A personalized computational model predicts cancer risk level of oral potentially malignant disorders and its web application for promotion of non‐invasive screening

• Published in: Journal of oral pathology & medicine 2020-05, Vol.49 (5), p.417-426
• Main Authors: Wang, Xiangjian, Yang, Jin, Wei, Changlei, Zhou, Gang, Wu, Lanyan, Gao, Qinghong, He, Xin, Shi, Jiahong, Mei, Yingying, Liu, Ying, Shi, Xueke, Wu, Fanglong, Luo, Jingjing, Guo, Yiqing, Zhou, Qizhi, Yin, Jiaxin, Hu, Tao, Lin, Mei, Liang, Zhi, Zhou, Hongmei
• Format: Article
• Language: English
• Subjects: Biopsy; Cancer; cancer risk level prediction; Computer applications; Dentistry; Learning algorithms; Machine learning; non‐invasive screening; oral potentially malignant disorders; Patients; personalized model; web application
• ISSN: 0904-2512; 1600-0714
• DOI: 10.1111/jop.12983

Background Despite their high accuracy to recognize oral potentially malignant disorders (OPMDs) with cancer risk, non‐invasive oral assays are poor in discerning whether the risk is high or low. However, it is critical to identify the risk levels, since high‐risk patients need active intervention, while low‐risk ones simply need to be follow‐up. This study aimed at developing a personalized computational model to predict cancer risk level of OPMDs and explore its potential web application in OPMDs screening. Methods Each enrolled patient was subjected to the following procedure: personal information collection, non‐invasive oral examination, oral tissue biopsy and histopathological analysis, treatment, and follow‐up. Patients were randomly divided into a training set (N = 159) and a test set (N = 107). Random forest was used to establish classification models. A baseline model (model‐B) and a personalized model (model‐P) were created. The former used the non‐invasive scores only, while the latter was incremented with appropriate personal features. Results We compared the respective performance of cancer risk level prediction by model‐B, model‐P, and clinical experts. Our data suggested that all three have a similar level of specificity around 90%. In contrast, the sensitivity of model‐P is beyond 80% and superior to the other two. The improvement of sensitivity by model‐P reduced the misclassification of high‐risk patients as low‐risk ones. We deployed model‐P in web.opmd-risk.com, which can be freely and conveniently accessed. Conclusion We have proposed a novel machine‐learning model for precise and cost‐effective OPMDs screening, which integrates clinical examinations, machine learning, and information technology.