Background
Gastric cancer (GC) is the fifth most frequently diagnosed cancer and the fourth leading cause of cancer death worldwide [
1]. Nearly three-quarters of all new cases and deaths from GC occur in Asian countries, including China, Japan, and Korea [
2]. However, among these three countries, the incidence rates of GC are higher in Japan and Korea, whereas the mortality rate is higher in China [
2]. This disparity is mainly due to the differences in the early detection of GC, leading to high 5-year survival rates in Japan (60.3%) and Korea (68.9%) but a much lower rate in China (35.9%) [
3]. Therefore, screening is critical to improve early detection and treatment and to ultimately reduce GC mortality in China.
Endoscopic screening has been shown to reduce GC mortality by 40% in Asian countries [
4]. In Japan, a national GC screening was implemented in 1983, and endoscopic screening was recommended for individuals aged 50 years and older [
5]. In Korea, a nationwide screening program was launched in 1999 to screen individuals aged 40 years and older for GC by either upper endoscopy or upper gastrointestinal series examinations [
6]. However, in China, there is still no national screening policy or program, because screening in a huge population is cost-prohibitive and requires the capabilities of local doctors and access to available technology. Recently, an endoscopic screening program showed significant reductions in both incidence and mortality of upper gastrointestinal cancer among local permanent residents aged between 40 and 69 years from six high-risk areas of China [
7]. Thus, tailored endoscopic screening in high-risk populations represents a more feasible and cost-effective approach in China.
Currently, the consensus on the GC screening in China is to target the subpopulation aged 40 years or older [
8]. However, more than 300 million people in China meet the criteria of the consensus, making it impracticable at present [
9]. Several prescreening tools prior to a gastroscopy have been developed for GC, which usually combine
Helicobacter pylori (
H. pylori) serology tests, serum pepsinogen (PG) I and PG II, and gastrin-17 (G-17) levels [
9‐
11]. Although these tools are effective in identifying high-risk individuals for GC, these serum biomarkers need to be measured in hospitals or other professional institutions and have inconsistent performance in different populations, leading to additional costs and increased difficulty in screening settings.
A number of risk prediction models based on traditional risk factors have been developed for breast cancer [
12], colorectal cancer [
13], and lung cancer [
14]. However, to date, very few risk prediction models have been developed for GC [
9,
11,
15,
16], and none has been used for organized screening programs largely due to the lack of external validations required before translation into practice. Herein, leveraging a nationwide prospective cohort, the China Kadoorie Biobank (CKB), we developed a GC risk score (GCRS) based on examination-free variables from questionnaires. We further validated its effectiveness and usefulness in an independent prospective cohort and a real-world cross-sectional endoscopy screening program, respectively. Finally, based on the GCRS, we developed an online tool, named Risk Evaluation for Stomach Cancer by Yourself (RESCUE) [
17], to be utilized by the public for GC risk assessment.
Methods
Study design and subjects
A three-stage study design was used in the present study (Additional file
1: Fig. S1). In the first stage, the CKB, the largest prospective cohort in China, was used to develop the GCRS. Details of the CKB have been described previously [
18,
19]. Briefly, a total of 512,714 participants (aged 30–79 years) were recruited from 10 (5 urban and 5 rural) areas between June 2004 and July 2008. In the present study, we excluded those with GC diagnosed at baseline (
n = 264), outside the target age range of 40–75 years old (
n = 81,047), or with missing covariates (
n = 15,060) and finally included 416,343 eligible subjects in the construction of the GCRS.
In the second stage, the GCRS was validated in an independent prospective cohort from Changzhou of Jiangsu province, China. A total of 20,803 permanent residents aged 35 years or older were enrolled between April 2004 and August 2005 [
20]. In this cohort, a total of 13,982 eligible participants remained after excluding those diagnosed with GC at baseline (
n = 42), outside the age range of 40–75 years old (
n = 6520), with missing covariates (
n = 214), or loss to follow-up (
n = 45).
In the third stage, the GCRS was evaluated in an ongoing upper gastrointestinal disease screening program from Yangzhou of Jiangsu province, China. Permanent residents aged between 40 and 75 years old from eight administrative communities were invited to participate in the program since December 2017. Until March 2022, a total of 5718 participants were recruited. After a face-to-face questionnaire interview and physical examinations, each participant also underwent upper gastrointestinal endoscopy and pathological biopsy. Besides the aforementioned exclusion criteria (n = 117), those who lacked pathological biopsy reports (n = 175) or had missing covariates (n = 78) were also excluded, leaving a total of 5348 participants for the final analysis.
All participants signed a written informed consent on enrollment. Further information on the study details can be found in the Additional file
1: Appendix 1.0 [
18‐
20].
Procedures
Self-reported information on demographic characteristics, lifestyle, dietary pattern, and medical history was obtained through similar questionnaires in the CKB cohort, the Changzhou cohort, and the Yangzhou screening program. In preliminary analyses of the CKB cohort, the predefined candidate predictors for model derivation were included according to the following criteria: (1) established or probable risk factors of gastric cancer through systematic literature review, (2) established in reported gastric cancer risk prediction models, and (3) available in questionnaires of the CKB. As a result, age [
9,
15,
16]; sex [
15,
16]; education [
21]; smoking [
15,
22]; alcohol drinking [
22]; consumption of fresh fruits and vegetables [
23]; salty food intake [
9]; physical activity [
22], body mass index (BMI) [
22]; medical history of physician-diagnosed cancer [
24,
25], gastrointestinal diseases (e.g., peptic ulcer) [
26,
27], or diabetes [
28]; and family history of cancer in first-degree relatives [
22] were identified as candidate predictors.
The primary outcome of the CKB and Changzhou cohort analysis was incident GC as classified by the 10th Revision of the International Classification of Diseases (ICD-10 codes C16). The complete follow-up for the CKB was updated on December 31, 2016. For the Changzhou cohort, three follow-up investigations were performed in 2008–2009, 2012–2013, and 2018–2019, separately. In the Yangzhou screening program, the primary outcome was histopathologically diagnosed GC, and the secondary outcomes included dysplasia (DYS), intestinal metaplasia (IM), atrophic gastritis (AG), and chronic superficial gastritis (SG). All the diagnoses were based on the gastric epithelial neoplasia classification system from the Japanese Research Society for Gastric Cancer (JRSGC) [
29]. Detailed information about the definition of risk predictors and outcome assessment in the three studies is detailed in the Additional file
1: Appendix 2.0 and 3.0 [
23,
29‐
34]. Deidentified datasets of the Changzhou cohort and Yangzhou screening program analyzed during the current study are available in Additional file
2.
Statistical analyses
All participants were assessed for their GC risk since enrollment until the time of GC diagnosis, death, loss to follow-up, or the end of follow-up, whichever occurred first. Cox proportional hazards regression model was used to assess the association between each variable and incident GC risk and to estimate hazard ratios (HRs) with 95% confidence intervals (CIs) in the CKB cohort. Univariate analyses were performed to select potentially effective predictors firstly, and those with P < 0.20 were kept for building a multivariate Cox regression model, followed by backward stepwise regression analyses. Based on the final Cox regression model in the CKB cohort, a regression coefficient-based scoring method was adopted to calculate the GCRS. One point was assigned to the predictor with the minimum regression coefficient in the model, and other predictors were assigned with the ratios of corresponding coefficients against the minimum coefficient. The points of predictors were kept to one decimal place and then summed up to generate a GCRS for each participant.
The predicted risk was estimated by using the “predict” function with the type of “expected” from the “survival” package with GCRS as a predictor. The observed GC risk was calculated by the Kaplan-Meier method. Model calibration was assessed by plotting the mean of the predicted probability against the mean of the observed probability of GC at 10 years by the tenth of predicted risk.
R2 was calculated from the linear regression and used to assess the quantitative calibration [
35]. Model discrimination was assessed with Harrell’s concordance C (Harrell’s C-index). Receiver operating characteristic (ROC) curves were plotted with all possible GCRSs as cutoff points for the prediction of developing GC within 10 years of follow-up [
36]. We also evaluated the model performance separately for 10 study regions. Internal validation of model discrimination was assessed by using the tenfold cross-validation [
37,
38].
The absolute risk of GC was projected at three time points (3, 5, and 10 years) by the deciles of the GCRS. Participants were further categorized into low (bottom 20%), intermediate (20–80%), and high (top 20%) risk groups based on the distribution of the GCRS in the CKB cohort, and the corresponding 3-, 5-, and 10-year cumulative incidences were estimated. In the Changzhou cohort and Yangzhou screening program, we calculated the GCRS for each participant blinded to the outcome with the same method used in the CKB cohort. We also estimated the performance of the GCRS corresponding to the deciles as cutoffs in the Yangzhou screening program. Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and numbers needed to be screened (NNS, one divided by the PPV) were evaluated.
We conducted sensitivity analyses to assess the robustness of our results. Firstly, a simplified model was created based on a subset of strong predictors (the assigned points ≥ 4.0). Secondly, a healthy lifestyle index was generated by integrating five modifiable lifestyle factors (generally weak predictors being assigned points < 4.0), i.e., BMI, smoking, alcohol use, consumption of fresh vegetables and fruits, and salty food intake. Thirdly, we excluded participants who had GC diagnosis within the first year after recruitment to avoid detection bias. Fourthly, in order to avoid the potential interaction between different cancers, we excluded all cancer participants at baseline. Finally, a competing risk model by considering death as a competing event was conducted, since those participants might develop GC thereafter. Additionally, the above sensitivity analyses were conducted by reconstructing GCRS accordingly, and the discrimination and calibration abilities were investigated as well. All P-values were two-sided, and P < 0.05 was considered statistically significant unless specified otherwise. All statistical analyses were performed by using R version 3.6.3 (R Core Team, Vienna, Austria).
Discussion
In the present study, by using the largest nationwide prospective cohort in China, we developed a GC risk assessment tool of GCRS based on eleven variables that could be easily determined without any physical examinations. We validated the GCRS with good calibration and discrimination in the independent Changzhou cohort, demonstrating the great potential of GCRS for GC risk prediction and stratification. When applying the GCRS to a real-world endoscopy screening program, we detected approximately 80% of all the identified GC cases in about one-quarter of individuals with high GCRS. To the best of our knowledge, the present study is the first to provide a questionnaire-based GC risk assessment tool based on a large-scale cohort study that can be used for risk stratification in an endoscopic screening setting of the Chinese population.
To date, several risk-prediction models have been developed for GC, but few are translated into practice. For example, the Japan Public Health Center-based Prospective Study (JPHC Study) developed a prediction model including age, sex, smoking status, consumption of high-salt food, family history of gastric cancer,
H. pylori antibody, and serum pepsinogen, which resulted in a C-statistic of 0.768 for discrimination [
15]. In China, there are two risk prediction models for GC, predominantly based on serum PG I, PG II, gastrin-17 (G-17), and anti-
H. pylori antibody, which were developed in a population-based follow-up study [
11] and a hospital-based cross-sectional study [
9], respectively. These two models also showed good discrimination (C-statistic of 0.803 and area under the curve of 0.76, respectively). However, these risk prediction models, mainly based on one study population, have a potential risk of over-fitting and should be subjected to rigorous external validations in the future. Of note, these abovementioned models based on serology tests not only add additional costs but also increase the degree of screening complexity, which may decrease the overall participation and efficiency. In the present study, we developed the questionnaire-based GCRS by using the largest Chinese cohort and independently validated the tool in an external cohort with good discrimination (Harrell’s C-index: 0.736). The large sample size and rigorous design ensured the quality and applicability of our GC risk assessment tool, which may be useful for tailored screening practices in the general population.
Although screening by endoscopy could reduce the mortality of GC [
4,
7], the availability of endoscopic instruments and expertise for mass screening remains questionable and impractical. Even though some countries, such as Japan and Korea, have implemented a national GC screening program [
5,
6], most have adopted screening approaches for high-risk populations. The initial prescreening tools, generally based on risk prediction models, provide a tailored screening for the general population. In the present study, we evaluated the initial GCRS in the Yangzhou screening program and found that 81.6% of the identified GC cases were correctly allocated to undergo endoscopy in the at-high-risk individuals who accounted for only about one-quarter of all screenings; moreover, none of the GC cases was detected in participants at low risk, suggesting that the low-risk populations could also be identified reliably. Thus, the developed GCRS may be employed to a tailored endoscopy screening, which could substantially decrease endoscopy workload and cost, compared with endoscopy for all. However, the incidence rate of GC changed remarkably in different areas across China [
39], while the developed GCRS may represent the average level of the Chinese populations. Therefore, further external validation with re-calibrated estimates based on local incidence would be necessary for clinical use [
40], especially for setting actionable cut points in different areas of China.
Nevertheless, additional studies are warranted to address several concerns regarding the applicability of the GCRS. First, although the GCRS was developed, replicated, and evaluated in twelve geographic areas of China, the tool needs to be evaluated or optimized in other areas or populations. For example, efforts are required to evaluate the generalizability of the GCRS to hospital-based screening. Second, the GCRS may help inform decision-making for GC screening, but several questions remain to be addressed, including optimal cutoff points of risk stratification, starting and stopping ages, and intensity of screening. Third, the prevalence rate of GC in the CKB was lower than expected, which was probably due to volunteer bias that individuals with GC were not inclined to attend the survey in the CKB at baseline. Nevertheless, the prevalent GC cases might be undetected through questionnaires, which could also contribute to the low prevalence rate and lead to inaccurate estimates of predictors. Fourth, although previous cancer diagnosis was used in this study as a predictor for GC, which was in line with that in lung cancer [
41], additional studies are warranted to explore the potential benefit of endoscopic screening in prevalent cancer patients. Fifth, concern still exists regarding whether or how much the GCRS-directed screening can improve the cost-effectiveness of endoscopic screening, compared with the current “one-size-fits-all” approach, which needs to be further assessed in future studies. At last,
H. pylori infection is the most important risk factor of GC, and we have also reported that a polygenic risk score with 112 genetic variants is effective for risk stratification of GC [
23]. However, the information was not available in the discovery and validation cohorts in this study. Therefore, additional studies are needed to develop a comprehensive score with the GCRS,
H. pylori infection status, polygenic risk score, and other serum biomarkers (e.g., PG I, PG II, and gastrin-17) to further optimize the risk prediction of GC. Moreover, the utility of these scores needs to be evaluated in endoscopy screening practices.
Several limitations of the present study should be noted. First, the lifestyle and personal history information was self-reported at baseline, which may cause some misclassifications and have biased the risk estimates of variables included in the GCRS. Second, we only evaluated the overall GC risk, but the risk estimate might differ depending on tumor location, stage, and subtype that were not obtained with details in the follow-up of cohorts. Third,
H. pylori infection, the most important risk factor of gastric cancer [
42], and family history of upper gastrointestinal cancers [
43] were unavailable in the development and validation cohorts and thus not included in the GCRS.
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.