Article Text
PDF
Abstract
Objective This study aimed to develop and validate robust predictive models for patients with oesophageal cancer who achieved a pathological complete response (pCR) and those who did not (non-pCR) after neoadjuvant therapy and oesophagectomy.
Design Clinicopathological data of 6517 primary oesophageal cancer patients who underwent neoadjuvant therapy and oesophagectomy were obtained from the National Cancer Database for the training cohort. An independent cohort of 444 Chinese patients served as the validation set. Two distinct multivariable Cox models of overall survival (OS) were constructed for pCR and non-pCR patients, respectively, and were presented using web-based dynamic nomograms (graphical representation of predicted OS based on the clinical characteristics that a patient could input into the website). The calibration plot, concordance index and decision curve analysis were employed to assess calibration, discrimination and clinical usefulness of the predictive models.
Results In total, 13 and 15 variables were used to predict OS for pCR and non-pCR patients undergoing neoadjuvant therapy followed by oesophagectomy, respectively. Key predictors included demographic characteristics, pretreatment clinical stage, surgical approach, pathological information and postoperative treatments. The predictive models for pCR and non-pCR patients demonstrated good calibration and clinical utility, with acceptable discrimination that surpassed that of the current tumour, node, metastases staging system.
Conclusions The web-based dynamic nomograms for pCR (https://predict-survival.shinyapps.io/pCR-eso/) and non-pCR patients (https://predict-survival.shinyapps.io/non-pCR-eso/) developed in this study can facilitate the calculation of OS probability for individual patients undergoing neoadjuvant therapy and radical oesophagectomy, aiding clinicians and patients in making personalised treatment decisions.
- OESOPHAGEAL CANCER
- SURGICAL ONCOLOGY
- OESOPHAGEAL SURGERY
- SURVEILLANCE
Data availability statement
Data may be obtained from a third party and are not publicly available. The data will not be made available publicly.
This is an open access article distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited, appropriate credit is given, any changes made indicated, and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/.
Statistics from Altmetric.com
WHAT IS ALREADY KNOWN ON THIS TOPIC
For oesophageal cancer patients receiving neoadjuvant therapy followed by oesophagectomy, those who achieved pathological complete response (pCR) had better prognoses than those who did not (non-pCR).
Clinical tools are still lacking to predict the long-term survival for those patients.
WHAT THIS STUDY ADDS
Distinct predictive models with robust performance were constructed and externally validated for pCR (involving 13 clinical variables) and non-pCR (involving 15 variables) patients.
For pCR patients, age, sex, year of diagnosis, comorbidity score, histology type, differentiation grade, pretreatment cT, cN stage and cM stage, surgical approach, total number of harvested lymph nodes (LNs), length of postoperative hospital stay and readmission within 30 days after discharge should be considered to predict the overall survival (OS).
For non-pCR patients, age, sex, year of diagnosis, histology type, differentiation grade, ypT stage, LN ratio, ypM stage, lymphovascular invasion, surgical margin, surgical approach, total number of harvested LNs, length of postoperative hospitalisation, readmission within 30 days of discharge and adjuvant therapy were associated with the OS.
HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY
The web-based dynamic nomograms constructed and validated could easily calculate the survival probability for pCR and non-pCR patients, enabling clinicians and patients to decide on individualised medical treatments.
Introduction
Oesophageal cancer ranked sixth in mortality and seventh in incidence among all malignancies in 2020.1 Based on the short-term and long-term results of several randomised controlled trials,2–6 neoadjuvant therapy consisting of chemoradiation followed by oesophagectomy has become the established standard treatment for locally advanced oesophageal cancer. Recently, the combination of immunotherapy7 8 and removal of radiotherapy9–17 have been explored in the field of neoadjuvant therapy.
Pathological complete response (pCR), widely defined as ypT0N0M0 with R0 resection, is considered a promising surrogate endpoint for evaluating the effectiveness of neoadjuvant therapy.18 A previous study has demonstrated that achieving pCR in patients with oesophageal cancer is associated with significantly improved survival compared with non-pCR patients.19 According to the latest evidence and current guidelines,20 active surveillance is recommended for pCR patients, while non-pCR patients who received chemoradiation before oesophagectomy may benefit from adjuvant immunotherapy.21
However, retrospective studies have shown that 23.4%–29.3% of pCR patients still experience recurrence or metastasis,22 23 highlighting the need to identify high-risk pCR patients who require more intensive follow-up and/or adjuvant treatment. Concurrently, the high cost and extended duration of adjuvant immunotherapy pose considerable challenges for non-pCR patients. Thus, investigating whether adjuvant therapy is necessary for all patients who do not achieve a pCR is crucial.
Therefore, this study aimed to develop and validate robust predictive models for pCR and non-pCR patients using a large-scale database and an independent external cohort. Additionally, the study aimed to enhance clinical utility by implementing web-based dynamic nomograms, which are expected to improve the accuracy of medical decisions for patients with oesophageal cancer.
Methods
Data extraction
This study included patients with primary oesophageal cancer who received neoadjuvant therapy (radiation and/or systemic therapy) followed by radical oesophagectomy. The inclusion and exclusion criteria were listed in online supplemental eTable 1. Demographic characteristics (age, sex, race, Charlson-Deyo Comorbidity Score24–26 (online supplemental eTable 2, ignoring cancer condition), history of cancer and year of diagnosis), neoadjuvant and adjuvant treatment details (modality, total radiation dose and regimen), perioperative information (pretreatment clinical tumour, node, metastases (TNM) status, surgical approach, total harvested lymph nodes (LNs), length of stay (LOS) after surgery and readmission within 30 days after discharge), pathological data (tumour location, histology, differentiation grade, ypTNM status, surgical margin, lymphovascular invasion and number of positive LNs) and active follow-up information were retrieved. Records with missing data were excluded from the analysis.
Supplemental material
The primary outcome, overall survival (OS), was measured from the time of oesophageal cancer diagnosis until death or the last follow-up. Patients who survived the last follow-up were considered censored observations. The sixth or seventh American Joint Committee on Cancer (AJCC) editions (AJCC-6 and AJCC-7) of the TNM stages were converted into the latest AJCC-8 edition.27 28 The pCR was defined as ypT0N0M0R0 without lymphovascular invasion. The lymph node ratio (LNR) was calculated as the ratio of pathologically confirmed metastatic LNs to total harvested LNs.29
Patient population
The training cohort was derived from the National Cancer Database (NCDB), a nationwide institution-based programme covering 72% of newly diagnosed cancer cases in the USA.30 A total of 20 963 patients diagnosed between 2004 and 2015 were screened. Based on the inclusion and exclusion criteria (online supplemental eTable 1), 6517 patients were included in the training cohort. Among them, 1051 patients (16.1%) achieved pCR after neoadjuvant therapy (online supplemental eFigure 1).
External validation used an independent dataset consisting of prospectively collected cases from four multicentre clinical trials conducted in China (NCT03792347,7 NCT04435197,31 NCT0451341832 and NCT0300159614 15 33) and retrospective data of patients diagnosed between 2013 and 2022 at Shanghai Ruijin Hospital. A total of 444 patients were included, 92 (20.7%) of whom achieved pCR (online supplemental eFigure 2).
Patient and public involvement
Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.
Statistical analysis
Statistical analyses were conducted using R software (V.4.2.2, R Foundation for Statistical Computing, Vienna, Austria). Categorical variables were presented as frequencies and percentages. Continuous variables were presented as median and IQR owing to non-normal distribution. A significance level of p<0.05 was considered statistically significant; a 95% CI was used.
Univariable survival analyses were conducted using the Kaplan-Meier method, providing estimates of median survival times.34 Survival curves were compared using the log-rank test. Multivariable survival analyses were performed using the Cox proportional hazard regression model and HRs were calculated for each variable.35 The proportionality of hazards was assessed by the Schoenfeld residual test. The distribution of residuals was evaluated by the visualisation of Diagnostics for Beta Changes (DFBETAS) values, referring to the change in the estimated coefficients of the model when each observation was omitted one at a time.
Covariates for the predictive model were selected through a three-step process. First, the least absolute shrinkage and selection operator (LASSO) algorithm used cross-validation to determine the optimal tuning parameter λ and corresponding variables.36 The λmin value was determined as the one with the smallest mean squared error, while the λ1se value was obtained within one SE of λmin. If the λmin resulted in a model with an excessive number of covariates, the λ1se value was chosen to obtain a more regularised model. Second, backward stepwise regression was applied to simplify the model until the Akaike Information Criterion (AIC)37 no longer decreased. This process aimed to derive a model with improved goodness-of-fit and fewer parameters. The variables included in the final predictive model were adjusted based on a combination of statistical significance (p<0.05) and clinical relevance (factors improved to be related with the prognosis based on previous publications). Collinearity between the covariates was tested by calculating the variance inflation factor (VIF). An VIF >5 would indicate the existence of collinearity.
Construction and validation of the nomograms
The nomogram, a pictorial representation of a complex mathematical formula,38 was constructed to visually represent the predictive model, with each variable listed separately and assigned a corresponding score. The cumulative scores of all variables were used to calculate the total score, which led to the predicted probability of the endpoint.38 Additionally, a web-based dynamic nomogram was developed to allow automatic calculation of the predicted OS probability at a specific follow-up time by entering the patient characteristics.
The predictive performance of the models was evaluated in terms of calibration, discrimination and clinical usefulness. The calibration plots were used to assess the agreement between predicted and actual outcomes, with closer alignment to the diagonal indicating better concordance between predicted and observed OS rates. Discrimination was quantitatively evaluated using the area under the receiver operating characteristic curve (AUROC) and Harrell’s concordance index (C-index, the proportion of pairs of patients whose predicted survival times are correctly ordered among all pairs that can actually be ordered), which indicated the ability of the model to correctly rank the predicted survival times of patients relative to the actual order.39 40 Higher values of these two indices indicate better discriminant ability. Regarding clinical usefulness, decision curve analysis (DCA) was performed to demonstrate the net benefit of the predictive models at different threshold probabilities, representing the minimum probability for considering further intervention for the disease.41 42
Results
Distinct survival between pCR and non-pCR patients
In the training cohort, the median follow-up time was 49.2 months (IQR, 32.4–68.6), and at the last follow-up, 57.9% (3775/6517) of patients had died. The median OS was 67.8 months (95% CI 59.4 to 84.3) for pCR patients and 30.8 months (95% CI 29.2 to 31.8) for non-pCR patients.
In the external validation cohort, the median follow-up duration was 40.2 months (IQR, 31.6–51.7), and 34.9% (155/444) of patients had died at the last follow-up. Among pCR patients, the median OS was not reached owing to the limited number of deaths (11/92 died at the last follow-up), while non-pCR patients had a median OS of 51.0 months (95% CI 43.2 to not reached).
Kaplan-Meier curves and log-rank tests demonstrated that pCR patients had statistically significantly better OS than that of non-pCR patients in both the training (p<0.001) and external validation cohorts (p<0.001) (figure 1).
Kaplan-Meier survival curves of the included patients. PCR, pathological complete response.
Prognostic nomogram for pCR patients
A total of 1051 and 92 pCR patients were included in the training and external validation cohorts, respectively. The clinicopathological characteristics and median OS of each subgroup are presented in table 1 and online supplemental eTable 3. In the validation set, the median OS of pCR patients was not reached, as 11 of 92 of these patients died at the last follow-up.
Clinicopathological characteristics of patients achieving pathological complete response status
The process of identifying and selecting covariates was as follows: first, LASSO regression and cross-validation derived λmin=0.017 with 15 variables were included in the model (online supplemental eFigure 3 and eTable 4). Second, the backward stepwise regression was applied, leading to the retention of 10 variables (age, sex, year of diagnosis, comorbidity score, pretreatment cT stage, cN stage, cM stage, total harvested LNs, LOS after surgery and readmission within 30 days after discharge), with a minimum AIC of 5316.19 (online supplemental eTable 5). Finally, three variables (histology type,43 differentiation grade43 and surgical approach44–46) were added to the model considering their clinical association with prognosis of oesophageal cancer based on previous publications.
After variable selection, the OS predictive model for pCR patients included 13 covariates, whose HRs and 95% CIs are presented in online supplemental eTable 6. Significant risk factors associated with poorer prognosis included older age, male sex, comorbidity score ≥1, poorly differentiated or undifferentiated grade (G3-4), advanced pretreatment cT and cN stages, prolonged postoperative LOS, and readmission within 30 days after discharge. Significant protective factors included a later year of diagnosis and higher number of resected LNs. Although not statistically significant, the histology type,43 pretreatment cM stage47 and surgical approach44–46 were included in the final model owing to their clinical relevance. The proportionality of hazards was confirmed by the Schoenfeld residual test (p=0.079). The DFBETAS values uniformly distributed around zero and were relatively symmetrical (online supplemental eFigure 4), indicating that the distribution of residuals was normal. The VIFs of all the variables were lower than 2, indicating the absence of collinearity between the predictors.
Online supplemental eFigure 5 shows the final nomogram developed for predicting OS in patients who achieved pCR after neoadjuvant therapy. A web-based dynamic nomogram was constructed (https://predict-survival.shinyapps.io/pCR-eso/) to simplify the calculation process and facilitate its clinical application. As shown in figure 2, after entering the individual characteristics of the patient (left panel), the nomogram automatically calculates and intuitively displays both the Kaplan-Meier survival curve (upper right panel) and the estimated OS rate at a specific follow-up time with the corresponding 95% CI (bottom right panel).
Web-based dynamic nomogram for pCR patients of oesophageal cancer (https://predict-survival.shinyapps.io/pCR-eso/). OS, overall survival; PCR, pathological complete response.
Prognostic nomogram for non-pCR patients
Among the patients who did not achieve pCR after neoadjuvant therapy for oesophageal cancer, a total of 5466 records in the training and 352 records in the external validation set were included. The clinicopathological characteristics of the non-pCR patients and the median OS of each subgroup are shown in table 2 and online supplemental eTable 7.
Clinicopathological characteristics of non-pathological complete response patients
The variable selection process for the non-pCR model differed slightly from that of the pCR model. Initially, LASSO regression and cross-validation were applied, resulting in λmin=0.006, including 25 variables. To simplify the model, a λ1se=0.038 was further chosen, reducing the number of variables to 14 (online supplemental eFigure 6 and eTable 8). Subsequently, backward stepwise regression analysis retained the 13-variable model with a minimum AIC of 52 644.85 (online supplemental eTable 9). The final model was optimised by considering both clinical relevance and statistical significance. Similar to the model of pCR patients, the histology type,43 differentiation grade43 and surgical approach44–46 were included as they were clinically related to the prognosis of patients with oesophageal cancer. The number of positive LNs was excluded, while the LNR group was retained as a covariate, as our previous study has demonstrated the superior predictive performance of LNR in patients undergoing neoadjuvant therapy for oesophageal cancer.29 To avoid redundant information indicated by ypT stage and LNR, the ypTNM stage was excluded while the ypM stage was included to reflect the metastatic status.27 The optimal number of harvested LNs remained controversial.47–53 Thus, it was included as a continuous variable in the final model.
A total of 15 variables were retained for the OS predictive model in non-pCR patients, and online supplemental eTable 10 shows their HRs and 95% CIs. The risk factors affecting OS included older age, male sex, poor differentiation, advanced ypT stage, higher LNR, distant metastasis, lymphovascular invasion, residual disease at the surgical margin, prolonged hospital stay and readmission within 30 days of discharge. Protective factors included later year of diagnosis, robot-assisted surgery, higher number of harvested LNs and adjuvant therapy after surgery. The proportionality of hazards was rejected by the Schoenfeld residual test (p<0.001). The symmetrical DFBETAS values around zero (online supplemental eFigure 7) illustrated the normal distribution of residuals. The VIFs of all the variables were lower than 2, denoting the absence of collinearity between the predictors.
The final nomogram for predicting OS in non-pCR patients with oesophageal cancer after neoadjuvant therapy and oesophagectomy is shown in online supplemental eFigure 8. Owing to the complexity of the model, a web-based dynamic nomogram (https://predict-survival.shinyapps.io/non-pCR-eso/) was constructed (figure 3). By entering information of the patient in the left panel, long-term survival can be predicted. The nomogram generates a Kaplan-Meier survival curve (upper right panel) and estimates the OS rate with a 95% CI at a specific time point (bottom right panel).
Web-based dynamic nomogram for non-pCR patients with oesophageal cancer (https://predict-survival.shinyapps.io/non-pCR-eso/). OS, overall survival; PCR, pathological complete response.
Predictive performance and external validation of the nomograms
Calibration curves of 1-year, 3-year and 5-year OS for pCR (online supplemental eFigure 9A,B) and non-pCR patients (online supplemental eFigure 9C,D) were plotted in the training set (online supplemental eFigure 9A,C) and the external validation set (online supplemental eFigure 9B,D), respectively. Good calibration of both models was indicated in the training cohort as the predicted OS was close to the actual survival probability with narrow 95% CIs (online supplemental eFigure 9A,C). For external validation, the 95% CIs of both models were wide owing to the limited sample size, but the calibration curves were generally close to the diagonal (online supplemental eFigure 9B,D).
Discrimination analysis was performed using time-dependent AUROCs for pCR (online supplemental eFigure 10A,B) and non-pCR patients (online supplemental eFigure 10C,D) in the training (online supplemental eFigure 10A,C) and external validation set (online supplemental eFigure 10B,D). In both the training and external validation sets, the AUROC for the predictive model outperformed that of the current TNM staging system. The C-index of the OS predictive model for pCR patients was 0.645 (95% CI 0.618 to 0.672) in the training cohort and 0.573 (95% CI 0.408 to 0.738) in the validation cohort. For the non-pCR predictive model, the C-index was 0.652 (95% CI 0.642 to 0.662) in the training set and 0.692 (95% CI 0.649 to 0.735) in the external validation set.
To evaluate clinical utility, DCA (online supplemental eFigure 11A) demonstrated that the benefits of the clinical strategy based on this predictive model were greater than the currently applied strategy of active follow-up without intervention (grey line, intervention for none) for pCR patients when the threshold probability was below 70%. For non-pCR patients (online supplemental eFigure 11B), similar benefits were observed compared with the current recommendation of intervention for all non-pCR patients (dashed lines, intervention for all).
Discussion
Our study constructed distinct web-based nomograms for predicting the OS of pCR and non-pCR patients after neoadjuvant therapy followed by oesophagectomy with the largest sample size. Although several clinical variables have been identified to influence the prognosis of pCR patients in previous studies,22 23 54 models based on a limited population are far from comprehensive clinical utility. Furthermore, the survival outcomes and stratification information differed significantly between pCR and non-pCR patients, necessitating separate predictive models.
Following careful variable selection based on statistical significance and clinical relevance, our final Cox regression models included 13 and 15 factors for pCR and non-pCR patients, respectively. The identification of risk factors such as older age, male sex, higher comorbidity burden, poor tumour differentiation and advanced pretreatment stages is in accordance with previous studies22 23 55 56 and clinical practice. Moreover, the LOS after surgery and readmission within 30 days of discharge were included in the model as surrogate measures of postoperative complications. Our study revealed important factors, including the surgical approach and ypT stage, which call for a re-evaluation of current strategies in the diagnosis and treatment of oesophageal cancer.
Previous randomised controlled studies, including the MIRO trial44 and ROBOT trial,45 have demonstrated that both hybrid minimally invasive oesophagectomy and robotic oesophagectomy had lower complication rates than those of open surgery. Recently, an updated meta-analysis conducted by our team comparing robot-assisted versus conventional minimally invasive oesophagectomy found that robotic surgery resulted in a higher 3-year disease-free survival rate.46 Considering these results, the surgical approach was included in our predictive models. Multivariable Cox regression analyses showed that robotic surgery tended to prolong OS compared with open surgery for both non-pCR (HR, 0.73; 95% CI 0.60 to 0.88; p=0.001) and pCR (HR, 0.78; 95% CI 0.51 to 1.19; p=0.250) patients, although the protective effect was not statistically significant in the latter population.
Concurrently, the higher the ypT stage, the worse the survival of non-pCR patients observed in the multivariable Cox regression. However, the OS of ypTIS/T1/T2 patients was not significantly inferior to ypT0 patients (HR, 1.04; 95% CI 0.86 to 1.24; p=0.706). This raises a real problem in current clinical practice: for an oesophageal cancer patient with no regional LN or distant metastasis (ypN0M0) who undergoes R0 resection but has residual tumour in the submucosa (ypT1b) after neoadjuvant therapy, the necessity of adjuvant therapy should be questioned. Therefore, more rigorous prospective studies are warranted to address this issue.
Complex statistical models can be effectively visualised using a nomogram, but the clinical utility of a nomogram is of paramount importance in precision medicine. However, performing manual calculations in a clinical setting is not practical for physicians and patients. To address this, we have developed user-friendly web-based dynamic nomograms. By inputting readily accessible clinical characteristics, the predicted survival curve can be generated within seconds. Furthermore, the web-based platform allows for the simultaneous display of different survival curves driven by various adjuvant strategies, offering a clear explanation of the survival benefit associated with specific treatments (figure 3, upper right panel). Traditionally, nomograms assume that outcomes remain constant over time, which has been controversial.38 However, clinical practice has demonstrated that advancements in diagnostic methods, treatment approaches, and surgical techniques have led to improved prognoses for many malignancies over time. For oesophageal cancer, the CROSS study has ushered in a new era of neoadjuvant therapy, significantly enhancing treatment effectiveness and long-term prognosis for locally advanced cases. Consequently, we incorporated the year of diagnosis as a continuous variable in our predictive model.
In this study, we used a training cohort (NCDB database from the USA) to develop the predictive models and an external cohort (multicentre data from China) to validate the predictive performance of the models. Indeed, several differences in the distribution of race, histology type, tumour location and neoadjuvant modality were observed between the training and validation sets, as they were derived from different populations. We selected the NCDB as the training set owing to its large size, comprehensive clinical variables, and extensive follow-up records, enabling us to draw more comprehensive conclusions based on a larger scale of data from the USA. The validation set in the Chinese cohort allowed us to explore the generalisability of the results to a population with a high incidence of oesophageal cancer.57 The external validation demonstrated comparative calibration and even superior discrimination compared with the training cohort. This finding may have partially solved the gap between regional differences in Eastern and Western populations,58 enhancing the applicability of the predictive models.
This study also had several limitations. First, although our models based on clinical characteristics showed good calibration, their discrimination was not excellent (ie, C-index >0.75). We expect that future predictive models can be optimised by incorporating multidimensional data such as molecular biomarkers and radiomics. Second, the proportionality of hazard was confirmed in the Cox model for pCR patients, but not the case for the non-pCR model. Such result might lead to developing a non-proportional hazards model (ie, adding an interaction between covariates and time) for non-pCR patients. But to maintain the consistency of methodology and to simplify the utility of the predictive model, the proportional Cox model was retained for non-pCR patients. Moreover, our final models included the readmission within 30 days after discharge as a predictor. The condition of being alive 30 days after discharge might lead to a selection bias towards survivors, although the proportion of patients censored before or died within 30 days after discharge was relatively low (1.35%, 6 of 444 patients in the external validation cohort).
Summary
In conclusion, our study demonstrated that patients with oesophageal cancer who achieved pCR after neoadjuvant therapy and oesophagectomy had significantly better OS than that of non-pCR patients. To facilitate personalised medical treatment decisions, we have developed and validated distinct predictive models for these two populations. The nomograms we constructed are web-based and dynamically calculate the OS probability for pCR and non-pCR patients, enabling clinicians and patients to decide on individualised medical treatments.
Data availability statement
Data may be obtained from a third party and are not publicly available. The data will not be made available publicly.
Ethics statements
Patient consent for publication
Ethics approval
This study involves human participants and this study was approved by the Institutional Review Board (IRB) of Shanghai Ruijin Hospital (KY2023-140) on 21 June 2023. The informed consent was waived due to the retrospective characteristic of this study.
Acknowledgments
We thank Ming Du, MD (The First Affiliated Hospital of Chongqing Medical University), Changhong Lian, MD (Heping Hospital Affiliated to Changzhi Medical College), Qiang Zhao, MD (Heping Hospital Affiliated to Changzhi Medical College), Hongjing Jiang, MD (Tianjin Medical University Cancer Institute and Hospital), Lei Gong, MD (Tianjin Medical University Cancer Institute and Hospital), Zhigang Li, MD (Shanghai Chest Hospital, Shanghai Jiao Tong University), Jun Liu, MD (Shanghai Chest Hospital, Shanghai Jiao Tong University), Deyao Xie, MD (The First Affiliated Hospital of Wenzhou Medical University), Wenfeng Li, MD (The First Affiliated Hospital of Wenzhou Medical University), Chun Chen, MD (Fujian Medical University Union Hospital), Bin Zheng, MD (Fujian Medical University Union Hospital), Keneng Chen, MD (Peking University Cancer Hospital and Institute), Liang Dai, MD (Peking University Cancer Hospital and Institute), Yongde Liao, MD (Union Hospital, Tongji Medical College, Huazhong University of Science and Technology) and Kuo Li, MD (Union Hospital, Tongji Medical College, Huazhong University of Science and Technology) for their contributions to the CMISG1701 trial (NCT03001596) which comprised part of the external validation data.
References
Supplementary materials
Supplementary Data
This web only file has been produced by the BMJ Publishing Group from an electronic file supplied by the author(s) and has not been edited for content.
Footnotes
YC, BH and HT are joint first authors.
YC, BH and HT contributed equally.
Contributors YZ is responsible for the overall content as guarantor. YZ, HL, JZ(JieZ) and LT had full access to all data in the study and take responsibility for the integrity of the data and accuracy of the data analysis. YC, BH and HT contributed equally to this work and are first coauthors. Concept and design: YC, YZ, BH, HT, LT, JZ(JieZ) and HL. Acquisition, analysis or interpretation of data: YC, BH, HT, DD, TS, XC, XF, JZ(JiahaoZ), LS, CL and HJ. Drafting of the manuscript: YC, YZ, BH and HT. Critical revision of the manuscript for important intellectual content: YZ, LT, JZ(JieZ) and HL. Statistical analysis: YC, YZ, BH and HT. Obtained funding: YZ and HL. Administrative, technical or material support: YZ, JH, LT and JZ(JieZ).
Funding This study was supported by the National Natural Science Foundation of China (82072557), National Key Research and Development Programme of China (2021YFC2500900), Shanghai Municipal Education Commission - Gaofeng Clinical Medicine Grant (20172005, the 2nd round of disbursement), programme of Shanghai Academic Research Leader from Science and Technology Commission of Shanghai Municipality (20XD1402300), Novel Interdisciplinary Research Project from Shanghai Municipal Health Commission (2022JC023), Interdisciplinary Programme of Shanghai Jiao Tong University (YG2023ZD04) and Clinical Research Project in Health Services of Shanghai Municipal Health Commission (202240089).
Competing interests None declared.
Provenance and peer review Not commissioned; externally peer reviewed.
Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.