Prognostic Factors in Survivors of Gastric Cancer: A Comparative Study of Cox Proportional Hazards Model and Machine Learning Approaches Using Korean National Cohort Data

Prognostic Factors in Survivors of Gastric Cancer: A Comparative Study of Cox Proportional Hazards Model and Machine Learning Approaches Using Korean National Cohort Data | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Prognostic Factors in Survivors of Gastric Cancer: A Comparative Study of Cox Proportional Hazards Model and Machine Learning Approaches Using Korean National Cohort Data Joonki Lee, Aesun Shin This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7022055/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract Background As the prognosis of gastric cancer has improved, the exploration of prognostic factors has become increasingly important. This study aimed to identify prognostic factors of gastric cancer using machine learning and statistical methods and to compare the effectiveness of different methodologies in identifying prognostic factors. Methods We conducted a retrospective cohort study of cancer research data from survivors of gastric cancer in Korea. Patients were followed up from the date of curative treatment of gastric cancer to the date of recurrence, cancer-specific death, or censoring. The Cox proportional hazards, random survival forest, XGBoost, and DeepSurv models were used to calculate the risk of recurrence and cancer-specific death. All the models were trained on 80% of the training set, and the concordance index was used for comparison with 20% of the test set. The SHAP value was used for variable interpretation in the machine learning models. Results A total of 11,029 gastric cancer survivors with a median follow-up time of 6.19 years were included. Remnant stomach after gastric cancer treatment, T stage and N stage were the most important features for recurrence and mortality according to both the Cox model and the machine learning model. All the models had a concordance index greater than 0.7 without large differences. Conclusions The machine learning model is not inferior to conventional statistical analysis models and offers greater flexibility, especially when statistical assumptions are violated. The key prognostic factors identified through this approach include residual stomach after treatment and cancer stage. Gastric cancer Prognosis Cohort Machine learning SHAP Figures Figure 1 Figure 2 Figure 3 Introduction Gastric cancer remains a significant health concern in Korea because of its high prevalence and incidence( 1 ). Although the incidence of gastric cancer has been steadily decreasing, the number of survivors continues to grow, underscoring the importance of surveillance and identification of prognostic factors. Tumor stage is a well-established prognostic indicator( 2 ). Treatment strategies are typically guided by tumor stage, with substantial supporting evidence( 3 , 4 ). Helicobacter pylori ( H. pylori ) eradication has demonstrated a significant role in reducing recurrence( 5 ), and its implementation following endoscopic resection is included in the Korean guidelines( 3 ). However, other potential prognostic indicators, such as lifestyle factors, have been less studied in relation to gastric cancer. Randomized controlled trials investigating these lifestyle factors are challenging, making observational cohort studies and survival analyses essential. However, conventional statistical models such as the Cox proportional hazards model rely on assumptions such as proportional hazards, which may not always be valid. Recently, machine learning (ML) methods have gained attention for their ability to handle complex data structures and facilitate personalized risk prediction( 6 ). Survival analysis is traditionally used as a tool for analyzing right-truncated data, and the Cox proportional hazards model is widely used. There are also ML models that can be used for survival analysis( 7 ). However, ML models are well known as black-box models, which makes interpretation difficult. Because of this limitation, ML models are rarely applicable in medical analyses in which the associated factors are identified. Despite concerns regarding the interpretability of ML models, methods such as SHAP (SHapley Additive exPlanations) offer promising solutions( 8 ). This study aimed to identify prognostic factors in survivors of gastric cancer by comparing conventional Cox models with ML approaches, including random survival forest( 9 ), XGBoost( 10 ), and DeepSurv( 11 ) models, in a large national cohort. The goal was to assess the clinical applicability and predictive accuracy of ML methods and explore understudied factors affecting prognosis. Materials and Methods Data Source This study utilized the Korean Cancer Research Data (K-CURE), which was launched in 2023( 12 ). The K-CURE integrates national cancer registry data, health insurance claim data, national health screening records, and mortality data. We analyzed data from a random sample of 23,717 patients diagnosed with gastric cancer. Study Design We conducted a retrospective cohort study involving patients who underwent curative treatment for gastric cancer between 2012 and 2019, following Korean guidelines( 3 ). Curative treatments included endoscopic submucosal dissection (ESD) and gastrectomy with or without adjuvant chemotherapy. Neoadjuvant chemotherapy, which is not a usual treatment in Korea, and nonguideline procedures, such as wedge resection or endoscopic mucosal resection, were excluded. Only patients who received capecitabine + oxaliplatin or tegafur + gimeracil + oteracil within three months of surgery were considered to have received guideline-compliant adjuvant therapy. Patients who received other types of chemotherapy, such as palliative chemotherapy or chemotherapy for cancers other than gastric cancer, were excluded from our analysis. Patients with distant metastasis at diagnosis, other malignancies, or incomplete data were also excluded. Study participants were followed from the date of ESD or gastrectomy, which was initially defined as the index date. If additional ESD or gastrectomy was performed within 90 days of this date, it was considered part of synchronous cancer treatment, incomplete resection, or a case not meeting ESD criteria. In such cases, the date of the last treatment within the initial course was designated as the revised index date. The study endpoints were gastric cancer recurrence, gastric cancer-specific death, the last recorded clinical visit, or death. Gastric cancer recurrence was defined as any treatment (ESD, gastrectomy, or chemotherapy) more than 90 days after index date. The variables included sex, age at diagnosis, AJCC 7th edition T and N stage, tumor size, carcinoembryonic antigen (CEA) level, and histologic differentiation. Tumor size was categorized as ≤ 1 cm or > 1 cm based on the data distribution. Patients who undergo ESD have a whole stomach, those who undergo distal gastrectomy have a proximal stomach, those who undergo proximal gastrectomy have a distal stomach, and those who undergo total gastrectomy have no stomach after treatment. Anatomical differences after treatment are potential prognostic factors; therefore, we included remnant stomach as a variable. A large proportion of the data concerning other factors, such as lifestyle, were missing. Therefore, we additionally performed subcohort analysis with patients who had complete data for smoking, alcohol use, physical activity, BMI, socioeconomic status, family history, body measurement value and tumor markers. For the subcohort analysis, prescriptions for H. pylori eradication were included. H. pylori eradication was defined as prescription of a proton pump inhibitor (PPI) + amoxicillin + clarithromycin for more than 7 days or a PPI + bismuth + metronidazole + tetracycline for more than 7 days before or after treatment ( 13 ). Metabolic syndrome was defined on the basis of waist circumference, triglyceride level, high-density lipoprotein (HDL) cholesterol level, blood pressure, and fasting glucose level in health screening records ( 14 ). Changes in modifiable factors after treatment were also analyzed by subcohort analysis. Modifiable behavior changes were categorized as ‘no change’, ‘good change’, or ‘bad change’ based on health screening data before and 1–2 years after treatment. The Charlson Comorbidity Index (CCI) was calculated using updated Quan’s codes and included in subcohort analysis( 15 ). The first and second diagnostic ICD-10 codes for admission for 5 years before the index date were used to calculate the CCI ( 16 ). Analysis We performed a Cox proportional hazards model for gastric cancer recurrence and mortality. Proportional hazard assumptions were checked via Kaplan‒Meier curves and Schoenfeld residual plots. In the analysis for gastric cancer mortality, age and BMI changes were categorized to address assumption violations. Machine learning analyses included random survival forest, XGBoost, and DeepSurv models. In the XGBoost and DeepSurv models, the Cox partial likelihood method was used to calculate the risk score and loss function. An 80:20 train‒test split was used for all the models. The concordance index (C-index) was calculated on the test set to assess model performance and comparison ( 7 ). SHAP values were computed for the ML models to assess variable importance. All analyses were conducted using SAS 9.4 and Python with the Scikit survival, XGBoost, and PyTorch packages. Results After all exclusions were applied, 11,029 patients were included in the main analysis, and 4,945 were included in the subcohort analysis (Fig. 1 ). The median follow-up was 6.19 years. Baseline characteristics are shown in Table 1 and Supplementary Table 1. Table 1 Baseline characteristics of the study population categorized by gastric cancer recurrence and mortality Recurrence Gastric cancer mortality Overall Disease free Recurrence p-value Survived Death p-value n 11029 10080 949 10471 558 Age (years), mean (SD) 62.7 (10.7) 62.6 (10.7) 63.0 (10.4) 0.373 62.6 (10.6) 64.9 (12.0) < 0.001 Age group, n (%) < 60yrs 4204 (38.12) 3878 (38.47) 326 (34.35) 0.013 4033 (38.52) 171 (30.65) < 0.001 ≥ 60yrs 6825 (61.88) 6202 (61.53) 623 (65.65) 6438 (61.48) 387 (69.35) Sex, n (%) Female 3321 (30.1) 3097 (30.7) 224 (23.6) < 0.001 3176 (30.3) 145 (26.0) 0.033 Male 7708 (69.9) 6983 (69.3) 725 (76.4) 7295 (69.7) 413 (74.0) T stage, n (%) T1 8942 (81.1) 8337 (82.7) 605 (63.8) < 0.001 8806 (84.1) 136 (24.4) < 0.001 T2 711 (6.4) 671 (6.7) 40 (4.2) 664 (6.3) 47 (8.4) T3 805 (7.3) 687 (6.8) 118 (12.4) 657 (6.3) 148 (26.5) T4 571 (5.2) 385 (3.8) 186 (19.6) 344 (3.3) 227 (40.7) N stage, n (%) N0 9744 (88.3) 9089 (90.2) 655 (69.0) < 0.001 9541 (91.1) 203 (36.4) < 0.001 N1 369 (3.3) 318 (3.2) 51 (5.4) 313 (3.0) 56 (10.0) N2 422 (3.8) 343 (3.4) 79 (8.3) 330 (3.2) 92 (16.5) N3 494 (4.5) 330 (3.3) 164 (17.3) 287 (2.7) 207 (37.1) Differentiation, n (%) Well-differentiated 3545 (32.1) 3201 (31.8) 344 (36.2) 0.008 3482 (33.3) 63 (11.3) < 0.001 Moderately differentiated 4138 (37.5) 3819 (37.9) 319 (33.6) 3953 (37.8) 185 (33.2) Poorly differentiated 3346 (30.3) 3060 (30.4) 286 (30.1) 3036 (29.0) 310 (55.6) Tumor size, n (%) ≤ 1 cm 10866 (98.5) 9959 (98.8) 907 (95.6) < 0.001 10367 (99.0) 499 (89.4) 1 cm 163 (1.5) 121 (1.2) 42 (4.4) 104 (1.0) 59 (10.6) CEA, n (%) Normal 7180 (65.1) 6659 (66.1) 521 (54.9) < 0.001 6848 (65.4) 332 (59.5) < 0.001 Elevated 764 (6.9) 677 (6.7) 87 (9.2) 675 (6.4) 89 (15.9) Missing 3085 (28.0) 2744 (27.2) 341 (35.9) 2948 (28.2) 137 (24.6) Remnant stomach, n (%) Whole 4139 (37.5) 3622 (35.9) 517 (54.5) < 0.001 4103 (39.2) 36 (6.5) < 0.001 Proximal 5416 (49.1) 5153 (51.1) 263 (27.7) 5124 (48.9) 292 (52.3) Distal 180 (1.6) 171 (1.7) 9 (0.9) 173 (1.7) 7 (1.3) None 1294 (11.7) 1134 (11.2) 160 (16.9) 1071 (10.2) 223 (40.0) The Cox model results (Table 2 ) revealed that remnant stomach was the strongest predictor of recurrence. Patients with a whole stomach after ESD had an approximately 5-fold higher recurrence rate than those with partial or no stomach after gastrectomy. Conversely, gastric cancer mortality was higher among patients who underwent gastrectomy, particularly those who underwent total gastrectomy. The T and N stages were also significant for both outcomes. Older age, male sex and elevated CEA levels were associated with increased mortality risk. However, tumor size and histologic differentiation were not significant. In the subcohort analysis (Supplementary Table 2), most factors were not significantly associated with prognosis. However, a decrease in BMI after treatment was associated with decreased mortality. Patients who initiated smoking after treatment had universally poor outcomes, precluding HR estimation. H. pylori eradication was significantly associated with lower mortality but not recurrence. However, Cox analysis stratified by treatment method revealed that H. pylori eradication significantly reduced both gastric cancer recurrence and mortality only in patients who underwent ESD (Supplementary Table 3). Patients who had a family history of gastric cancer had favorable outcomes in terms of both recurrence and mortality. The SHAP results of the ML models were similar (Figs. 2 , 3 ). In contrast to the other ML models, XGBoost identified age as the most important variable in recurrence. Model comparison using the C-index revealed no large differences; all the models had a C-index > 0.7 (Table 3 ). Table 2. Hazard ratio of gastric cancer recurrence and mortality Recurrence Mortality Univariate Multivariate Univariate Multivariate HR (95% CI) HR (95% CI) HR (95% CI) HR (95% CI) Sex Male (vs. Female) 1.43 (1.23-1.66)* 1.37 (1.18-1.59)* 1.24 (1.03-1.50)* 1.35 (1.11-1.63)* Age † cont./≥60yrs (vs. <60yrs) 1.01 (1.00-1.01)* 1.00 (1.00-1.01) 1.45 (1.21-1.73)* 1.80 (1.50-2.17)* T stage T2 (vs. T1) 0.85 (0.62-1.17) 2.08 (1.40-3.11)* 4.51 (3.24-6.28)* 2.15 (1.49-3.11)* T3 (vs. T1) 2.43 (1.99-2.96)* 4.16 (2.94-5.87)* 13.46 (10.66-16.99)* 4.53 (3.31-6.22)* T4 (vs. T1) 6.41 (5.44-7.56)* 7.73 (5.40-11.06)* 33.43 (27.01-41.37)* 7.48 (5.37-10.43)* N stage N1 (vs. N0) 2.28 (1.72-3.04)* 2.14 (1.48-3.10)* 7.92 (5.89-10.64)* 1.79 (1.27-2.53)* N2 (vs. N0) 3.18 (2.52-4.02)* 2.94 (2.11-4.09)* 11.74 (9.17-15.02)* 2.46 (1.82-3.33)* N3 (vs. N0) 6.69 (5.63-7.94)* 4.39 (3.21-5.99)* 26.29 (21.64-31.92)* 4.06 (3.07-5.37)* Remnant stomach Distal (vs. Whole) 0.39 (0.20-0.75)* 0.17 (0.08-0.33)* 4.53 (2.02-10.17)* 2.07 (0.90-4.74) None (vs. Whole) 1.04 (0.87-1.24) 0.17 (0.13-0.23)* 21.72 (15.27-30.88)* 3.87 (2.53-5.93)* Proximal (vs. Whole) 0.37 (0.32-0.43)* 0.13 (0.10-0.17)* 6.28 (4.45-8.88)* 2.27 (1.51-3.41)* Size >1 cm (vs. ≤1 cm) 3.94 (2.89-5.37)* 1.19 (0.85-1.66) 9.90 (7.56-12.97)* 1.33 (0.99-1.77) Differentiation Moderately differentiated (vs. WD) 0.81 (0.69-0.94)* 0.96 (0.81-1.13) 2.57 (1.93-3.42)* 0.92 (0.67-1.26) Poorly differentiated (vs. WD) 0.90 (0.77-1.05) 1.02 (0.82-1.27) 5.44 (4.15-7.13)* 1.02 (0.75-1.41) CEA Elevated (vs. Normal) 1.70 (1.35-2.13)* 1.24 (0.98-1.56) 2.69 (2.13-3.40)* 1.42 (1.12-1.81)* Missing (vs. Normal) 1.56 (1.36-1.79)* 1.11 (0.96-1.28) 0.70 (0.96-0.79)* 1.11 (0.90-1.36) *p value < 0.05 †Age was treated as a categorical variable in mortality analysis Table 3 Comparison between machine learning models and statistical analysis Model Recurrence Mortality c-index Important variables c-index Important variables Cox proportional hazard model 0.7560 Remnant stomach, T stage, N stage, Sex 0.8979 T stage, N stage, Remnant stomach, CEA level, Sex, Age Random survival forest model 0.7648 Remnant stomach, T stage, N stage 0.9196 N stage, T stage, Age, Remnant stomach XGBoost model 0.7300 Age, T stage, Remnant stomach, Grade, N stage 0.8916 Age, N stage, T stage, Remnant stomach, Grade DeepSurv model 0.7745 Remnant stomach, T stage, N stage 0.8907 Age, Remnant stomach, T stage, N stage *The selected variables were age, sex, T stage, N stage, tumor size, Remnant stomach after treatment, and the CEA level and c-index, which were calculated with a 20% randomly sampled test dataset. **Variables were selected by significance and SHAP value importance and ordered by importance Discussion This study demonstrated that ML models can produce comparable results to conventional Cox models, even in the presence of assumption violations. Age and BMI changes, which violated proportional hazard assumptions, were effectively handled in the ML models without requiring categorization, highlighting the flexibility of ML models. SHAP values provided interpretability to ML models, aligning closely with Cox model findings. This suggests that SHAP can effectively address the "black box" nature of ML, making it more suitable for clinical applications. However, the SHAP value does not provide the risk ratio or significance presented in most medical studies. The SHAP value indicates the importance of the feature in the model ( 8 ). The important features were discovered with the SHAP value, but the interpretation was less clear than that of the statistical model. Although the SHAP’s ability to interpret ML models is limited, its application is more flexible than that of a statistical model. SHAP can be applied for both linear and nonlinear models, even for the Cox model and deep learning model. As shown in our study, SHAP can be used as a good alternative option when statistical analysis cannot be performed. We used random survival forest, XGBoost, and DeepSurv models. Although there was no significant difference between the models, they yielded different SHAP results. This might be due to the different modeling processes among the models. Thus, although the models showed similar c-indices SHAP value should be interpreted with caution. Several medical studies have compared conventional statistical models and machine learning models ( 17 – 19 ). Compared with conventional statistical models, ML has strong advantages in terms of application to high-dimensional data, accuracy, and flexible modeling. Previous studies suggested the possibility of applying ML in medical studies, which is supported by our findings. Furthermore, the usefulness of the ML model was emphasized by the SHAP results in our study. Our findings confirmed known prognostic factors, such as remnant stomach status, tumor stage, and sex( 2 , 20 – 22 ). In 2018, a randomized controlled study revealed that H. pylori eradication in post-ESD patients decreased the recurrence of gastric cancer( 5 ), which led to H. pylori eradication as a standard treatment after ESD in Korea. In our study, H. pylori eradication was effective only in patients who underwent ESD. Patients whose stomach anatomy was altered after gastrectomy did not show an association between H. pylori eradication and either recurrence or mortality. Because our data did not include H. pylori status, we could only consider prescriptions for H. pylori eradication. Although the assessment of H. pylori was limited, the prognosis associated with its eradication was similar to that reported in previous studies. Inconsistent results concerning the effects of H. pylori eradication have been reported in patients undergoing gastrectomy( 23 , 24 ). The role of H. pylori eradication was evident in patients who underwent ESD but not in those who underwent gastrectomy; therefore, further studies with better designs are needed for patients undergoing gastrectomy. Tumor size is also known as an independent prognostic factor( 25 ). However, our study revealed a greater risk of recurrence and mortality in patients with tumors larger than 1 cm than in those with tumors smaller than 1 cm, but the difference was not statistically significant, which was likely related to the data distribution. Differentiation and family history are factors whose associations have not been clearly identified, as previous studies have shown inconsistent results. However, recent studies indicated that there was no significant association with differentiation, which was consistent with our study ( 26 ). A Korean study of 1,273 patients with gastric cancer reported favorable outcomes of patients with a family history of gastric cancer, which was also shown in our study ( 27 ). The CCI is a factor that is designed for adjusting mortality. However, a previous study indicated that the CCI may not be associated with gastric cancer mortality( 28 ), which was consistent with our study. Modifiable risk factors before treatment have limited prognostic value, but posttreatment changes, particularly smoking, may influence outcomes. Other factors also showed favorable outcomes with favorable behavior changes, but these findings were not statistically significant. Further studies with more comprehensive behavioral data are needed. Our study had limitations, which included the use of administrative data lacking clinical test results. This leads to potential misclassification of recurrence or H. pylori status and the absence of important variables. Pathology results, such as depth of invasion, lymphatic invasion, and margin involvement, which are well-known risk factors, were absent ( 29 ). Atrophic gastritis and intestinal metaplasia are also known risk factors for gastric cancer and its recurrence( 30 ), which were not included in our data. In addition, patient-related factors such as morbidity were not fully accessible. Patients may not have received treatment when they experienced recurrence because of their health condition. The detailed information of patients was not recorded in the claim data, which led to misclassification of the endpoint. However, there were no missing data for mortality in our study; therefore, analysis with gastric cancer mortality as the endpoint revealed no exceptional misclassification. Additionally, findings may not be generalizable outside of Korea because of the different natures of gastric cancer ( 31 , 32 ). To generalize these findings worldwide, differences in ethnicity should be considered. In conclusion, machine learning offers a viable alternative to conventional survival analysis, particularly when statistical assumptions are violated. SHAP enhances model interpretability, supporting its integration into precision medicine. Key prognostic factors such as the remnant stomach status and tumor stage should guide the level of follow-up. Posttreatment behavioral modifications may offer additional survival benefits, which warrants further study. Declarations Author Contributions: JL designed and conducted the study; acquired, analyzed, and interpreted the data; drafted the manuscript. AS provided administrative, technical, and financial support; reviewed the manuscript, and approved the final draft submitted. Acknowledgements: We thank American Journal Experts (AJE) for English language editing of this manuscript (certificate verification code: 19BC-7A1E-FC0B-98F7-1F8E). This study used the K-CURE cancer public library database, which was established by the National Cancer Data Center as part of the K-CURE project organized by the Ministry of Health and Welfare (Research number: :KC20240712002) Funding : This study was supported by Cancer Research Institute, Seoul National University. Ethics approval and consent to participate This study was approved by the Institutional Review Board of Seoul National University Hospital (IRB No. E-2401-057-1501). The need for informed consent was waived by the IRB, as only anonymized data were used. This study was conducted in accordance with the principles of the Declaration of Helsinki. Consent for publication Not applicable. Competing interests The authors declare no other conflict of interest. Data availability The data used in this study are available through the K-CURE project (https://k-cure.mohw.go.kr/) upon approval by the Institutional Review Board (IRB) and Data Review Board (DRB) of participating institutions. Data can only be accessed and analyzed in a secure Central Data Center; only analysis results may be exported. References Park EH, Jung KW, Park NJ, Kang MJ, Yun EH, Kim HJ, et al. Cancer Statistics in Korea: Incidence, Mortality, Survival, and Prevalence in 2022. Cancer Res Treat. 2025;57(2):312-30. In H, Solsky I, Palis B, Langdon-Embry M, Ajani J, Sano T. Validation of the 8th Edition of the AJCC TNM Staging System for Gastric Cancer using the National Cancer Database. Annals of Surgical Oncology. 2017;24(12):3683-91. Kim TH, Kim IH, Kang SJ, Choi M, Kim BH, Eom BW, et al. Korean Practice Guidelines for Gastric Cancer 2022: An Evidence-based, Multidisciplinary Approach. J Gastric Cancer. 2023;23(1):3-106. Eom SS, Ryu KW, Han HS, Kong S-H. A Comprehensive and Comparative Review of Global Gastric Cancer Treatment Guidelines: 2024 Update. J Gastric Cancer. 2025;25(1):153-76. Choi IJ, Kook MC, Kim YI, Cho SJ, Lee JY, Kim CG, et al. Helicobacter pylori Therapy for the Prevention of Metachronous Gastric Cancer. N Engl J Med. 2018;378(12):1085-95. Rajkomar A, Dean J, Kohane I. Machine Learning in Medicine. New England Journal of Medicine. 2019;380(14):1347-58. Wang P, Li Y, Reddy CK. Machine Learning for Survival Analysis: A Survey. ACM Comput Surv. 2019;51(6):Article 110. Lundberg SM, Lee S-I. A unified approach to interpreting model predictions. Advances in neural information processing systems. 2017;30. Ishwaran H, Kogalur UB, Blackstone EH, Lauer MS. Random survival forests. 2008. Chen T, Guestrin C. XGBoost: A Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; San Francisco, California, USA: Association for Computing Machinery; 2016. p. 785–94. Katzman JL, Shaham U, Cloninger A, Bates J, Jiang T, Kluger Y. DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Medical Research Methodology. 2018;18(1):24. Choi D-W, Guk MY, Kim HR, Ryu KS, Kong H-J, Cha HS, et al. Data Resource Profile: The Cancer Public Library Database in South Korea. Cancer Res Treat. 2024;56(4):1014-26. Jung H-K, Kang SJ, Lee YC, Yang H-J, Park S-Y, Shin CM, et al. Evidence-based Guidelines for the Treatment of Helicobacter pylori Infection in Korea: 2020 Revised Edition. Korean J Helicobacter Up Gastrointest Res. 2020;20(4):261-87. Huang PL. A comprehensive definition for metabolic syndrome. Disease Models & Mechanisms. 2009;2(5-6):231-7. Quan H, Li B, Couris CM, Fushimi K, Graham P, Hider P, et al. Updating and validating the Charlson comorbidity index and score for risk adjustment in hospital discharge abstracts using data from 6 countries. American journal of epidemiology. 2011;173(6):676-82. Kim K-H. Comparative study on three algorithms of the ICD-10 Charlson comorbidity index with myocardial infarction patients. Journal of Preventive Medicine and Public Health. 2010;43(1):42-9. Rajula HSR, Verlato G, Manchia M, Antonucci N, Fanos V. Comparison of Conventional Statistical Methods with Machine Learning in Medicine: Diagnosis, Drug Development, and Treatment. Medicina. 2020;56(9):455. Spooner A, Chen E, Sowmya A, Sachdev P, Kochan NA, Trollor J, et al. A comparison of machine learning methods for survival analysis of high-dimensional clinical data for dementia prediction. Scientific Reports. 2020;10(1):20410. Premsagar P, Aldous C, Esterhuizen TM, Gomes BJ, Gaskell JW, Tabb DL. Comparing conventional statistical models and machine learning in a small cohort of South African cardiac patients. Informatics in Medicine Unlocked. 2022;34:101103. Mantziari S, St Amour P, Abboretti F, Teixeira-Farinha H, Gaspar Figueiredo S, Gronnier C, et al. A Comprehensive Review of Prognostic Factors in Patients with Gastric Adenocarcinoma. Cancers. 2023;15(5):1628. Sun B, Zhang H, Wang J, Cai H, Xuan Y, Xu D. Tumor Location Causes Different Recurrence Patterns in Remnant Gastric Cancer. J Gastric Cancer. 2022;22(4):369-80. Liu Q, Ding L, Qiu X, Meng F. Updated evaluation of endoscopic submucosal dissection versus surgery for early gastric cancer: A systematic review and meta-analysis. International Journal of Surgery. 2020;73:28-41. Kim Y-I, Cho S-J, Lee JY, Kim CG, Kook M-C, Ryu KW, et al. Effect of Helicobacter pylori Eradication on Long-Term Survival after Distal Gastrectomy for Gastric Cancer. Cancer Res Treat. 2016;48(3):1020-9. Zhao Z, Zhang R, Chen G, Nie M, Zhang F, Chen X, et al. Anti–Helicobacter pylori Treatment in Patients With Gastric Cancer After Radical Gastrectomy. JAMA Network Open. 2024;7(3):e243812-e. Im WJ, Kim MG, Ha TK, Kwon SJ. Tumor size as a prognostic factor in gastric cancer patient. J Gastric Cancer. 2012;12(3):164-72. Feng F, Liu J, Wang F, Zheng G, Wang Q, Liu S, et al. Prognostic value of differentiation status in gastric cancer. BMC Cancer. 2018;18(1):865. Han MA, Oh MG, Choi IJ, Park SR, Ryu KW, Nam B-H, et al. Association of Family History With Cancer Recurrence and Survival in Patients With Gastric Cancer. Journal of Clinical Oncology. 2012;30(7):701-8. Kyung M-H, Yoon S-J, Ahn H-S, Hwang S-m, Seo H-J, Kim K-H, et al. Prognostic impact of Charlson comorbidity index obtained from medical records and claims data on 1-year mortality and length of stay in gastric cancer patients. Journal of Preventive Medicine and Public Health. 2009;42(2):117-22. Hatta W, Gotoda T, Oyama T, Kawata N, Takahashi A, Yoshifuku Y, et al. A Scoring System to Stratify Curability after Endoscopic Submucosal Dissection for Early Gastric Cancer: “eCura system”. Official journal of the American College of Gastroenterology | ACG. 2017;112(6):874-81. Factors influencing occurrence of metachronous gastric cancer after endoscopic resection: a systematic review and meta-analysis FAU - Choe, Younghee FAU - Park, Jae Myung FAU - Kim, Joon Sung FAU - Cho, Yu Kyung FAU - Kim, Byung-Wook FAU - Choi, Myung-Gyu. Korean J Intern Med. 2023;38(6):831-43. Kendrick P, Kelly YO, Baumann MM, Compton K, Blacker BF, Daoud F, et al. The burden of stomach cancer mortality by county, race, and ethnicity in the USA, 2000–2019: a systematic analysis of health disparities. The Lancet Regional Health – Americas. 2023;24. Kim J, Sun CL, Mailey B, Prendergast C, Artinyan A, Bhatia S, et al. Race and ethnicity correlate with survival in patients with gastric adenocarcinoma. Annals of Oncology. 2010;21(1):152-60. Additional Declarations No competing interests reported. Supplementary Files SupplementaryTables.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 04 Sep, 2025 Reviews received at journal 28 Aug, 2025 Reviews received at journal 20 Aug, 2025 Reviewers agreed at journal 11 Aug, 2025 Reviewers agreed at journal 06 Aug, 2025 Reviewers agreed at journal 05 Aug, 2025 Reviewers invited by journal 05 Aug, 2025 Editor assigned by journal 02 Aug, 2025 Submission checks completed at journal 29 Jul, 2025 First submitted to journal 29 Jul, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7022055","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":496345996,"identity":"7b445266-d174-4c34-8201-1e6dd60b84e2","order_by":0,"name":"Joonki Lee","email":"","orcid":"","institution":"Seoul National University College of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Joonki","middleName":"","lastName":"Lee","suffix":""},{"id":496345997,"identity":"8cbbbbe6-c3d1-4f88-928f-48d393c4a3cc","order_by":1,"name":"Aesun Shin","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1ElEQVRIiWNgGAWjYDACCcYGCIMdSjPwEK2F5wDRWuCMBCK1yM9ubnvwo6Iucbvk82ePeRjs5Bl4zj7Aq8XgzsF2w54zhxN3zs4xN+ZhSDZs4G03wK9FIrFNgrftQOKG2zls0jwMzAkM/GwEHDYjsU3y77+6xA03jz8DaqknrIXhRmKbNG8Dc+KGGwxmQC2HExh42/DrMABpkTl22HjDmRwzyTkGxw3beI4Rclj6M8k3NXWyG44ffybxpqJanp8njYDD0CxlYCDkk1EwCkbBKBgFRAAA40w/Aj0x+KIAAAAASUVORK5CYII=","orcid":"","institution":"Seoul National University College of Medicine","correspondingAuthor":true,"prefix":"","firstName":"Aesun","middleName":"","lastName":"Shin","suffix":""}],"badges":[],"createdAt":"2025-07-01 15:53:15","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7022055/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7022055/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":88646083,"identity":"70ba1a0e-0496-436e-a77a-a4929524090c","added_by":"auto","created_at":"2025-08-08 16:32:57","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":427624,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStudy population of the main analysis of patients with gastric cancer with curative treatment and subcohort analysis of these patients without missing data.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7022055/v1/01457523a4a5065fb3de4917.png"},{"id":88647399,"identity":"41516bd0-0fc6-4d1f-b61f-a0162ea0a3e3","added_by":"auto","created_at":"2025-08-08 16:40:57","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":300166,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSHAP results (summary plots) of gastric cancer recurrence and mortality from machine learning models\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7022055/v1/d6d6df666fb2b2167cb6415e.png"},{"id":88646084,"identity":"ba7708f4-ebd0-4ff1-831f-6460a58b71fa","added_by":"auto","created_at":"2025-08-08 16:32:57","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":315029,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSHAP results (summary plots) of gastric cancer recurrence and mortality from machine learning models\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-7022055/v1/9585d1c71b8663bac85e8137.png"},{"id":88649088,"identity":"bc7a0b8c-639b-40dd-8323-7319298caee2","added_by":"auto","created_at":"2025-08-08 16:56:59","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2028953,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7022055/v1/51c7f12b-02e9-4527-941f-e9551635bcff.pdf"},{"id":88646082,"identity":"ff9dd187-34be-4707-8adf-d675cb73f889","added_by":"auto","created_at":"2025-08-08 16:32:57","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":34141,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTables.docx","url":"https://assets-eu.researchsquare.com/files/rs-7022055/v1/8f2f962e59cce54bf72d08b2.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Prognostic Factors in Survivors of Gastric Cancer: A Comparative Study of Cox Proportional Hazards Model and Machine Learning Approaches Using Korean National Cohort Data","fulltext":[{"header":"Introduction","content":"\u003cp\u003eGastric cancer remains a significant health concern in Korea because of its high prevalence and incidence(\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). Although the incidence of gastric cancer has been steadily decreasing, the number of survivors continues to grow, underscoring the importance of surveillance and identification of prognostic factors. Tumor stage is a well-established prognostic indicator(\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). Treatment strategies are typically guided by tumor stage, with substantial supporting evidence(\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e). \u003cem\u003eHelicobacter pylori\u003c/em\u003e (\u003cem\u003eH. pylori\u003c/em\u003e) eradication has demonstrated a significant role in reducing recurrence(\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e), and its implementation following endoscopic resection is included in the Korean guidelines(\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). However, other potential prognostic indicators, such as lifestyle factors, have been less studied in relation to gastric cancer.\u003c/p\u003e\u003cp\u003eRandomized controlled trials investigating these lifestyle factors are challenging, making observational cohort studies and survival analyses essential. However, conventional statistical models such as the Cox proportional hazards model rely on assumptions such as proportional hazards, which may not always be valid. Recently, machine learning (ML) methods have gained attention for their ability to handle complex data structures and facilitate personalized risk prediction(\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e). Survival analysis is traditionally used as a tool for analyzing right-truncated data, and the Cox proportional hazards model is widely used. There are also ML models that can be used for survival analysis(\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e). However, ML models are well known as black-box models, which makes interpretation difficult. Because of this limitation, ML models are rarely applicable in medical analyses in which the associated factors are identified. Despite concerns regarding the interpretability of ML models, methods such as SHAP (SHapley Additive exPlanations) offer promising solutions(\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eThis study aimed to identify prognostic factors in survivors of gastric cancer by comparing conventional Cox models with ML approaches, including random survival forest(\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e), XGBoost(\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e), and DeepSurv(\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e) models, in a large national cohort. The goal was to assess the clinical applicability and predictive accuracy of ML methods and explore understudied factors affecting prognosis.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003eData Source\u003c/p\u003e\u003cp\u003eThis study utilized the Korean Cancer Research Data (K-CURE), which was launched in 2023(\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e). The K-CURE integrates national cancer registry data, health insurance claim data, national health screening records, and mortality data. We analyzed data from a random sample of 23,717 patients diagnosed with gastric cancer.\u003c/p\u003e\u003cp\u003eStudy Design\u003c/p\u003e\u003cp\u003eWe conducted a retrospective cohort study involving patients who underwent curative treatment for gastric cancer between 2012 and 2019, following Korean guidelines(\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). Curative treatments included endoscopic submucosal dissection (ESD) and gastrectomy with or without adjuvant chemotherapy. Neoadjuvant chemotherapy, which is not a usual treatment in Korea, and nonguideline procedures, such as wedge resection or endoscopic mucosal resection, were excluded. Only patients who received capecitabine\u0026thinsp;+\u0026thinsp;oxaliplatin or tegafur\u0026thinsp;+\u0026thinsp;gimeracil\u0026thinsp;+\u0026thinsp;oteracil within three months of surgery were considered to have received guideline-compliant adjuvant therapy. Patients who received other types of chemotherapy, such as palliative chemotherapy or chemotherapy for cancers other than gastric cancer, were excluded from our analysis. Patients with distant metastasis at diagnosis, other malignancies, or incomplete data were also excluded.\u003c/p\u003e\u003cp\u003eStudy participants were followed from the date of ESD or gastrectomy, which was initially defined as the index date. If additional ESD or gastrectomy was performed within 90 days of this date, it was considered part of synchronous cancer treatment, incomplete resection, or a case not meeting ESD criteria. In such cases, the date of the last treatment within the initial course was designated as the revised index date.\u003c/p\u003e\u003cp\u003eThe study endpoints were gastric cancer recurrence, gastric cancer-specific death, the last recorded clinical visit, or death. Gastric cancer recurrence was defined as any treatment (ESD, gastrectomy, or chemotherapy) more than 90 days after index date.\u003c/p\u003e\u003cp\u003eThe variables included sex, age at diagnosis, AJCC 7th edition T and N stage, tumor size, carcinoembryonic antigen (CEA) level, and histologic differentiation. Tumor size was categorized as \u0026le;\u0026thinsp;1 cm or \u0026gt;\u0026thinsp;1 cm based on the data distribution. Patients who undergo ESD have a whole stomach, those who undergo distal gastrectomy have a proximal stomach, those who undergo proximal gastrectomy have a distal stomach, and those who undergo total gastrectomy have no stomach after treatment. Anatomical differences after treatment are potential prognostic factors; therefore, we included remnant stomach as a variable.\u003c/p\u003e\u003cp\u003eA large proportion of the data concerning other factors, such as lifestyle, were missing. Therefore, we additionally performed subcohort analysis with patients who had complete data for smoking, alcohol use, physical activity, BMI, socioeconomic status, family history, body measurement value and tumor markers. For the subcohort analysis, prescriptions for \u003cem\u003eH. pylori\u003c/em\u003e eradication were included. \u003cem\u003eH. pylori\u003c/em\u003e eradication was defined as prescription of a proton pump inhibitor (PPI)\u0026thinsp;+\u0026thinsp;amoxicillin\u0026thinsp;+\u0026thinsp;clarithromycin for more than 7 days or a PPI\u0026thinsp;+\u0026thinsp;bismuth\u0026thinsp;+\u0026thinsp;metronidazole\u0026thinsp;+\u0026thinsp;tetracycline for more than 7 days before or after treatment (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). Metabolic syndrome was defined on the basis of waist circumference, triglyceride level, high-density lipoprotein (HDL) cholesterol level, blood pressure, and fasting glucose level in health screening records (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e). Changes in modifiable factors after treatment were also analyzed by subcohort analysis. Modifiable behavior changes were categorized as \u0026lsquo;no change\u0026rsquo;, \u0026lsquo;good change\u0026rsquo;, or \u0026lsquo;bad change\u0026rsquo; based on health screening data before and 1\u0026ndash;2 years after treatment. The Charlson Comorbidity Index (CCI) was calculated using updated Quan\u0026rsquo;s codes and included in subcohort analysis(\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e). The first and second diagnostic ICD-10 codes for admission for 5 years before the index date were used to calculate the CCI (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e).\u003c/p\u003e\u003cp\u003eAnalysis\u003c/p\u003e\u003cp\u003eWe performed a Cox proportional hazards model for gastric cancer recurrence and mortality. Proportional hazard assumptions were checked via Kaplan‒Meier curves and Schoenfeld residual plots. In the analysis for gastric cancer mortality, age and BMI changes were categorized to address assumption violations.\u003c/p\u003e\u003cp\u003eMachine learning analyses included random survival forest, XGBoost, and DeepSurv models. In the XGBoost and DeepSurv models, the Cox partial likelihood method was used to calculate the risk score and loss function. An 80:20 train‒test split was used for all the models. The concordance index (C-index) was calculated on the test set to assess model performance and comparison (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e). SHAP values were computed for the ML models to assess variable importance. All analyses were conducted using SAS 9.4 and Python with the Scikit survival, XGBoost, and PyTorch packages.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eAfter all exclusions were applied, 11,029 patients were included in the main analysis, and 4,945 were included in the subcohort analysis (Fig. \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003e). The median follow-up was 6.19 years. Baseline characteristics are shown in Table \u003cspan class=\"InternalRef\"\u003e1\u003c/span\u003eand Supplementary Table 1.\u003c/p\u003e\n\u003ctable id=\"Tab1\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eBaseline characteristics of the study population categorized by gastric cancer recurrence and mortality\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003eRecurrence\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"3\"\u003e\n \u003cp\u003eGastric cancer mortality\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eOverall\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDisease free\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRecurrence\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ep-value\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eSurvived\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDeath\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ep-value\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003en\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e11029\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10080\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e949\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10471\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e558\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAge (years), mean (SD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e62.7 (10.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e62.6 (10.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e63.0 (10.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.373\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e62.6 (10.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e64.9 (12.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAge group, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;60yrs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4204 (38.12)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3878 (38.47)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e326 (34.35)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.013\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4033 (38.52)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e171 (30.65)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026ge;\u0026thinsp;60yrs\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6825 (61.88)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6202 (61.53)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e623 (65.65)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6438 (61.48)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e387 (69.35)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSex, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eFemale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3321 (30.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3097 (30.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e224 (23.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3176 (30.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e145 (26.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.033\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMale\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7708 (69.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6983 (69.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e725 (76.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7295 (69.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e413 (74.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eT stage, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eT1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8942 (81.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8337 (82.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e605 (63.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e8806 (84.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e136 (24.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eT2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e711 (6.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e671 (6.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e40 (4.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e664 (6.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e47 (8.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eT3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e805 (7.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e687 (6.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e118 (12.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e657 (6.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e148 (26.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eT4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e571 (5.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e385 (3.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e186 (19.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e344 (3.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e227 (40.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN stage, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN0\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9744 (88.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9089 (90.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e655 (69.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9541 (91.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e203 (36.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e369 (3.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e318 (3.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e51 (5.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e313 (3.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e56 (10.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e422 (3.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e343 (3.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e79 (8.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e330 (3.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e92 (16.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e494 (4.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e330 (3.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e164 (17.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e287 (2.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e207 (37.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDifferentiation, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWell-differentiated\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3545 (32.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3201 (31.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e344 (36.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.008\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3482 (33.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e63 (11.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModerately differentiated\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4138 (37.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3819 (37.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e319 (33.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3953 (37.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e185 (33.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePoorly differentiated\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3346 (30.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3060 (30.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e286 (30.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3036 (29.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e310 (55.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTumor size, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026le;\u0026thinsp;1 cm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10866 (98.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9959 (98.8)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e907 (95.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e10367 (99.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e499 (89.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026gt;\u0026thinsp;1 cm\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e163 (1.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e121 (1.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e42 (4.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e104 (1.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e59 (10.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCEA, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNormal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7180 (65.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6659 (66.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e521 (54.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e6848 (65.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e332 (59.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eElevated\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e764 (6.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e677 (6.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e87 (9.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e675 (6.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e89 (15.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMissing\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3085 (28.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2744 (27.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e341 (35.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2948 (28.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e137 (24.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRemnant stomach, n (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWhole\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4139 (37.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e3622 (35.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e517 (54.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e4103 (39.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e36 (6.5)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eProximal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5416 (49.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5153 (51.1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e263 (27.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e5124 (48.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e292 (52.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDistal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e180 (1.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e171 (1.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e9 (0.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e173 (1.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e7 (1.3)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNone\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1294 (11.7)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1134 (11.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e160 (16.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e1071 (10.2)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e223 (40.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003c/p\u003e\n\u003cp\u003eThe Cox model results (Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e) revealed that remnant stomach was the strongest predictor of recurrence. Patients with a whole stomach after ESD had an approximately 5-fold higher recurrence rate than those with partial or no stomach after gastrectomy. Conversely, gastric cancer mortality was higher among patients who underwent gastrectomy, particularly those who underwent total gastrectomy. The T and N stages were also significant for both outcomes. Older age, male sex and elevated CEA levels were associated with increased mortality risk. However, tumor size and histologic differentiation were not significant.\u003c/p\u003e\n\u003cp\u003eIn the subcohort analysis (Supplementary Table 2), most factors were not significantly associated with prognosis. However, a decrease in BMI after treatment was associated with decreased mortality. Patients who initiated smoking after treatment had universally poor outcomes, precluding HR estimation. \u003cem\u003eH. pylori\u003c/em\u003e eradication was significantly associated with lower mortality but not recurrence. However, Cox analysis stratified by treatment method revealed that \u003cem\u003eH. pylori\u003c/em\u003e eradication significantly reduced both gastric cancer recurrence and mortality only in patients who underwent ESD (Supplementary Table 3). Patients who had a family history of gastric cancer had favorable outcomes in terms of both recurrence and mortality.\u003c/p\u003e\n\u003cp\u003eThe SHAP results of the ML models were similar (Figs. \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e). In contrast to the other ML models, XGBoost identified age as the most important variable in recurrence. Model comparison using the C-index revealed no large differences; all the models had a C-index\u0026thinsp;\u0026gt;\u0026thinsp;0.7 (Table \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2. Hazard ratio of gastric cancer recurrence and mortality\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\n \u003cp\u003e \u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e \u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"bottom\" style=\"width: 26px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecurrence\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd colspan=\"2\" valign=\"bottom\" style=\"width: 28px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMortality\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eUnivariate\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMultivariate\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eUnivariate\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMultivariate\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\n \u003cp\u003e \u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e \u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHR (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHR (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHR (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eHR (95% CI)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\n \u003cp\u003eSex\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eMale (vs. Female)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e1.43 (1.23-1.66)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1.37 (1.18-1.59)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e1.24 (1.03-1.50)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1.35 (1.11-1.63)*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\n \u003cp\u003eAge\u0026nbsp;\u0026dagger;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003econt./\u0026ge;60yrs (vs. \u0026lt;60yrs)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e1.01 (1.00-1.01)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1.00 (1.00-1.01)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e1.45 (1.21-1.73)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1.80 (1.50-2.17)*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\n \u003cp\u003eT stage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eT2 (vs. T1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e0.85 (0.62-1.17)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e2.08 (1.40-3.11)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e4.51 (3.24-6.28)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e2.15 (1.49-3.11)*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eT3 (vs. T1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e2.43 (1.99-2.96)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e4.16 (2.94-5.87)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e13.46 (10.66-16.99)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e4.53 (3.31-6.22)*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eT4 (vs. T1)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e6.41 (5.44-7.56)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e7.73 (5.40-11.06)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e33.43 (27.01-41.37)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e7.48 (5.37-10.43)*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\n \u003cp\u003eN stage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eN1 (vs. N0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e2.28 (1.72-3.04)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e2.14 (1.48-3.10)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e7.92 (5.89-10.64)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1.79 (1.27-2.53)*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eN2 (vs. N0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e3.18 (2.52-4.02)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e2.94 (2.11-4.09)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e11.74 (9.17-15.02)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e2.46 (1.82-3.33)*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eN3 (vs. N0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e6.69 (5.63-7.94)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e4.39 (3.21-5.99)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e26.29 (21.64-31.92)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e4.06 (3.07-5.37)*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\n \u003cp\u003eRemnant stomach\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eDistal (vs. Whole)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e0.39 (0.20-0.75)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.17 (0.08-0.33)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e4.53 (2.02-10.17)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e2.07 (0.90-4.74)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eNone (vs. Whole)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e1.04 (0.87-1.24)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.17 (0.13-0.23)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e21.72 (15.27-30.88)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e3.87 (2.53-5.93)*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eProximal (vs. Whole)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e0.37 (0.32-0.43)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.13 (0.10-0.17)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e6.28 (4.45-8.88)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e2.27 (1.51-3.41)*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\n \u003cp\u003eSize\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003e\u0026gt;1 cm (vs. \u0026le;1 cm)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e3.94 (2.89-5.37)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1.19 (0.85-1.66)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e9.90 (7.56-12.97)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1.33 (0.99-1.77)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\n \u003cp\u003eDifferentiation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eModerately differentiated (vs. WD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e0.81 (0.69-0.94)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.96 (0.81-1.13)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e2.57 (1.93-3.42)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e0.92 (0.67-1.26)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003ePoorly differentiated (vs. WD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e0.90 (0.77-1.05)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1.02 (0.82-1.27)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e5.44 (4.15-7.13)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1.02 (0.75-1.41)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\n \u003cp\u003eCEA\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eElevated (vs. Normal)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e1.70 (1.35-2.13)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1.24 (0.98-1.56)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e2.69 (2.13-3.40)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1.42 (1.12-1.81)*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 21px;\"\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 23px;\"\u003e\n \u003cp\u003eMissing (vs. Normal)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 12px;\"\u003e\n \u003cp\u003e1.56 (1.36-1.79)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1.11 (0.96-1.28)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15px;\"\u003e\n \u003cp\u003e0.70 (0.96-0.79)*\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 13px;\"\u003e\n \u003cp\u003e1.11 (0.90-1.36)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e*p value \u0026lt; 0.05\u003c/p\u003e\n\u003cp\u003e\u0026dagger;Age was treated as a categorical variable in mortality analysis\u003cstrong\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eComparison between machine learning models and statistical analysis\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eModel\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eRecurrence\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\" colspan=\"2\"\u003e\n \u003cp\u003eMortality\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\u0026nbsp;\u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ec-index\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eImportant variables\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003ec-index\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eImportant variables\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCox proportional hazard model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7560\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRemnant stomach, T stage, N stage, Sex\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8979\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eT stage, N stage, Remnant stomach, CEA level, Sex, Age\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRandom survival forest model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7648\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRemnant stomach, T stage, N stage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.9196\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eN stage, T stage, Age, Remnant stomach\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eXGBoost model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7300\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAge, T stage, Remnant stomach, Grade, N stage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8916\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAge, N stage, T stage, Remnant stomach, Grade\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDeepSurv model\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.7745\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRemnant stomach, T stage, N stage\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"char\"\u003e\n \u003cp\u003e0.8907\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eAge, Remnant stomach, T stage, N stage\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003ctfoot\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\"\u003e*The selected variables were age, sex, T stage, N stage, tumor size, Remnant stomach after treatment, and the CEA level and c-index, which were calculated with a 20% randomly sampled test dataset.\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd colspan=\"5\"\u003e**Variables were selected by significance and SHAP value importance and ordered by importance\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tfoot\u003e\n\u003c/table\u003e\n"},{"header":"Discussion","content":"\u003cp\u003eThis study demonstrated that ML models can produce comparable results to conventional Cox models, even in the presence of assumption violations. Age and BMI changes, which violated proportional hazard assumptions, were effectively handled in the ML models without requiring categorization, highlighting the flexibility of ML models. SHAP values provided interpretability to ML models, aligning closely with Cox model findings. This suggests that SHAP can effectively address the \"black box\" nature of ML, making it more suitable for clinical applications. However, the SHAP value does not provide the risk ratio or significance presented in most medical studies. The SHAP value indicates the importance of the feature in the model (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e). The important features were discovered with the SHAP value, but the interpretation was less clear than that of the statistical model. Although the SHAP\u0026rsquo;s ability to interpret ML models is limited, its application is more flexible than that of a statistical model. SHAP can be applied for both linear and nonlinear models, even for the Cox model and deep learning model. As shown in our study, SHAP can be used as a good alternative option when statistical analysis cannot be performed.\u003c/p\u003e\u003cp\u003eWe used random survival forest, XGBoost, and DeepSurv models. Although there was no significant difference between the models, they yielded different SHAP results. This might be due to the different modeling processes among the models. Thus, although the models showed similar c-indices SHAP value should be interpreted with caution.\u003c/p\u003e\u003cp\u003eSeveral medical studies have compared conventional statistical models and machine learning models (\u003cspan additionalcitationids=\"CR18\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e). Compared with conventional statistical models, ML has strong advantages in terms of application to high-dimensional data, accuracy, and flexible modeling. Previous studies suggested the possibility of applying ML in medical studies, which is supported by our findings. Furthermore, the usefulness of the ML model was emphasized by the SHAP results in our study.\u003c/p\u003e\u003cp\u003eOur findings confirmed known prognostic factors, such as remnant stomach status, tumor stage, and sex(\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan additionalcitationids=\"CR21\" citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e). In 2018, a randomized controlled study revealed that \u003cem\u003eH. pylori\u003c/em\u003e eradication in post-ESD patients decreased the recurrence of gastric cancer(\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e), which led to \u003cem\u003eH. pylori\u003c/em\u003e eradication as a standard treatment after ESD in Korea. In our study, \u003cem\u003eH. pylori\u003c/em\u003e eradication was effective only in patients who underwent ESD. Patients whose stomach anatomy was altered after gastrectomy did not show an association between \u003cem\u003eH. pylori\u003c/em\u003e eradication and either recurrence or mortality. Because our data did not include \u003cem\u003eH. pylori\u003c/em\u003e status, we could only consider prescriptions for \u003cem\u003eH. pylori\u003c/em\u003e eradication. Although the assessment of \u003cem\u003eH. pylori\u003c/em\u003e was limited, the prognosis associated with its eradication was similar to that reported in previous studies. Inconsistent results concerning the effects of \u003cem\u003eH. pylori\u003c/em\u003e eradication have been reported in patients undergoing gastrectomy(\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e). The role of \u003cem\u003eH. pylori\u003c/em\u003e eradication was evident in patients who underwent ESD but not in those who underwent gastrectomy; therefore, further studies with better designs are needed for patients undergoing gastrectomy. Tumor size is also known as an independent prognostic factor(\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e). However, our study revealed a greater risk of recurrence and mortality in patients with tumors larger than 1 cm than in those with tumors smaller than 1 cm, but the difference was not statistically significant, which was likely related to the data distribution. Differentiation and family history are factors whose associations have not been clearly identified, as previous studies have shown inconsistent results. However, recent studies indicated that there was no significant association with differentiation, which was consistent with our study (\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e). A Korean study of 1,273 patients with gastric cancer reported favorable outcomes of patients with a family history of gastric cancer, which was also shown in our study (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e). The CCI is a factor that is designed for adjusting mortality. However, a previous study indicated that the CCI may not be associated with gastric cancer mortality(\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e), which was consistent with our study.\u003c/p\u003e\u003cp\u003eModifiable risk factors before treatment have limited prognostic value, but posttreatment changes, particularly smoking, may influence outcomes. Other factors also showed favorable outcomes with favorable behavior changes, but these findings were not statistically significant. Further studies with more comprehensive behavioral data are needed.\u003c/p\u003e\u003cp\u003eOur study had limitations, which included the use of administrative data lacking clinical test results. This leads to potential misclassification of recurrence or \u003cem\u003eH. pylori\u003c/em\u003e status and the absence of important variables. Pathology results, such as depth of invasion, lymphatic invasion, and margin involvement, which are well-known risk factors, were absent (\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e). Atrophic gastritis and intestinal metaplasia are also known risk factors for gastric cancer and its recurrence(\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e), which were not included in our data. In addition, patient-related factors such as morbidity were not fully accessible. Patients may not have received treatment when they experienced recurrence because of their health condition. The detailed information of patients was not recorded in the claim data, which led to misclassification of the endpoint. However, there were no missing data for mortality in our study; therefore, analysis with gastric cancer mortality as the endpoint revealed no exceptional misclassification. Additionally, findings may not be generalizable outside of Korea because of the different natures of gastric cancer (\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e). To generalize these findings worldwide, differences in ethnicity should be considered.\u003c/p\u003e\u003cp\u003eIn conclusion, machine learning offers a viable alternative to conventional survival analysis, particularly when statistical assumptions are violated. SHAP enhances model interpretability, supporting its integration into precision medicine. Key prognostic factors such as the remnant stomach status and tumor stage should guide the level of follow-up. Posttreatment behavioral modifications may offer additional survival benefits, which warrants further study.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAuthor Contributions:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJL designed and conducted the study; acquired, analyzed, and interpreted the data; drafted the manuscript. AS provided administrative, technical, and financial support; reviewed the manuscript, and approved the final draft submitted.\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eAcknowledgements:\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank American Journal Experts (AJE) for English language editing of this manuscript (certificate verification code: 19BC-7A1E-FC0B-98F7-1F8E). This study used the K-CURE cancer public library database, which was established by the National Cancer Data Center as part of the K-CURE project organized by the Ministry of Health and Welfare (Research number: :KC20240712002)\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e:\u003c/p\u003e\n\u003cp\u003eThis study was supported by Cancer Research Institute, Seoul National University.\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Institutional Review Board of Seoul National University Hospital (IRB No. E-2401-057-1501). The need for informed consent was waived by the IRB, as only anonymized data were used. This study was conducted in accordance with the principles of the Declaration of Helsinki.\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no other conflict of interest.\u003c/p\u003e\n\n\u003cp\u003e\u003cstrong\u003eData availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data used in this study are available through the K-CURE project (https://k-cure.mohw.go.kr/) upon approval by the Institutional Review Board (IRB) and Data Review Board (DRB) of participating institutions. Data can only be accessed and analyzed in a secure Central Data Center; only analysis results may be exported.\u003c/p\u003e\n"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003ePark EH, Jung KW, Park NJ, Kang MJ, Yun EH, Kim HJ, et al. Cancer Statistics in Korea: Incidence, Mortality, Survival, and Prevalence in 2022. Cancer Res Treat. 2025;57(2):312-30.\u003c/li\u003e\n\u003cli\u003eIn H, Solsky I, Palis B, Langdon-Embry M, Ajani J, Sano T. Validation of the 8th Edition of the AJCC TNM Staging System for Gastric Cancer using the National Cancer Database. Annals of Surgical Oncology. 2017;24(12):3683-91.\u003c/li\u003e\n\u003cli\u003eKim TH, Kim IH, Kang SJ, Choi M, Kim BH, Eom BW, et al. Korean Practice Guidelines for Gastric Cancer 2022: An Evidence-based, Multidisciplinary Approach. J Gastric Cancer. 2023;23(1):3-106.\u003c/li\u003e\n\u003cli\u003eEom SS, Ryu KW, Han HS, Kong S-H. A Comprehensive and Comparative Review of Global Gastric Cancer Treatment Guidelines: 2024 Update. J Gastric Cancer. 2025;25(1):153-76.\u003c/li\u003e\n\u003cli\u003eChoi IJ, Kook MC, Kim YI, Cho SJ, Lee JY, Kim CG, et al. Helicobacter pylori Therapy for the Prevention of Metachronous Gastric Cancer. N Engl J Med. 2018;378(12):1085-95.\u003c/li\u003e\n\u003cli\u003eRajkomar A, Dean J, Kohane I. Machine Learning in Medicine. New England Journal of Medicine. 2019;380(14):1347-58.\u003c/li\u003e\n\u003cli\u003eWang P, Li Y, Reddy CK. Machine Learning for Survival Analysis: A Survey. ACM Comput Surv. 2019;51(6):Article 110.\u003c/li\u003e\n\u003cli\u003eLundberg SM, Lee S-I. A unified approach to interpreting model predictions. Advances in neural information processing systems. 2017;30.\u003c/li\u003e\n\u003cli\u003eIshwaran H, Kogalur UB, Blackstone EH, Lauer MS. Random survival forests. 2008.\u003c/li\u003e\n\u003cli\u003eChen T, Guestrin C. XGBoost: A Scalable Tree Boosting System. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; San Francisco, California, USA: Association for Computing Machinery; 2016. p. 785\u0026ndash;94.\u003c/li\u003e\n\u003cli\u003eKatzman JL, Shaham U, Cloninger A, Bates J, Jiang T, Kluger Y. DeepSurv: personalized treatment recommender system using a Cox proportional hazards deep neural network. BMC Medical Research Methodology. 2018;18(1):24.\u003c/li\u003e\n\u003cli\u003eChoi D-W, Guk MY, Kim HR, Ryu KS, Kong H-J, Cha HS, et al. Data Resource Profile: The Cancer Public Library Database in South Korea. Cancer Res Treat. 2024;56(4):1014-26.\u003c/li\u003e\n\u003cli\u003eJung H-K, Kang SJ, Lee YC, Yang H-J, Park S-Y, Shin CM, et al. Evidence-based Guidelines for the Treatment of Helicobacter pylori Infection in Korea: 2020 Revised Edition. Korean J Helicobacter Up Gastrointest Res. 2020;20(4):261-87.\u003c/li\u003e\n\u003cli\u003eHuang PL. A comprehensive definition for metabolic syndrome. Disease Models \u0026amp; Mechanisms. 2009;2(5-6):231-7.\u003c/li\u003e\n\u003cli\u003eQuan H, Li B, Couris CM, Fushimi K, Graham P, Hider P, et al. Updating and validating the Charlson comorbidity index and score for risk adjustment in hospital discharge abstracts using data from 6 countries. American journal of epidemiology. 2011;173(6):676-82.\u003c/li\u003e\n\u003cli\u003eKim K-H. Comparative study on three algorithms of the ICD-10 Charlson comorbidity index with myocardial infarction patients. Journal of Preventive Medicine and Public Health. 2010;43(1):42-9.\u003c/li\u003e\n\u003cli\u003eRajula HSR, Verlato G, Manchia M, Antonucci N, Fanos V. Comparison of Conventional Statistical Methods with Machine Learning in Medicine: Diagnosis, Drug Development, and Treatment. Medicina. 2020;56(9):455.\u003c/li\u003e\n\u003cli\u003eSpooner A, Chen E, Sowmya A, Sachdev P, Kochan NA, Trollor J, et al. A comparison of machine learning methods for survival analysis of high-dimensional clinical data for dementia prediction. Scientific Reports. 2020;10(1):20410.\u003c/li\u003e\n\u003cli\u003ePremsagar P, Aldous C, Esterhuizen TM, Gomes BJ, Gaskell JW, Tabb DL. Comparing conventional statistical models and machine learning in a small cohort of South African cardiac patients. Informatics in Medicine Unlocked. 2022;34:101103.\u003c/li\u003e\n\u003cli\u003eMantziari S, St Amour P, Abboretti F, Teixeira-Farinha H, Gaspar Figueiredo S, Gronnier C, et al. A Comprehensive Review of Prognostic Factors in Patients with Gastric Adenocarcinoma. Cancers. 2023;15(5):1628.\u003c/li\u003e\n\u003cli\u003eSun B, Zhang H, Wang J, Cai H, Xuan Y, Xu D. Tumor Location Causes Different Recurrence Patterns in Remnant Gastric Cancer. J Gastric Cancer. 2022;22(4):369-80.\u003c/li\u003e\n\u003cli\u003eLiu Q, Ding L, Qiu X, Meng F. Updated evaluation of endoscopic submucosal dissection versus surgery for early gastric cancer: A systematic review and meta-analysis. International Journal of Surgery. 2020;73:28-41.\u003c/li\u003e\n\u003cli\u003eKim Y-I, Cho S-J, Lee JY, Kim CG, Kook M-C, Ryu KW, et al. Effect of Helicobacter pylori Eradication on Long-Term Survival after Distal Gastrectomy for Gastric Cancer. Cancer Res Treat. 2016;48(3):1020-9.\u003c/li\u003e\n\u003cli\u003eZhao Z, Zhang R, Chen G, Nie M, Zhang F, Chen X, et al. Anti\u0026ndash;Helicobacter pylori Treatment in Patients With Gastric Cancer After Radical Gastrectomy. JAMA Network Open. 2024;7(3):e243812-e.\u003c/li\u003e\n\u003cli\u003eIm WJ, Kim MG, Ha TK, Kwon SJ. Tumor size as a prognostic factor in gastric cancer patient. J Gastric Cancer. 2012;12(3):164-72.\u003c/li\u003e\n\u003cli\u003eFeng F, Liu J, Wang F, Zheng G, Wang Q, Liu S, et al. Prognostic value of differentiation status in gastric cancer. BMC Cancer. 2018;18(1):865.\u003c/li\u003e\n\u003cli\u003eHan MA, Oh MG, Choi IJ, Park SR, Ryu KW, Nam B-H, et al. Association of Family History With Cancer Recurrence and Survival in Patients With Gastric Cancer. Journal of Clinical Oncology. 2012;30(7):701-8.\u003c/li\u003e\n\u003cli\u003eKyung M-H, Yoon S-J, Ahn H-S, Hwang S-m, Seo H-J, Kim K-H, et al. Prognostic impact of Charlson comorbidity index obtained from medical records and claims data on 1-year mortality and length of stay in gastric cancer patients. Journal of Preventive Medicine and Public Health. 2009;42(2):117-22.\u003c/li\u003e\n\u003cli\u003eHatta W, Gotoda T, Oyama T, Kawata N, Takahashi A, Yoshifuku Y, et al. A Scoring System to Stratify Curability after Endoscopic Submucosal Dissection for Early Gastric Cancer: \u0026ldquo;eCura system\u0026rdquo;. Official journal of the American College of Gastroenterology | ACG. 2017;112(6):874-81.\u003c/li\u003e\n\u003cli\u003eFactors influencing occurrence of metachronous gastric cancer after endoscopic resection: a systematic review and meta-analysis FAU - Choe, Younghee FAU - Park, Jae Myung FAU - Kim, Joon Sung FAU - Cho, Yu Kyung FAU - Kim, Byung-Wook FAU - Choi, Myung-Gyu. Korean J Intern Med. 2023;38(6):831-43.\u003c/li\u003e\n\u003cli\u003eKendrick P, Kelly YO, Baumann MM, Compton K, Blacker BF, Daoud F, et al. The burden of stomach cancer mortality by county, race, and ethnicity in the USA, 2000\u0026amp;#x2013;2019: a systematic analysis of health disparities. The Lancet Regional Health \u0026ndash; Americas. 2023;24.\u003c/li\u003e\n\u003cli\u003eKim J, Sun CL, Mailey B, Prendergast C, Artinyan A, Bhatia S, et al. Race and ethnicity correlate with survival in patients with gastric adenocarcinoma. Annals of Oncology. 2010;21(1):152-60.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"discover-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"dion","sideBox":"Learn more about [Discover Oncology](https://www.springer.com/12672)","snPcode":"","submissionUrl":"","title":"Discover Oncology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Gastric cancer, Prognosis, Cohort, Machine learning, SHAP","lastPublishedDoi":"10.21203/rs.3.rs-7022055/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7022055/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eAs the prognosis of gastric cancer has improved, the exploration of prognostic factors has become increasingly important. This study aimed to identify prognostic factors of gastric cancer using machine learning and statistical methods and to compare the effectiveness of different methodologies in identifying prognostic factors.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eWe conducted a retrospective cohort study of cancer research data from survivors of gastric cancer in Korea. Patients were followed up from the date of curative treatment of gastric cancer to the date of recurrence, cancer-specific death, or censoring. The Cox proportional hazards, random survival forest, XGBoost, and DeepSurv models were used to calculate the risk of recurrence and cancer-specific death. All the models were trained on 80% of the training set, and the concordance index was used for comparison with 20% of the test set. The SHAP value was used for variable interpretation in the machine learning models.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eA total of 11,029 gastric cancer survivors with a median follow-up time of 6.19 years were included. Remnant stomach after gastric cancer treatment, T stage and N stage were the most important features for recurrence and mortality according to both the Cox model and the machine learning model. All the models had a concordance index greater than 0.7 without large differences.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eThe machine learning model is not inferior to conventional statistical analysis models and offers greater flexibility, especially when statistical assumptions are violated. The key prognostic factors identified through this approach include residual stomach after treatment and cancer stage.\u003c/p\u003e","manuscriptTitle":"Prognostic Factors in Survivors of Gastric Cancer: A Comparative Study of Cox Proportional Hazards Model and Machine Learning Approaches Using Korean National Cohort Data","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-08 16:32:52","doi":"10.21203/rs.3.rs-7022055/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-09-04T08:07:33+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-28T15:34:01+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-20T14:03:46+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"103300700378927021697890909771523928207","date":"2025-08-12T03:48:35+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"182780003202531372046865269290273514292","date":"2025-08-06T04:57:04+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"265417332552355095407205583633571679470","date":"2025-08-05T11:02:42+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-08-05T09:53:14+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-08-02T07:01:25+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-07-29T14:16:16+00:00","index":"","fulltext":""},{"type":"submitted","content":"Discover Oncology","date":"2025-07-29T14:13:16+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"discover-oncology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"dion","sideBox":"Learn more about [Discover Oncology](https://www.springer.com/12672)","snPcode":"","submissionUrl":"","title":"Discover Oncology","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Discover Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f183312d-9214-4e5c-8b33-b8f057dc3306","owner":[],"postedDate":"August 8th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2025-11-10T05:23:04+00:00","versionOfRecord":[],"versionCreatedAt":"2025-08-08 16:32:52","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7022055","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7022055","identity":"rs-7022055","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}