He B J, Chen W Y, Liu L L, Zhu H Y, Cheng H Z, Zhang Y X, Wang S F, Zhan S Y
Department of Epidemiology and Biostatistics, School of Public Health, Peking University, Beijing 100191, China.
School of Public Health, Peking University, Beijing 100191, China.
Zhonghua Liu Xing Bing Xue Za Zhi. 2021 Oct 10;42(10):1855-1862. doi: 10.3760/cma.j.cn112338-20200806-01031.
To systematically summarize and assess risk prediction models for occurrence of cervical cancer and to provide evidence for selecting the most reliable model for practice, and guide cervical cancer screening. Two groups of keywords related to cervical cancer and risk prediction model were searched on Chinese databases (CNKI, and Wanfang) and English databases (PubMed, Embase, and Cochrane Library). Original articles that developed or validated risk prediction models and published before November 21, 2019, were selected. Information form was created based on the CHARMS checklist. The PROBAST was used to assess the risk of bias. 12 eligible articles were identified, describing 15 prediction models, of which five were established in China. The predicted outcomes included multiple stages from cervical precancerous lesions to cancer occurrence, ., abnormal Pap smear (1), occurrence or recurrence of CIN (9), and occurrence of cervical cancer (5), . The most frequently used predictors were HPV infection (12), age (7), smoking (5), and education (5). There were two models using machine learning to develop models. In terms of model performance, the discrimination ranged from 0.53 to 0.87, while only two models assessed the calibration correctly. Only two models were externally validated in Taiwan of China, using people in different periods. All of the models were at high risk of bias, especially in the analysis domain. The problems were concentrated in the improper handling of missing data (13), preliminary evaluation of model performance (13), improper use of internal validation (12), and insufficient sample size (11). In addition, the problems of inconsistency measurements of predictors and outcomes (8) and the flawed report of the use of blindness for outcome measures (8) were also severe. Compared with the other models, the Rothberg (2018) model had relatively high quality. There are a certain number of cervical cancer risk prediction models, but the quality is poor. It is urgent to improve the measurement of predictors and outcomes, the statistical analysis details such as handling missing data and evaluation of model performance and externally validate existing models to better guide screening.
系统总结和评估宫颈癌发生风险预测模型，为选择最可靠的模型用于实践提供依据，并指导宫颈癌筛查。在中国数据库（知网、万方）和英文数据库（PubMed、Embase、Cochrane图书馆）上搜索了两组与宫颈癌和风险预测模型相关的关键词。选择了在2019年11月21日前发表的开发或验证风险预测模型的原始文章。根据CHARM清单创建信息表。使用PROBAST评估偏倚风险。共识别出12篇符合条件的文章，描述了15种预测模型，其中5种是在中国建立的。预测结果包括从宫颈癌前病变到癌症发生的多个阶段，即巴氏涂片异常（1种）、CIN发生或复发（9种）以及宫颈癌发生（5种）。最常用的预测因素是HPV感染（12种）、年龄（7种）、吸烟（5种）和教育程度（5种）。有两种模型使用机器学习开发。在模型性能方面，辨别力范围为0.53至0.87，而只有两种模型正确评估了校准。只有两种模型在中国台湾进行了外部验证，使用了不同时期的人群。所有模型都存在较高的偏倚风险，尤其是在分析领域。问题集中在缺失数据处理不当（13种）、模型性能的初步评估（13种）、内部验证使用不当（12种）和样本量不足（11种）。此外，预测因素和结果测量不一致的问题（8种）以及结局测量中使用盲法的报告存在缺陷的问题（8种）也很严重。与其他模型相比，Rothberg（2018）模型质量相对较高。有一定数量的宫颈癌风险预测模型，但质量较差。迫切需要改进预测因素和结果的测量、处理缺失数据和评估模型性能等统计分析细节，并对现有模型进行外部验证，以更好地指导筛查。
