Identification of clinical biomarkers for adjuvant chemotherapy for gastric cancer after D2 dissection by analyses of individual patents’ data from large randomized controlled trials

1. Preliminary analysisWe will first verify the integrity of individual patient data (IPD) from ACTS-GC, CLASSIC, and SAMIT by comparing their descriptive statistics against the published figures. Relapse-free survival and overall survival according to treatment groups and trials will be described by Kaplan-Meier methods. Distributions of clinical and pathological factors, including age, sex, performance status, weight, times since surgery, cancer stage, tumor stage, tumor location, nodal status, and, histrologic type, will be described as well and will be compared across trials by ANOVA or the Fisher exact tests. For adjuvant groups, hazard ratios of capecitabine and oxaliplatin in CLASSIC vs. S-1 in ACTS-GC, and paclitaxel then S-1 in SAMIT vs. S-1 in ACTS-GC will be estimated by Cox regression adjusted for clinical and pathological factors. 2. Identification for prognostic factors of relapse-free survival and overall survivalIn surgery alone groups of ACTS-GC and CLASSIC trials, prognostic factors for relapse-free survival or overall survival will be identified by stratified Cox regression using trial as the stratified factor, which is correspond to fixed effects meta-analysis. The clinical and pathological factors will be screened through a backward variable selection with the critical value of p = 0.1. The hazard ratios, 95% confidence intervals, p values and the baseline survival functions of the final models will be estimated and a postoperative survival prediction model will be developed based on the estimates. The prognostic factors identified will be validated and may be modified in adjuvant groups of ACTS-GC, CLASSIC, and SAMIT by fitting stratified Cox regression with the same covariates. 3. Identification for predictive factors of efficacy of adjuvant chemotherapyPredictive factors for adjuvant chemotherapy will be identified by examining the difference in relapse-free survival or overall survival between treatment groups, i.e. adjuvant plus surgery vs. surgery alone, in all combined patients of ACTS-GC, CLASSIC, and SAMIT. Relapse-free survival and overall survival according to treatment groups and subgroups (defined based on the clinical and pathological factors) will be described by Kaplan-Meier methods. The hazard ratios, 95% confidence intervals, and p values of the treatment groups will be estimated by stratified Cox regression using trial as the stratified factor. Tests for treatment-subgroup interactions will be examined by stratified Cox regression including a treatment-subgroup interaction as a covariate and trial as the stratified factor. The hazard ratios, 95% confidence intervals, p values and the baseline survival functions of the final models which includes significant interactions will be estimated and a postoperative survival prediction model will be developed based on the estimates. 4. Identification for predictive factors of efficacy of specific regimensIf the hazard ratio of capecitabine and oxaliplatin in CLASSIC vs. S-1 in ACTS-GC, or paclitaxel then S-1 in SAMIT vs. S-1 in ACTS-GC are significant in preliminary analysis, we further explore predictive factors of efficacy of these regimens. Relapse-free survival and overall survival according to treatment groups and subgroups (defined based on the clinical and pathological factors) will be described by Kaplan-Meier methods. The hazard ratios, 95% confidence intervals, and p values of the treatment groups will be estimated by Cox regression adjusted for clinical and pathological factors. Tests for treatment-subgroup interactions will be examined by Cox regression including a treatment-subgroup interaction and clinical and pathological factors as covariates. The hazard ratios, 95% confidence intervals, p values and the baseline survival functions of the final models which includes significant interactions will be estimated and a postoperative survival prediction model will be developed based on the estimates.5. Validation of prediction modelsWe will assess the predictive accuracy of the prediction models using 10-fold cross-validation, i.e., we performed 10 rounds of cross-validation using different partitions. One round of cross-validation involved randomly partitioning data on the entire population into complementary subsets, fitting the final Cox regression models to one subset of 90% of patients, and validating the model on the remaining subset. Calibration, namely, how closely the prediction reflected observed events, was assessed for each event by the HosmerLemeshow test and the mean of observedto-predicted (O/P) ratios, which was calculated as the mean of ratios of the observed-to-expected events across the strata used in the Hosmer-Lemeshow test. Discrimination, the ability to distinguish between those who experienced the event and those who did not, was evaluated using Harrell C statistics, the proportion of all patient pairs in which the predictions of the model and observed events were concordant. 6. GeneralAll reported p values will be two-tailed and p<0.05 is chosen as the threshold for statistical significance. An academic statistician (ST) will conduct all statistical analyses using SAS version 9.3 (SAS Institute, Cary, NC, USA). Missing data will be substituted using the multiple imputation method.