Risk prediction model for precancerous cervical lesions by integrating multimodal data: development and validation of a nomogram based on modified Poisson regression

Cervical cancer, as the fourth most common malignancy among women worldwide, poses a serious threat to global women’s health. Early identification and intervention of high-grade squamous intraepithelial lesions (HSIL) are of great significance for the prevention and control of cervical cancer and precancerous lesions. Current screening methods are relatively singular, leading to a coexistence of missed diagnosis risk and overtreatment. How to achieve individualized precision diagnosis and treatment for lesions staged ≥HSIL has become a vital clinical issue. This study aims to integrate multimodal information, including demographics, social behavior, molecular biology, and radiology, identify independent risk factors for high-grade lesions, and construct an individualized risk prediction nomogram model based on modified Poisson regression (MPR) to achieve precision diagnosis and treatment and individual stratification of cervical cancer and precancerous lesions.

A retrospective observational study was conducted on 741 patients who underwent biopsies at Cervical Disease Diagnosis and Treatment Center of Women and Children’s Hospital of Chongqing Medical University between July 2022 and July 2024. Based on the histopathological gold standards, they were divided into a low-grade squamous intraepithelial lesion group (≤LSIL group, n=484, including 206 cases of chronic cervicitis and 278 cases of LSIL) and a high-grade squamous intraepithelial lesion group (≥HSIL group, n=257, including 243 cases of HSIL, 8 cases of cervical adenocarcinoma in situ, and 6 cases of cervical cancer). Multimodal data were systematically collected, including age, parity, human papillomavirus (HPV) vaccination history, thin-prep cytologic test (TCT) results, HPV genotypes, duration of persistent HPV infection, and colposcopic impression (such as acetowhite epithelial thickness and vascular patterns). Independent samples t-test (for continuous variables with normal distribution and homogeneity of variance) and Pearson Chi-square test (Fisher’s exact test when theoretical frequency <5) were used for univariate analysis. After incorporating variables with P<0.1 from univariate analysis into the candidate set and after forcibly controlling confounders such as age, gravidity and parity, a backward elimination method was used within the MPR framework to identify independent risk factors and calculate risk ratios (RR) with 95%CI. A nomogram was constructed using R software, and its discrimination and consistency were evaluated via the AUC value of ROC curve analysis and calibration curve analysis (Bootstrap method, B=500).

Univariate analysis showed statistically significant differences between the 2 groups in terms of (t=-2.264, P=0.024), gravidity (t=-3.527, P<0.001), parity (t=-3.222, P<0.001), contraceptive methods (χ² =10.712, P=0.013), TCT results (χ² =44.883, P<0.001), HPV genotypes (χ²=15.692, P=0.001), cumulative duration of HPV infection (χ²=11.670, P=0.009), HPV vaccination status (χ²=15.685, P<0.001), colposcopic impression (χ² =87.284, P<0.001), acetowhite findings (χ²=72.771, P<0.001), vascular patterns (punctation/mosaicism) (χ²=67.099, P<0.001), and type of cervical transformation zone (χ² =20.316, P<0.001). Multivariate MPR identified 5 independent predictors: ① age (RR=1.015, 95%CI: 1.000 to 1.030, P=0.048); ② gravidity (RR=1.066, 95%CI: 1.002 to 1.134, P=0.043); ③ duration of persistent HPV infection (RR=1.405, 95%CI: 1.010 to 1.956, P=0.044); ④ colposcopic impression (RR=2.108, 95%CI: 1.545 to 2.878, P<0.001); and ⑤ HPV vaccination (protective factor, RR=0.631, 95%CI: 0.510 to 0.781, P<0.001). The nomogram model constructed based on these factors achieved an AUC value of 0.74 (95%CI: 0.702 to 0.778) in predicting ≥HSIL risk. The calibration curve showed good consistency between predicted probabilities and actual observed values, with a calibration slope of 0.9580 and a Brier score of 0.1914.

Our nomogram model constructed by integrating multimodal information demonstrates good discrimination and clinical reliability. It enables intuitive and visualized quantification of individualized risk, facilitating optimized efficiency in cervical cancer prevention and control.