUNOCHEMICAL TEST RESULTS — A STATISTICAL SOLUTION TO THE COLONOSCOPY CAPACITY ISSUE

Background
Colorectal cancer (CRC) is the third leading cause of cancer-related death in the United States. CRC screening can reduce mortality by removal of precancerous lesions during colonoscopy and increasing detection of CRC at an early stage. Fecal immunochemical test (FIT) is a recommended approach to CRC screening but requires annual testing. Currently, each screening episode is treated as an independent event, considering only the most recent measurement when labeling the test as normal or abnormal. A more efficient approach would combine longitudinal screening results statistically. This has been shown to improve performance in screening for other cancers but has not yet been applied to FIT results.

Methods We examined FIT results accumulated by participants in a large, mailed-outreach program from 11/1/2013 to 12/31/2021. We leverage the pattern mixture model (PMM), originally designed for risk prediction, and apply it to the early detection of advanced neoplasia (AN). Here we model cases and controls separately, where cases are defined as those patients diagnosed with AN and controls are those without AN. Model results are then used to derive the optimal combination of quantitative FIT measurements to yield the conditional probability for AN. We model the current screening approach using the most recent FIT test in a univariate logistic regression to predict AN. We then compare PMM performance to the current screening approach using single threshold qualitative assessment by the area under the ROC curve (AUC). Model AUCs are reported as range due to the sensitivity analysis performed, reflecting the various methods used to address quantitative FIT measurements equal to zero and accommodate the necessary log transformation used in response to the right skewedness of the data.

Results We analyzed 29,398 FIT results for 16,637 patients. Participants had between 1 and 8 repeated FIT results. Median age was 57 years, 31.9% were male and the cohort was racially diverse (49.1% White, 18.0% Black, of which 38.7% were Hispanic) (Table 1). Of the 1301 patients with a positive FIT, as defined by the current screening approach, only 225 had AN, resulting in a true positive rate of 17.3%. Figure 1 illustrates how FIT score trajectories differ markedly between those participants with and without AN, suggesting an advantage to the PMM approach. PMM approach consistently outperformed the single-test logistic regression, with AUCs ranging from 0.925 to 0.956 compared to 0.905 to 0.927 for the standard approach.

Conclusions
We recommend a longitudinal approach to evaluating FIT scores to more efficiently triage patients for colonoscopy scheduling, thereby addressing the resource problem of limited colonoscopy capacity.