A risk prediction model to allow personalized screening for cervical cancer.

Abstract

Cervical cancer screening guidelines are in evolution. Current guidelines do not differentiate recommendations based on individual patient risk. To derive and validate a tool for predicting individualized probability of cervical intraepithelial neoplasia grade 2 or higher (CIN2+) at a single time point, based on demographic factors and medical history. The study design consisted of an observational cohort with hierarchical generalized linear regression modeling. The study was conducted in a setting of 33 primary care practices from 2004 to 2010. The participants of the study were women aged ≥ 30years. CIN2+ was the main outcome on biopsy, and the following predictors were included: age, race, marital status, insurance type, smoking history, median income based on zip code, prior human papilloma virus (HPV) results. The final dataset included 99,319 women. Of these, 745 (0.75%) had CIN2+. The multivariable model had a C-statistic of 0.81. All factors but race were independently associated with CIN2+. The model categorized women as having below-average CIN2+ risk (0.15% predicted vs. 0.12% observed risk), average CIN2+ risk (0.42% predicted vs. 0.36% observed), and above-average CIN2+ risk (1.76% predicted vs. 1.85% observed). Before screening, women at below-average risk had a risk of CIN2+ well below that of women with ASCUS and HPV negative (0.12 vs. 0.20%). A multivariable model using data from the electronic health record was able to stratify women across a 50-fold gradient of risk for CIN2+. After further validation, use of a similar model could enable more targeted cervical cancer screening.