Use of Virus Genotypes in Machine Learning Diagnostic Prediction Models for Cervical Cancer in Women With High-Risk Human Papillomavirus Infection.

Importance: High-risk human papillomavirus (hrHPV) is recognized as an etiologic agent for cervical cancer, and hrHPV DNA testing is recommended as the preferred method of cervical cancer screening in recent World Health Organization guidelines. Cervical cancer prediction models may be useful for screening and monitoring, particularly in low-resource settings with unavailable cytological and colposcopic examination results, but previous studies did not include women infected with hrHPV. Objectives: To develop and validate a cervical cancer prediction model that includes women positive for hrHPV infection and examine whether the inclusion of HPV genotypes improves the cervical cancer prediction ability. Design, Setting, and Participants: This diagnostic study included diagnostic data from 314 587 women collected from 136 primary care centers in China between January 15, 2017, and February 28, 2018. The data set was separated geographically into data from 100 primary care centers in 6 districts for model development (training data set) and 36 centers in 3 districts for model validation. A total of 24 391 women identified with positive hrHPV test results in the cervical cancer screening program were included in the study. Data were analyzed from January 1, 2022, to July 14, 2022. Main Outcomes and Measures: Cervical intraepithelial neoplasia grade 3 or worse (CIN3+) was the primary outcome, and cervical intraepithelial neoplasia grade 2 or worse (CIN2+) was the secondary outcome. The ability of the prediction models to discriminate CIN3+ and CIN2+ was evaluated using the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, positive likelihood ratio, and negative likelihood ratio. The calibration and clinical utility of the models were assessed using calibration plots and decision curves, respectively. Results: After excluding women without screening outcomes, the study included 21 720 women (median [IQR] age, 50 [44-55] years). Of 14 553 women in the training data set, 349 (2.4%) received a diagnosis of CIN3+ and 673 (4.6%) of CIN2+. Of 7167 women in the validation set, 167 (2.3%) received a diagnosis of CIN3+ and 228 (3.2%) of CIN2+. Including HPV genotype in the model improved the AUROC by 35.9% for CIN3+ and 41.7% for CIN2+. With HPV genotype, epidemiological factors, and pelvic examination as predictors, the stacking model had an AUROC of 0.87 (95% CI, 0.84-0.90) for predicting CIN3+. The sensitivity was 80.1%, specificity was 83.4%, positive likelihood ratio was 4.83, and negative likelihood ratio was 0.24. The model for predicting CIN2+ had an AUROC of 0.85 (95% CI, 0.82-0.88), with a sensitivity of 80.4%, specificity of 81.0%, positive likelihood ratio of 4.23, and negative likelihood ratio of 0.24. The decision curve analysis indicated that the stacking model provided a superior standardized net benefit when the threshold probability for clinical decision was lower than 23% for CIN3+ and lower than 17% for CIN2+. Conclusions and Relevance: This diagnostic study found that inclusion of HPV genotypes markedly improved the ability of a stacking model to predict cervical cancer among women who tested positive for hrHPV infection. This prediction model may be an important tool for screening and monitoring cervical cancer, particularly in low-resource settings.