Predictive modelling of duodenal stump leakage after gastric cancer and long-term oncological outcomes

Introduction

As the fifth most common cancer worldwide, gastric cancer accounts for the fourth leading cause of death among cancer types (1). Despite significant advancements in multimodal therapies, including radiotherapy, chemotherapy, and immunotherapy, the overall survival (OS) rates for gastric cancer remain unsatisfactory (2,3). Currently, the comprehensive treatment of gastric cancer is still centered on radical gastric cancer surgery, but the poor prognosis brought by postoperative complications of gastric cancer should not be neglected (4,5).

Duodenal stump leakage (DSL), as one of the most serious complications after gastric cancer surgery, has a low incidence (from 1.6% to 5%) but still has a high mortality rate (from 7% to 67%), and the spontaneous closure rate is reported to be 28–92% (6,7). It is also prone to other associated complications such as intra-abdominal abscesses, sepsis, and diffuse peritonitis (8). Although DSL poses significant postoperative challenges, there is no predictive model for DSL after radical gastric cancer treatment reported, and it remains unclear whether DSL impacts the survival prognosis of gastric cancer patients.

Therefore, it is essential to develop a predictive model for DSL and to observe the impact of DSL on oncologic outcomes in gastric cancer. In this study, we collected clinical characteristics from 2,511 patients who underwent radical gastric cancer surgery and followed them up. Subsequently, we constructed a series of predictive models to evaluate the risk of DSL in patients, with the objective of identifying effective predictive model to inform clinical decision-making. We present this article in accordance with the TRIPOD reporting checklist (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0047/rc).

Methods

Study population

Patients who underwent radical gastric cancer surgery at The First Affiliated Hospital of Nanchang University from January 2016 to September 2019 were retrospectively collected. The surgical approach was based on the guideline for the diagnosis and treatment of gastric cancer. All surgeries were performed by experienced senior attending surgeons, each with over 100 independent radical gastrectomies and at least 5 years of specialized gastric cancer practice. All patients underwent radical distal gastrectomy (open or laparoscopic-assisted) with D1+ or D2–3 lymphadenectomy, and the duodenal stump was routinely reinforced with embedded sutures. Neoadjuvant chemotherapy [typically (SOX) or (XELOX) for 2–4 cycles] was given for locally advanced gastric cancer (cT3–4 or N+), with completion defined as ≥80% of planned cycles. Preoperative gastric outlet obstruction was defined as vomiting, endoscopic narrowing at the antrum/duodenum, or imaging showing gastric retention. The study population flow chart is illustrated in Figure 1. The inclusion criteria are as follows: (I) patients with primary gastric adenocarcinoma; (II) patients with M1 metastasis or R1/R2 resection were excluded to avoid confounding from non-curative surgery; (III) patients underwent gastrectomy with duodenal stump closure, using Billroth II or Roux-en-Y reconstruction. Billroth I cases were excluded because no duodenal stump is created; (IV) complete clinical and follow-up data. The diagnostic criteria for DSL were as follows: (I) early postoperative patients presenting with fever and right lower pressure pain and rebound pain; (II) bile-like fluid draining from abdominal drainage tubes; (III) confirmed by sinusography, abdominal ultrasound or CT examination; (IV) confirmed by re-operation exploration (9-11). The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This retrospective study was approved by the Ethics Committee of The First Affiliated Hospital of Nanchang University (IIT [2024] Clinical Ethics Review No. 497). All patients had previously signed the general admission consent form authorizing the anonymized use of their clinical data for future research purposes. All patient data were anonymized before analysis.

Figure 1 Flow chart of the study design. DCA, decision curve analysis; LASSO, least absolute shrinkage and selection operator; NPV, negative predictive value; PPV, positive predictive value; ROC, receiver operating characteristic.

Predictive model construction

Predictive models are a common tool in gastric cancer research, yet the predictive efficacy of different models varies. Here we selected six prediction models for comparison. Here we selected six prediction models for comparison: (I) the logistic regression (LR) diagnostic model was developed in the training cohort using stepwise regression with the ‘rms’ package in R, and the Akaike information criterion was used to select the optimal model; (II) the ‘glmnet’ package in R was utilized to implement the least absolute shrinkage and selection operator (LASSO), and the lambda value with the lowest mean square error was chosen; (III) the support vector machine (SVM) model was optimized for cost and gamma values using the ‘e1071’ package in R; (IV) the ‘knn’ package in R was utilized to develop the optimal K-nearest neighbors (KNN) model by varying the k-value from 0 to 100; (V) using the ‘rpart’ package in R, an initial decision tree (DT) was constructed with all independent variables from the training cohort, followed by the development of the optimal DT; (VI) random forest (RF) consists of multiple DTs that are weak classifiers on their own but collectively provide strong predictions, and hyper-parameter optimization was used to select the optimal model.

Long-term outcomes definition

Patients received standard oncologic follow-up according to gastric cancer guidelines. Patients received quarterly follow-up during the first 2 years, semiannual follow-up from 2 to 5 years, and annual follow-up after 5 years. Disease recurrence was assessed clinically and confirmed by endoscopic or radiologic methods.

Statistical analysis

Data analysis was conducted using SPSS and R software. Frequencies were used to express count data, and comparisons were made using the Chi-squared test. Data following a normal distribution were presented as mean ± standard deviation (SD) and analyzed using the t-test. Non-parametric tests were applied to data lacking normal distribution or exhibiting variance differences. A one-way analysis of variance (ANOVA) was conducted to compare multiple groups. A P value below 0.05 was deemed significant.

Results

Clinical characteristics of patients with NDSL and DSL

As illustrated in Figure 1, of the 2,511 patients who met the eligibility criteria, 63 cases of DSL were identified, representing an incidence rate of 2.51%. Baseline characteristics of the NDSL and DSL groups showed no significant differences (P>0.05, Table S1). However, a difference was noted in the perioperative portion of the partial characteristics component, which suggests that these portion of the characteristics may be a risk factor for DSL (Table S2). The 2,511 patients were randomly divided in a 7:3 ratio into a training cohort of 1,759 patients, including 45 DSLs, and a test cohort of 752 patients, including 18 DSLs. The training and testing cohorts showed no statistically significant differences in baseline data and preoperative characteristics (Tables S1,S2), ensuring the reliability of the test and validation cohort results.

LR-based predictive model

The univariate analysis, conducted using LR modelling, demonstrated that the following clinical characteristics were identified as risk factors for DSL: preoperative albumin, preoperative hemoglobin, American Society of Anesthesiologists-Physical Status (ASA-PS), reconstruction methods, operation time, intraoperative blood loss, tumor size, T4, total positive lymph nodes, N3, 4th postoperative-day C-reactive protein (CRP), and the distal margin. The multivariate analysis identified preoperative albumin, preoperative hemoglobin, operation time, intraoperative blood loss, 4th postoperative-day CRP, and the distal margin as independent risk factors for DSL (Table 1). A nomogram was constructed utilizing the identified independent risk factors to estimate the probability of DSL occurrence (Figure 2A). The model demonstrated area under the curve (AUC) values of 0.980 in the training cohort and 0.954 in the test cohort, indicating high discriminatory power (Figure 2B,2E). Additionally, calibration curves for the model exhibited excellent alignment across both cohorts (Figure 2C,2F), and decision curve analysis (DCA) demonstrated a significant net benefit (Figure 2D,2G).

Figure 2 The assessment of LR-based predictive model. (A) Nomogram to predict the occurrence of DSL. (B) ROC curve of nomogram in the training cohort. (C) Calibration curve of nomogram in the training cohort. (D) Decision curve analysis curve of training cohort. (E) ROC curve of nomogram in the test cohort. (F) Calibration curve of nomogram in the test cohort. (G) Decision curve analysis curve of test cohort. AUC, area under the curve; CRP, C-reactive protein; DSL, duodenal stump leakage; LR, logistic regression; ROC, receiver operating characteristic.

LASSO-based predictive model

LASSO regression was conducted on the clinical characteristics of the training cohort with the objective of identifying the most pertinent clinical characteristics. The clinical characteristics identified using the optimal lambda value include preoperative albumin, preoperative hemoglobin, CRP levels on the 1st and 4th postoperative days, ASA-PS, operation time, intraoperative blood loss, tumor size, degree of differentiation, and the distal margin (Figure 3A). Furthermore, the model demonstrated robust predictive capacity, as evidenced by AUC value of 0.977 in the training cohort and 0.945 in the test cohort (Figure 3B,3C).

Figure 3 The assessment of LASSO-based, SVM-based and KNN-based predictive models. (A) The LASSO model finds the best parameters and shows how deviance and log parameters relate. (B) ROC curve of LASSO model in training cohort. (C) ROC curve of LASSO model in the test cohort. (D) The SVM model with eight features had the best accuracy. (E) ROC curve of SVM model in the training cohort. (F) ROC curve of SVM model in the test cohort. (G) The KNN model was built with the best kernel and k-value. (H) ROC curve of KNN model in the training cohort. (I) ROC curve of KNN model in the test cohort. AUC, area under the curve; KNN, K-nearest neighbors; LASSO, least absolute shrinkage and selection operator; ROC, receiver operating characteristic; SVM, support vector machine.

SVM-based predictive model

The SVM prediction model achieved optimal accuracy with 8 vectors, as shown in Figure 3D. These features were the distal margin, 4th post-operative day CRP, preoperative albumin, tumor size, preoperative hemoglobin, operation time, T category, and reconstruction methods. Surprisingly, the AUC value of the model reaches 1 in the training cohort (Figure 3E), which is relatively low at 0.895 in the testing cohort (Figure 3F). However, this further demonstrates that the SVM model exhibits excellent predictive efficacy.

KNN-based predictive model

The hyperparameter optimization of the KNN function demonstrated that the triangular kernel was the most effective for the prediction model, with an optimal K value of 14 (Figure 3G). The construction of the prediction model with these features revealed that the AUC was relatively low at 0.885 in the training cohort (Figure 3H) and even lower at 0.719 in the validation cohort (Figure 3I) . This suggests that the KNN model is not an effective predictor of the DSL.

DT-based predictive model

The DT predictive model was employed to ascertain the relative importance of various variables. The top eight variables, as determined by the model, were as follows: the distal margin, 4th post-operative-day CRP, preoperative albumin, body mass index (BMI), preoperative obstruction, intraoperative blood loss, operation time, and 1st post-operative-day CRP (Figure 4A). In the DT model, the number of split nodes was determined to be 4 (Figure 4B). Consequently, four key features were selected for modelling the DT prediction: the distal margin ≥1.8 cm; 4th post-operative-day CRP <148 mg/L; the distal margin ≥3.8 cm and the preoperative albumin ≥30 g/L (Figure 4C). It was thus determined that the distal margin plays a pivotal role in the DT model. However, the AUC of the predictive model was not as high as 0.863 in the training cohort (Figure 4D) and even lower at 0.724 in the test cohort (Figure 4E), indicating that the DT model may not be the optimal model for distinguishing between DSL and NDSL.

Figure 4 The assessment of DT-based and RF-based predictive models. (A) The importance of clinical features in the construction of DT model. (B) The tree size evaluates the relative errors when the cp is set to an optimal value. (C) The DT model was constructed with the selected clinical indicators as inputs. (D) ROC curve of DT model in the training cohort. (E) ROC curve of DT model in the test cohort. (F) Error variation for randomly selected predictor variables for the random forest classifier. (G) Clinical features improve precision and reduce the Gini coefficient. (H) ROC curve of RF model in the training cohort. (I) ROC curve of RF model in the test cohort. ASA-PS, American Society of Anesthesiologists-Physical Status; AUC, area under the curve; BMI, body mass index; cp, complexity parameter; CRP, C-reactive protein; DT, decision tree; M, metastasis; N, node; RF, random forest; ROC, receiver operating characteristic; T, tumor.

RF-based predictive model

In the RF predictive model, the optimal number of tree-splitting nodes was identified, and it was observed that the error reached a minimum value of 0.167 when the number of splitting nodes reached 25 (Figure 4F). Furthermore, a feature importance analysis was conducted, which demonstrated that among the 25 selected features, the distal margin, 4th post-operative-day CRP and preoperative albumin were particularly crucial for prediction accuracy and played a similarly pivotal role in reducing the Gini coefficient of the prediction model (Figure 4G). It is noteworthy that the AUC of the RF prediction model reached a value of 1.000 in the training cohort (Figure 4H) and was also high at 0.944 in the test cohort (Figure 4I). This indicates that the RF model exhibits the superior predictive efficacy among the six models previously described.

Comparison of predictive abilities of various models

Subsequently, the predictive efficacy of the aforementioned six models was assessed, with the findings presented in Table S3. It was observed that the SVM and RF models exhibited superior accuracy, AUC, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) compared to the LR, DT, LASSO, and KNN models within the training cohort. It is noteworthy that the RF prediction model exhibited optimal performance, with a perfect 1 for accuracy, AUC, sensitivity, specificity, PPV, and NPV (Figure 5A). In the test cohort, the PPV of all predictive models was relatively low, but the accuracy, AUC, sensitivity, specificity, and NPV exhibited minimal variation, suggesting that these models were also relatively stable (Figure 5B). The analysis revealed an intriguing observation: the traditional LR model exhibited significantly greater stability, particularly in terms of PPV, compared to machine learning algorithms. This suggests that, although the traditional LR model may not achieve optimal performance in the training set, it demonstrates notable superiority in the test set.

Figure 5 Evaluation of predictive models and long-term outcomes. (A) Comprehensive assessment of each model in the training cohort. (B) Comprehensive assessment of each model in the test cohort. (C) The OS for the DSL and NDSL cohort. (D) The RFS for the DSL and NDSL cohort. AUC, area under the curve; DSL, duodenal stump leakage; DT, decision tree; KNN, K-nearest neighbors; LASSO, least absolute shrinkage and selection operator; NDSL, non-duodenal stump leakage; NPV, negative predictive value; OS, overall survival; PPV, positive predictive value; RF, random forest; RFS, recurrence-free survival; SVM, support vector machine.

Long-term oncologic outcomes of non-duodenal stump leakage (NDSL) and DSL

The median survival time was 41 months in the NDSL group and 38 months in the DSL group. The OS was not significantly different between the two groups (P=0.13, Figure 5C). However, there was a difference in recurrence-free survival (RFS) between the two groups (P=0.03, Figure 5D). These findings suggest that DSL does not affect long-term survival but may affect tumor recurrence. Recurrence occurred in 44.44% of the DSL group versus 27.74% of the non-DSL group (P=0.004), with peritoneal dissemination being significantly more frequent in the DSL group (71.43% vs. 48.01%, P=0.02), whereas rates of hematogenous metastasis, locoregional recurrence, and distant lymph node metastasis were comparable. With respect to mortality, deaths directly attributable to DSL occurred in only 10 patients, while the majority of deaths in both groups were due to gastric cancer progression.

Discussion

DSL represents a significant postoperative complication following radical gastric cancer surgery (12,13). It poses a considerable threat to the life and health of patients and presents a significant challenge to the management of clinicians (14). In light of the aforementioned considerations, our study represents a pioneering effort in the construction of multiple prediction models, with the objective of identifying the optimal model for predicting DSL. The predictive models may be utilized for the early identification or preventive intervention of DSL, thereby reducing the incidence or severity of this complication and conferring benefits to patient health.

Risk factors

A comprehensive analysis of all predictive models revealed four primary categories of risk factors associated with DSL. The analysis identified the following key findings: (I) among the baseline factors, preoperative hemoglobin and albumin had the greatest influence; (II) among the tumor factors, the distal margin had the greatest influence; (III) among the surgical factors, intraoperative blood loss and operation time had the greatest influence; (IV) among the infectious factors, the 4th postoperative-day CRP had the greatest influence. It should be emphasized, however, that the observed association between 4th postoperative-day CRP and DSL is intended to be interpreted as a diagnostic and prognostic association rather than a causal risk factor. Among all the prediction models, preoperative albumin, the distal margin, and the 4th postoperative-day CRP were the most important, suggesting that these three factors seriously affect the occurrence of DSL. This indicates that clinicians must prioritize the identification and management of these risk factors to prevent DSL. Among them, preoperative albumin and CRP have been reported in other studies (14-16), while the effect of the distal margin on DSL is the first time reported to the best of our knowledge. The distal margin distance contributes to the occurrence of DSL for two possible reasons: firstly, too short a margin severely affects the reinforced anastomosis of the duodenal stump (17,18); secondly, too close proximity of the tumor to the duodenal stump causes inflammatory edema of the surrounding tissues that affects the healing of the duodenal stump (19). In fact, a short duodenal stump resulting from a necessary oncological distal margin can hinder safe stump closure and increase the risk of DSL. Although Roux-en-Y reconstruction reduces this risk, it does not eliminate it. In cases where the duodenal stump is particularly short or difficult to close securely, an alternative technique—direct duodenojejunal anastomosis—may be considered, as it relies on an anastomosis rather than a suture closure, potentially offering greater safety (20).

Models evaluation

The results indicate that the overall efficacy of SVM and RF prediction models is superior to that of LR, LASSO, DT and KNN in the training cohort. However, the improvement in PPV is not as pronounced in the test cohort as it is in the training cohort. The observed instability in machine learning models may stem from their heightened sensitivity to variations in data distribution, the presence of outliers, and other extraneous factors, which can lead to substantial variability in predictive performance across different datasets (21,22). In contrast, traditional LR models appear to offer greater stability, particularly concerning PPV. This stability may stem from the relatively straightforward structure of LR models and their clear-cut assumptions regarding the underlying data (23,24). Therefore, in practical applications where models are deployed in dynamic and complex clinical endeavour, the focus should extend beyond mere accuracy to include considerations of stability, interpretability, and performance on specific metrics relevant to individual circumstances of patients (25,26). Looking forward, a promising avenue for subsequent research and practice is to further explore the inherent stability of LR models. Additionally, investigating strategies to combine these stability advantages with the performance strengths of machine learning models could lead to the development of more powerful and stable predictive models. Such models would be better poised to handle various complex clinical endeavours.

Long-term prognosis

Our findings revealed a novel observation: patients with DSL and NDSL exhibited no significant difference in OS outcomes. However, there was a discernible divergence in RFS. This pattern suggests that the adverse effect of DSL on RFS is primarily driven by an increased risk of peritoneal dissemination, whereas OS remains comparable due to effective salvage therapies and the limited number of DSL-related fatalities. Numerous studies over the past decades have explored the potential link between gastrointestinal leakage and long-term prognosis following gastrointestinal tumor surgery (27,28). However, the results remain controversial due to the presence of a number of confounding factors that need to be considered in these patients (29,30). This study revealed a disparity in tumor RFS between the DSL and NDSL groups. This result may be attributed to the imbalance in the number of patients in the two groups (31,32), as these differences may also be responsible for the higher rate of lymph node metastasis and larger tumor size observed in the DSL group, which may affect the RFS of the patients (33,34).

Limitations

While the predictive model yielded promising results, it is important to acknowledge the inherent limitations of the study. Firstly, this retrospective study is inherently susceptible to confounding bias, including the inability to assess comorbidity severity; future work should utilize objective measures such as the Charlson Comorbidity Index to clarify this relationship (35). Secondly, the study was conducted at a single center, which may have resulted in a smaller number of cases of duodenal stump fistula being included due to the low prevalence of the disease. To enhance the accuracy and statistical efficacy of the risk prediction model, external validation of a multicenter, large-sample database is required (36,37).

Clinical implications and actionable interventions

Based on our predictive model, several modifiable factors across the perioperative period offer actionable targets to reduce DSL risk and improve outcomes. Preoperatively, nutritional support should correct hypoalbuminemia and low hemoglobin. Intraoperatively, surgeons should secure an adequate distal margin (e.g., ≥1.8 cm), avoid excessive devascularization, and minimize operation time and blood loss. Postoperatively, elevated CRP on day 4 warrants early imaging [computed tomography (CT) or sinography] for prompt DSL diagnosis. Finally, as DSL impairs recurrence-free survival, affected patients need intensified surveillance (e.g., more frequent imaging or tumor markers) for early recurrence detection.

Conclusions

A series of preliminary risk predictive models for DSL have been developed based on data from our center. These models are expected to aid clinicians in improving early diagnosis and optimizing treatment strategies for DSL. Among the models developed, the RF predictive model demonstrated the most optimal predictive efficacy. Additionally, the data highlight DSL’s impact on RFS, underscoring its role in oncological outcomes.

Acknowledgments

The authors greatly thank Yi Tu and Yi Cao for their support. The authors would like to thank Zongfeng Feng and Qingwen Zeng for their valuable help.

Footnote

Reporting Checklist: The authors have completed the TRIPOD reporting checklist. Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0047/rc

Data Sharing Statement: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0047/dss

Peer Review File: Available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0047/prf

Funding: This study was supported by a grant from the National Natural Science Foundation of China (No. 82103165); Jiangxi Provincial Health Technology Project (Project Grant No. 202310020); The First Affiliated Hospital of Nanchang University Young Talent Research and Cultivation Project (Project Grant No. YFYPY202272); the Science and Technology Project of Jiangxi Health Commission (No. 202410191); the Science and Technology Project of Jiangxi Provincial Department of Education (No. GJJ200225); and the “Talent 555 Project” of Jiangxi Province, People’s Republic of China.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://tcr.amegroups.com/article/view/10.21037/tcr-2026-1-0047/coif). All authors report that this study was supported by a grant from the National Natural Science Foundation of China (No. 82103165); Jiangxi Provincial Health Technology Project (Project Grant No. 202310020); The First Affiliated Hospital of Nanchang University Young Talent Research and Cultivation Project (Project Grant No. YFYPY202272); the Science and Technology Project of Jiangxi Health Commission (No. 202410191); the Science and Technology Project of Jiangxi Provincial Department of Education (No. GJJ200225) and the “Talent 555 Project” of Jiangxi Province, People’s Republic of China. The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This retrospective study was approved by the Ethics Committee of The First Affiliated Hospital of Nanchang University (IIT [2024] Clinical Ethics Review No. 497). All patients had previously signed the general admission consent form authorizing the anonymized use of their clinical data for future research purposes. All patient data were anonymized before analysis.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229-63. [Crossref] [PubMed]
2. Zhan T, Betge J, Schulte N, et al. Digestive cancers: mechanisms, therapeutics and management. Signal Transduct Target Ther 2025;10:24. [Crossref] [PubMed]
3. Luo D, Liu Y, Lu Z, et al. Targeted therapy and immunotherapy for gastric cancer: rational strategies, novel advancements, challenges, and future perspectives. Mol Med 2025;31:52. [Crossref] [PubMed]
4. Ramos MFKP, Pereira MA, Dias AR, et al. Surgical treatment of gastric cancer: A single center's 15-year experience. Clinics (Sao Paulo) 2025;80:100828. [Crossref] [PubMed]
5. Patel AK, Sethi NS, Park H. Gastric Cancer: A Review. JAMA 2026;335:439-50. [Crossref] [PubMed]
6. Zizzo M, Ugoletti L, Manzini L, et al. Management of duodenal stump fistula after gastrectomy for malignant disease: a systematic review of the literature. BMC Surg 2019;19:55. [Crossref] [PubMed]
7. Paik HJ, Lee SH, Choi CI, et al. Duodenal stump fistula after gastrectomy for gastric cancer: risk factors, prevention, and management. Ann Surg Treat Res 2016;90:157-63. [Crossref] [PubMed]
8. Po Chu Patricia Y, Ka Fai Kevin W, Fong Yee L, et al. Duodenal stump leakage. Lessons to learn from a large-scale 15-year cohort study. Am J Surg 2020;220:976-81.
9. Liu HB. Diagnosis and treatment of duodenal stump leakage after laparoscopic gastrectomy. World Chinese Journal of Digestology 2017;25:399.
10. Liu J, Zhou S, Wang S, et al. Analysis of risk factors for duodenal leak after repair of a duodenal perforation. BMC Surg 2023;23:116. [Crossref] [PubMed]
11. Wu H, Liang M, Liao E, et al. Surgical replays-assisted diagnosis of duodenal stump fistula after laparoscopic radical distal gastrectomy (with video). Asian J Surg 2022;45:1350-1. [Crossref] [PubMed]
12. Ramos MFKP, Pereira MA, Barchi LC, et al. Duodenal fistula: The most lethal surgical complication in a case series of radical gastrectomy. Int J Surg 2018;53:366-70. [Crossref] [PubMed]
13. He H, Li H, Ye B, et al. Single Purse-String Suture for Reinforcement of Duodenal Stump During Laparoscopic Radical Gastrectomy for Gastric Cancer. Front Oncol 2019;9:1020. [Crossref] [PubMed]
14. Gu L, Zhang K, Shen Z, et al. Risk Factors for Duodenal Stump Leakage after Laparoscopic Gastrectomy for Gastric Cancer. J Gastric Cancer 2020;20:81-94. [Crossref] [PubMed]
15. Cozzaglio L, Coladonato M, Biffi R, et al. Duodenal fistula after elective gastrectomy for malignant disease: an italian retrospective multicenter study. J Gastrointest Surg 2010;14:805-11. [Crossref] [PubMed]
16. Orsenigo E, Bissolati M, Socci C, et al. Duodenal stump fistula after gastric surgery for malignancies: a retrospective analysis of risk factors in a single centre experience. Gastric Cancer 2014;17:733-44. [Crossref] [PubMed]
17. Sano A, Imai Y, Yamaguchi T, et al. Importance of duodenal stump reinforcement to prevent stump leakage after gastrectomy: a large-scale multicenter retrospective study (KSCC DELICATE study). Gastric Cancer 2024;27:1320-30. [Crossref] [PubMed]
18. Sun L, Wang W, Zhou J, et al. Modified Q-type purse-string suture duodenal stump embedding method for laparoscopic gastrectomy for gastric cancer. BMC Surg 2024;24:123. [Crossref] [PubMed]
19. Cook WH, Gillespie CS, Bakhsh A, et al. Epileptogenesis in meningioma: Theories, putative biomarkers, and postoperative risk. Epilepsia 2025;66:4079-90. [Crossref] [PubMed]
20. Manenti A, Melegari G, Farinetti A. A direct duodeno-jejunal anastomosis: A solution for a difficult duodenal stump. J Visc Surg 2017;154:469-70. [Crossref] [PubMed]
21. Wu A, Luo L, Zeng Q, et al. Comparative assessment of the capability of machine learning-based radiomic models for predicting omental metastasis in locally advanced gastric cancer. Sci Rep 2024;14:16208. [Crossref] [PubMed]
22. Zhang P, Sun W, Wei D, et al. PDA-PRGCN: identification of Piwi-interacting RNA-disease associations through subgraph projection and residual scaling-based feature augmentation. BMC Bioinformatics 2023;24:18. [Crossref] [PubMed]
23. Grant SW, Hickey GL, Head SJ. Statistical primer: multivariable regression considerations and pitfalls. Eur J Cardiothorac Surg 2019;55:179-85. [Crossref] [PubMed]
24. Varady NH, Pareek A, Eckhardt CM, et al. Multivariable regression: understanding one of medicine's most fundamental statistical tools. Knee Surg Sports Traumatol Arthrosc 2023;31:7-11. [Crossref] [PubMed]
25. Bao QR, Pellino G, Spolverato G, et al. The impact of anastomotic leak on long-term oncological outcomes after low anterior resection for mid-low rectal cancer: extended follow-up of a randomised controlled trial. Int J Colorectal Dis 2022;37:1689-98. [Crossref] [PubMed]
26. Theodorou B, Danek B, Tummala V, et al. Improving medical machine learning models with generative balancing for equity and excellence. NPJ Digit Med 2025;8:100. [Crossref] [PubMed]
27. Kamarajah SK, Navidi M, Wahed S, et al. Anastomotic Leak Does Not Impact on Long-Term Outcomes in Esophageal Cancer Patients. Ann Surg Oncol 2020;27:2414-24. [Crossref] [PubMed]
28. Andreou A, Biebl M, Dadras M, et al. Anastomotic leak predicts diminished long-term survival after resection for gastric and esophageal cancer. Surgery 2016;160:191-203. [Crossref] [PubMed]
29. Aiolfi A, Griffiths EA, Sozzi A, et al. Effect of Anastomotic Leak on Long-Term Survival After Esophagectomy: Multivariate Meta-analysis and Restricted Mean Survival Times Examination. Ann Surg Oncol 2023;30:5564-72. [Crossref] [PubMed]
30. Aiolfi A, Bona D, Cali M, et al. Impact of Thoracic Duct Resection on Long-Term Survival After Esophagectomy: Individual Patient Data Meta-analysis. Ann Surg Oncol 2024;31:6699-709. [Crossref] [PubMed]
31. Li Y, Hao J, Hu Z, et al. Current status of clinical trials assessing mesenchymal stem cell therapy for graft versus host disease: a systematic review. Stem Cell Res Ther 2022;13:93. [Crossref] [PubMed]
32. Woelber L, Hampl M, Eulenburg CZ, et al. Risk for Pelvic Metastasis and Role of Pelvic Lymphadenectomy in Node-Positive Vulvar Cancer-Results from the AGO-VOP.2 QS Vulva Study. Cancers (Basel) 2022;14:418. [Crossref] [PubMed]
33. Chen L, Yang F, Qi Z, et al. Predicting lymph node metastasis and recurrence in patients with early stage colorectal cancer. Front Med (Lausanne) 2022;9:991785. [Crossref] [PubMed]
34. Tu J, Shao S, Qin J. The number of mesogastria containing metastatic lymph nodes predicts gastric cancer prognosis. Surgery 2024;176:739-47. [Crossref] [PubMed]
35. Sun Y, Li Z, Tian Y, et al. Development and validation of nomograms for predicting overall survival and cancer-specific survival in elderly patients with locally advanced gastric cancer: a population-based study. BMC Gastroenterol 2023;23:117. [Crossref] [PubMed]
36. Xia M, Jin C, Pei B, et al. Development and validation of a nomogram for difficult laryngoscopy at visual laryngoscopy: a prospective nested case-control study. Quant Imaging Med Surg 2023;13:4663-75. [Crossref] [PubMed]
37. Pérez-Sáez MJ, Redondo-Pachón D, Arias-Cabrales CE, et al. Outcomes of Frail Patients While Waiting for Kidney Transplantation: Differences between Physical Frailty Phenotype and FRAIL Scale. J Clin Med 2022;11:672. [Crossref] [PubMed]