Diagnostic potential of serum biomarkers for colorectal cancer statistical methods on retrospective data

Although the current colorectal cancer (CRC) screening strategy involves faecal immunochemical test combined with invasive colonoscopy, there is growing interest in developing non-invasive tests. In particular, for patients unwilling or for whom colonoscopy is contraindicated, the integration of blood-based markers with the fecal immunochemical test for CRC diagnosis is currently under investigation (Loktionov,2020 ). We analyzed C-reactive protein (CRP), lipopolysaccharide-binding protein (LBP) —the protein driver of the host response to lipopolysaccharide— and zonulin —the protein responsible for regulating tight junctions— to assess their association with CRC risk. We applied the statistical methods developed by Huang and Pepe (Huang,2007; Huang,2010 ) on predictiveness curves for continuous biomarkers to evaluate their diagnostic potential for CRC. We used data from an Italian multicentric case-control study comprising 75 histologically confirmed CRC cases and 146 controls. Serum CRP, zonulin, and LBP were assessed by ELISA kit. We used logistic regression models adjusted for centre, sex, age and education to estimate the odds ratio (OR) of CRC for tertiles of CRP, zonulin and LBP. We examined the potential interaction of CRP and LBP with levels of zonulin. We applied statistical techniques to evaluate the predictiveness of these biomarkers on data from a case-control study. We estimated a risk function using logistic regression models, including an offset term based on a literature-derived estimate of CRC prevalence p in the population our sample came from. Given that FCa and FCo are cumulative distribution functions (CDF) of a continuous marker Y , respectively, between cases and controls, the CDF of Y could be estimated as F∗ = pFCa + (1 − p)FCo. We generated the predictiveness curves R(v) = P[D = 1|Y = F∗−1(v)] to obtain a CRC risk distribution in our population, with v ∈ (0, 1), referring to the marker percentiles. To assess the joint diagnostic potential of zonulin and CRP, as well as zonulin and LBP, we extended this statistical technique to two markers. We constructed a predictiveness surface R(v, w) = P[D = 1|Y = F∗−1(v), Z = G∗−1(w)], with G∗the estimated CDF of the second marker Z and w ∈ (0, 1). To construct surfaces, we estimated risk function trough logistic models both with and without interaction term. A positive association was found between CRC risk and CRP (OR for the highest vs lowest tertile = 2.69, 95% Confidence Interval [CI]: 1.28 − 5.64), zonulin (OR= 2.10, 95% CI 1.04 − 4.26), and LBP (OR= 2.35, 95% CI: 1.14 − 4.85). The OR of CRC for the highest versus lowest tertiles of both CRP and zonulin was 5.19, while for LBP and zonulin, it was 5.15, without significant interaction (p=0.96 and p=0.11, respecitively). Our biomarkers exhibited increasing predictiveness curves, which intersected and exceeded the prevalence line —indicating that the markers became informative— for percentiles approximately greater than or equal to 85th for CRP, 65th for zonulin, and 75th for LBP. For each biomarker, the curve became steeper from the 95th percentile onward, reaching an estimated risk of approximately 10% at the 98th percentile. Considering zonulin and LBP together, we obtained a predictiveness surface that intersected and exceeded the prevalence plan for cut-off lower than those found for analysis of each marker, approximately at the 60th percentile of each. Approximately 10% of risk was found at the 90th percentile of both markers. The performance of the surface constructed including the interaction was better. We constructed predictiveness curves and surfaces for CRP, zonulin, and LBP, although the diagnostic capacities of our biomarkers were limited. Combining two biomarkers, particularly with the inclusion of an interaction term, resulted in improved performance. However, larger studies are needed to better understand the diagnostic potential of our biomarkers for CRC. New models can be applied to optimize the performance of continuous biomarkers.