Content area
Abstract
Background
Current risk scores for colorectal cancer (CRC) are based on demographic and behavioral factors and have limited predictive values.
Aim
To develop a novel risk prediction model for sporadic CRC using clinical and laboratory data in electronic medical records.
Methods
We conducted a nested case-control study in a UK primary care database. Cases included those with a diagnostic code of CRC, aged 50-85. Each case was matched with four controls using incidence density sampling. CRC predictors were examined using univariate conditional logistic regression. Variables with p value <0.25 in the univariate analysis were further evaluated in multivariate models using backward elimination. Discrimination was assessed using receiver operating curve. Calibration was evaluated using the McFadden's R2. Net reclassification index (NRI) associated with incorporation of laboratory results was calculated. Results were internally validated.
Results
A model similar to existing CRC prediction models which included age, sex, height, obesity, ever smoking, alcohol dependence, and previous screening colonoscopy had an AUC of 0.58 (0.57-0.59) with poor goodness of fit. A laboratory-based model including hematocrit, MCV, lymphocytes, and neutrophil-lymphocyte ratio (NLR) had an AUC of 0.76 (0.76-0.77) and a McFadden's R2 of 0.21 with a NRI of 47.6 %. A combined model including sex, hemoglobin, MCV, white blood cells, platelets, NLR, and oral hypoglycemic use had an AUC of 0.80 (0.79-0.81) with a McFadden's R2 of 0.27 and a NRI of 60.7 %. Similar results were shown in an internal validation set.
Conclusion
A laboratory-based risk model had good predictive power for sporadic CRC risk.