Nasopharyngeal carcinoma (NPC) is a rarely diagnosed disease in most parts of the world, with an age-standardized incidence rate usually lower than 1 case per 100,000 person-years. However, the rates are high in South China, South-East Asia, the Arctic, North Africa, and the Middle East among indigenous groups (1). Research has shown that the 5-year survival rate of stage I NPC in high-prevalence areas is as high as ≥85%. However, the 5-year survival rate of stage IV patients is only 20%, and the 5-year survival rate after distant metastasis is less than 5% (2). The proportion of early-stage patients among the patients attending the clinic is less than 30%. Most patients are already in advanced stages at the time of consultation because of the hidden location of the nasopharyngeal cavity and the lack of specificity of the early symptoms of NPC (3). Therefore, the early detection, diagnosis, and treatment of NPC are now central to its prevention and treatment. The Epstein-Barr virus (EBV) is a ubiquitous B-lymphotropic virus carried latently by almost all humans. It is usually first contracted in childhood, during which it either causes no symptoms or only mild ones (4). The International Agency for Research on Cancer has shown that EBV is an established cause of several human malignancies, including nasopharyngeal cancer (5). Approximately 95% of the global population is asymptomatically infected with EBV throughout their lives. However, EBV can be periodically reactivated in response to endogenous and environmental stresses (6, 7). During the transition of EBV from a latent infection state to a lytic replication phase, several latent and lytic gene products are expressed, contributing to epithelial cell genetic damage, immune system perturbation, and angiogenesis in the nasopharynx, thereby increasing the risk of nasopharyngeal carcinoma (8). Our previous studies have demonstrated that serum immunoglobulin A (IgA) antibodies against EBV nuclear antigen 1 (EBNA1/IgA) and viral capsid antigen (VCA/IgA) serve as serological biomarkers of EBV activation and are effective in predicting nasopharyngeal cancer (9, 10). Thus, understanding the risk factors for EBV reactivation associated with NPC is essential for preventing nasopharyngeal cancer.

It was previously found through the construction of logistic models that smoking (11), using solid fuel (12), and consuming salty food (12) were associated with EBV reactivation. However, the conventional logistic regression models fail to capture the complex network of interactions between multiple risk factors. Also, the assumption of linear additivity of the model may limit its use, leading to reduced efficacy of regression model tests and regression model failure (13–15). Furthermore, the model cannot identify direct or indirect risk factors. Accordingly, screening variables or constructing models based on the network structures is significant for analyzing the risk factors for EBV reactivation associated with NPC.

Bayesian networks (BNs), which were proposed by Judea Pearl (16), can better compensate for the shortcomings of logistic regression models. BNs comprise two components: a directed acyclic graph (DAG) that reflects the complex network of interactions among risk factors and a conditional probability table that depicts the correlation between variables (17, 18). The BN overcomes the correlation restrictions of the traditional regression models (i.e., the assumption of independence between variables), and it can also infer the probability of an unknown node when the node is known. An increasing number of studies have applied BN for predicting risk factors for diseases such as cardiovascular disease (19), stroke (20), and colorectal cancer prognosis (21). However, reports regarding the application of the BN in NPC studies, especially in the context of China, are limited. In addition, no BN modeling studies related to predicting the risk of EBV reactivation have been found. Therefore, considering the problems of the traditional prediction model, we attempted to construct the EBV reactivation prediction model based on the interactive network system and explore the predictive effect of the model, which is of great significance to public health.

Unbalanced datasets can lead to degraded model performance, so balancing the classes of EBV reactivation states is crucial for constructing BN models (22). Also, we used a nested case–control study approach and processed the data by matching 1:1 for several baseline characteristics and ill-defined variables to make the data more comparable and control for confounding bias as much as possible. Besides, in BN structure learning, we applied structural equation modeling tools, enhancing the stability of the constructed structures (23). In this study, we combined screening data from a high-prevalence area of NPC in South China, aiming to explore the risk factors for EBV reactivation by constructing a structural equation modeling (SEM)-treated BN model using a nested case–control study. The purpose was to intervene and reduce the risk of NPC in an early stage and provide a new idea for the prevention and treatment of NPC.

In previous studies, logistic regression was usually used to analyze NPC-associated risk factors for EBV reactivation, the results of which were better understood and more acceptable to researchers. However, the shortcomings of the regression model are magnified in a network framework (14). First, the assumption of linear additivity of the model may limit its use. Second, the performance of the regression model is more likely to be affected by covariance between variables in datasets with complex relationships. Third, the regression model cannot capture the complex network of interactions between multiple risk factors. Although this problem can be solved by adding interaction terms, the number of possible interactions increases exponentially with the increase in the number of variables, leading to a complex interaction process and thus reducing the effectiveness of the model test. However, the BN model can compensate for the aforementioned deficiencies of the regression model. As a result, BNs have been increasingly favored by clinical researchers in recent years as a risk assessment tool for large clinical datasets.

The BN model suggests that EBV may be reactivated due to poor daily living habits, unhealthy cooking environments, decreased host immunity, and having specific genes. Smoking directly exposes the nasopharyngeal epithelium to tobacco smoke. It has been suggested that smoking may promote nasopharyngeal carcinogenesis not only through the direct carcinogenic effect of tobacco smoke, but also indirectly by inducing EBV reactivation (35, 36). The rationale may be that exposure to nicotine promotes NPC cell proliferation and EBV replication and expression of its cleaved gene products (37). Similarly, in a prospective screening study based in southern China, Hu et al. (38) found that smoking was associated with EBV serum (VCA/IgA, EBNA1/IgA) positivity during a 3- to 5-year follow-up period, and that participants aged 60 to 69 years had a higher risk of EBNA1/IgA positivity. In addition, some studies have found an increased risk of EBV reactivation in people who have used solid fuels continuously for more than 40 years compared with those who do not cook or use cleaner fuels, and that ingestion of preserved foods may exacerbate the effect of solid fuels on the risk of EBV reactivation (39). During the combustion of solid fuels, incomplete combustion results in the emission of complex mixtures of gaseous and particulate pollutants; these products, such as polycyclic aromatic hydrocarbons (PAHs), have been found to cause immune damage (40, 41) and lead to changes in the reactivation of defense viruses (42). The nasopharynx is the first site of deposition of solid fuel combustion products, and the pro-inflammatory effect on epithelial cells may disrupt the inflammatory balance and affect the bacterial and viral flora of the nasopharynx (43).

This study also visually reasoned the constructed BN model using Netica software. Moreover, we compared the predictive performance of the two models using the test set. The results are depicted in Table 6 . The table demonstrates that the accuracy (56.36%) was consistent. The BN slightly outperformed the traditional logistic regression model in all indicators. The BN, being based on the interaction network, not only predicted the direct or indirect risk factors of the disease but also facilitated visual inference. Therefore, the BN model might be considered a novel predictive model in studying NPC-related risk factors for EBV reactivation.

NPC is a complex disease whose development is associated with genetic susceptibility, EBV infection, and environmental factors (44–46). Non-keratinizing NPC is the main type of NPC in China, and almost all patients with non-keratinizing NPC have EBV infection. Also, many other factors affect EBV reactivation. However, based on the South China Chronic Disease Cohort Study, we only considered the influence of the living habits of people in high-prevalence areas on EBV reactivation. Besides, the selection of different EBV-related markers might also have influenced the results. Under the hypothesis that antibodies against the remaining EBV proteins may perform better than those already reported, Li et al. (47) designed a peptide library containing highly sorted B-cell epitopes of EBV based on linear B-cell epitope prediction. The serum antibodies, including IgA, IgG, and Ab, to the peptide fragment P85 expressed by the BNLF2b gene differed significantly between cases and controls, with P85-Ab having the best performance (area under the curve = 0.97). This novel NPC screening marker might substantially improve the performance of screening methods for NPC populations and increase the cost-effectiveness of screening, thus shedding new light on our study of the risk factors for NPC-associated EBV reactivation.

To consider intricate interactions between factors, this study applied BN to analyze the risk factors associated with EBV reactivation in NPC and used SEM tools in constructing BN. The results of the study not only revealed the risk factors for EBV reactivation associated with NPC but also identified the direct and indirect influences on EBV reactivation and elucidated their complex interactive network relationships. In addition, the model, with Bayesian visualization of predictive inference, can raise our awareness of nasopharyngeal cancer prevention in our daily life, suggesting that we should maintain good living habits, such as quitting smoking, not staying up late, using ventilated and clean cooking environments, and exercising. This not only has a positive effect on the middle- and high-risk groups to actively participate in follow-up screening, but also helps in clinical decision-making, which can improve the early detection rate of nasopharyngeal cancer, which is of great significance to the prognosis of nasopharyngeal cancer patients, and provides a new way of thinking about the prevention of nasopharyngeal cancer. However, this study had some limitations. First, some of the data in this study were collected through structured questionnaires, and participants are also required to review their lifestyle habits from the past six months or even several years ago, such as eating pickled foods, smoking and drinking. So recall bias was inevitable. Second, the directed edges in BNs did not indicate causal relationships between connected nodes. Instead, they represented probabilistic dependencies. Third, most of the invalid variables were found in the construction of the logistic regression model, which could be partly explained by undifferentiated misclassification. In addition, we did not find an association between the consumption of salty food and EBV reactivation. This lack of association could be attributed to the extended duration of the NPC screening study in Sihui City, leading to a gradual shift in local residents’ dietary habits. Fourth, this study was only based on the investigation in Sihui City, Guangdong Province, and the extrapolation of data was affected. Fifth, this study did not attempt to use other antibody biomarkers of EBV reactivation for comparison.

EBV reactivation is also associated with a number of diseases. The present study is only a preliminary study of the relationship between lifestyle habits and EBV reactivation, and subsequent studies could further incorporate a number of chronic non-communicable diseases (e.g., hypertension, diabetes mellitus, stroke, etc.) to explore the relationship between them and EBV reactivation. Moreover, We plan to conduct multicenter studies with more EBV reactivation antibody biomarkers and more variable parameters in the future to construct more accurate and reliable BN models of EBV reactivation risk factors associated with NPC. And we will try to construct BN prediction models for other diseases in order to build a more mature and stable prediction network to provide methodological references for the prevention of more diseases.

This study demonstrated that BN based on SEM balanced mean TABU and MMHC combined algorithm could not only realize the complex network relationship between risk factors and NPC-associated EBV reactivation but also made it possible to predict the risk of EBV reactivation. The findings provided a scientific idea for preventing and treating NPC, contributing to the reduction of its prevalence. The specific findings were as follows: (1) The logistic regression model showed that NPC-related EBV reactivation was significantly associated with age and smoking status, whereas stable environmental factors had no association, suggesting that environmental factors might in turn influence NPC through other mechanisms. (2)The BN model of EBV reactivation was constructed using 8 nodes and 10 directed edges. Age and smoking status were the direct influencing factors for EBV reactivation. In contrast, sex, education level, tea drinking, cooking, and family history of cancer constituted the indirect influencing factors of EBV reactivation. (3) The BN using the SEM balanced average TABU and MMHC combined algorithm could achieve the probabilistic inference of unknown nodes through known nodes and flexibly demonstrated the influence of a risk factor on EBV reactivation. (4) Overall, the BN with SEM balanced average TABU and MMHC combined algorithm could be used as a new model to predict risk factors for NPC-associated EBV reactivation, with a broad prospect in the clinical practice of NPC.