基于机器学习算法的宫颈癌辅助诊断模型

Objective To explore the performance of machine learning models for cervical cancer diagnosis and discuss the implications and feasibility of further clinical replication based on real-world data from clinical settings. Methods From January 12,2018 to December 30,2021 ,data of patients with cervical lesions were collected from Shenzhen Maternal and Child Health Hospital (n= 1 294),Affiliated Hospital of Qingdao University (n= \ 336) and Chengdu Women and Children's Central Hospital (n = 384). The training and validation sets were divided by the ratio of 8 : 2. Related cervical cancer diagnosis models were established based on five machine learning algorithms,then validate the prediction performance of the models. Results A total of 3 014 patients were included,including 767 patients (25. 45%) with cervical intraepithelial neoplasia(CIN) grade 2 or above (CIN2 + ) ,and 2 247 patients (74. 55%) with CIN grade 1 or below (<CIN2). After K-fold cross validation (K = 5) , the optimal parameters were selected to establish the risk predictive models. Performance of the predictive models were good except decision tree model,among which the performance of neural network model a-chieved the highest value of area under curve which was more than 0. 92. Conclusions Several machine learning models including annual neural network perform well in cervical cancer diagnosis. It is suggested to integrate the superiority of each model to fully exploit its clinical roles and values.