Development of a Machine Learning Model for Nasopharyngeal Carcinoma Screening Based on Tongue Imaging

Overview

Nasopharyngeal cancer is common in China, Southeast Asia, and North Africa, and is usually associated with Epstein-Barr virus (EBV) infection. Using EBV specific antibodies or EBV DNA screening can increase the proportion of patients diagnosed with early nasopharyngeal carcinoma from approximately 20% to over 70%. However, the application of nasopharyngeal carcinoma screening in clinical practice is hindered by low positive predictive values, even in areas where the EB virus is prevalent in China, the positive predictive value is only 4.8%. Therefore, there is an urgent need to identify new biomarkers or strategies with high sensitivity and positive predictive value for nasopharyngeal carcinoma screening. A study published in the Lancet sub journal 《eClinicalMedicine》 in 2023 showed that a tongue image model based on machine learning can serve as a stable diagnostic method for gastric cancer (AUC=0.89), and has been clinically validated in multiple centers. This study inspires researchers to introduce artificial intelligence machine learning technology into the diagnosis and treatment of nasopharyngeal cancer. In summary, this plan explores the establishment of tongue image machine learning models in nasopharyngeal carcinoma patients to help improve the positive predictive value of nasopharyngeal carcinoma screening.

Nasopharyngeal cancer is common in China, Southeast Asia, and North Africa, and is generally associated with Epstein-Barr virus (EBV) infection. Using EBV specific antibodies or EBV DNA screening can increase the proportion of patients diagnosed with early nasopharyngeal carcinoma from approximately 20% to over 70%. In previous studies, researchers found that participants who underwent screening were more likely to achieve long-term survival after being diagnosed with nasopharyngeal carcinoma compared to the control group, and the risk of nasopharyngeal carcinoma specific death was lower among screened patients (relative risk 0.22). However, the application of nasopharyngeal carcinoma screening in clinical practice is hindered by low positive predictive values, even in areas where the EB virus is prevalent in China, the positive predictive value is only 4.8%. More than 95% of high-risk participants identified through primary serological screening underwent unnecessary and time-consuming clinical examinations and follow-up. The combination of various biomarkers, multi-step screening, and identification of new biomarkers are used to improve the performance of nasopharyngeal cancer screening strategies. However, the progress achieved so far is still unsatisfactory, characterized by low sensitivity, complex operation, or high cost. Therefore, there is an urgent need to identify new biomarkers or strategies with high sensitivity and positive predictive value for nasopharyngeal carcinoma screening. In 《The New England Journal of Medicine》 in 2023, Professor Xia Ningshao's team reported on the identification and validation of anti BNLF2 total antibody (P85Ab) as a new serological biomarker for nasopharyngeal cancer screening.The sensitivity of P85-Ab nasopharyngeal carcinoma is 97.9%, with a positive predictive value of 10.0%. Furthermore, on the basis of P85-Ab positivity, if further detection of EB double antibodies (EBV nuclear antigen 1 \[EBNA1\]-IgA and EBV-specific viral capsid antigen \[VCA\]-IgA) is carried out, intermediate or medium high risk individuals with EB double antibodies can undergo nasopharyngoscopy examination, which can increase the positive predictive value of nasopharyngeal carcinoma screening to 29.6% -44.6%, that is, for every 2-3 nasopharyngoscopes performed, one case of nasopharyngeal carcinoma can be diagnosed. The sensitivity of this study is very high, but the positive predictive value is only 10%. Even when combined with traditional EB dual antibody monitoring and nasal endoscopy, one-third to one-half of non nasopharyngeal carcinoma patients still undergo unnecessary and time-consuming clinical examinations. Therefore, it is still necessary to explore simple and cost-effective methods to improve the strategy of positive predictive value for nasopharyngeal carcinoma screening. A study published in the Lancet sub journal 《eClinicalMedicine》 in 2023 showed that a tongue image model based on machine learning can serve as a stable diagnostic method for gastric cancer (AUC=0.89), and has been clinically validated in multiple centers. This study inspires researchers to introduce artificial intelligence machine learning technology into the diagnosis and treatment of nasopharyngeal cancer. In summary, this plan explores the establishment of tongue image machine learning models in nasopharyngeal carcinoma patients to help improve the positive predictive value of nasopharyngeal carcinoma screening.