Predictive value of the preoperative controlling nutritional status score for anastomotic leakage after radical gastrectomy for gastric cancer and risk stratification based on predicted probability Authors: Ruiyun Chen, Jianqi Li, Shuqin Guan Published: 2026 ABSTRACT BackgroundAnastomotic leakage remains one of the major complications after gastrectomy in patients with gastric cancer and may increase postoperative morbidity and delay recovery. Preoperative nutritional status is closely associated with anastomotic healing; however, the value of the Controlling Nutritional Status (CONUT) score for predicting clinically relevant anastomotic leakage after gastric cancer surgery requires further clarification. This study aimed to evaluate the predictive value of the preoperative CONUT score for clinically relevant anastomotic leakage after radical gastrectomy and to construct an interpretable risk prediction model and a three-tier risk stratification system based on predicted probability.MethodsClinical data from 283 patients who underwent radical gastrectomy from January 2021 to January 2025 were retrospectively analyzed. The CONUT score was calculated from serum albumin, total lymphocyte count, and total cholesterol. The primary endpoint was clinically relevant anastomotic leakage of grade II or higher within 30 postoperative days. Logistic regression was used to identify risk factors. Receiver operating characteristic (ROC) curve analysis was used to evaluate the predictive value of the CONUT score and its components. Internal validation was performed using bootstrap resampling with 1000 repetitions, and exploratory temporal validation was conducted by using patients treated from January 2021 to December 2023 as the modeling cohort and those treated from January 2024 to January 2025 as the temporal validation cohort.ResultsAmong the 283 patients, 39 (13.8%) developed clinically relevant anastomotic leakage. Multivariable Logistic regression showed that CONUT score (odds ratio [OR] = 1.425, 95% confidence interval [CI]: 1.168–1.739, P < 0.001), diabetes mellitus (OR = 2.486, 95% CI: 1.052–5.876, P = 0.038), total gastrectomy (OR = 2.318, 95% CI: 1.028–5.228, P = 0.043), operative time (OR = 1.008, 95% CI: 1.002–1.014, P = 0.012), and intraoperative blood loss (OR = 1.004, 95% CI: 1.001–1.007, P = 0.018) were independent risk factors for anastomotic leakage. The area under the curve (AUC) of the CONUT score for predicting anastomotic leakage was 0.782 (95% CI: 0.716–0.848). When a CONUT score ≥ 3 was used as a sensitivity-prioritized screening threshold, the sensitivity, specificity, and negative predictive value were 74.4%, 62.3%, and 93.8%, respectively; when a score ≥ 4 was used as a higher-specificity alternative threshold, the sensitivity and specificity were 59.0% and 78.3%, respectively. Compared with serum albumin alone, the AUC of the CONUT score showed a modest increase of 0.058 (DeLong test P = 0.042). The risk prediction model based on five independent risk factors had an apparent concordance index (C-index) of 0.838 (95% CI: 0.779–0.897), a bootstrap-corrected C-index of 0.826, and a C-index of 0.803 (95% CI: 0.701–0.905) in the temporal validation cohort. After three-tier risk stratification based on predicted probability, the incidence of anastomotic leakage increased stepwise in the low-, intermediate-, and high-risk groups (χ² = 56.925, P < 0.001).ConclusionsThe preoperative CONUT score was associated with an increased risk of clinically relevant anastomotic leakage after gastric cancer surgery. A risk prediction model incorporating CONUT score, diabetes mellitus, total gastrectomy, operative time, and intraoperative blood loss may provide a reference for perioperative risk assessment, although multicenter external validation is still required before broader application.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12893-026-03938-5. Introduction Gastric cancer remains one of the major gastrointestinal malignancies worldwide, with a particularly substantial disease burden in East Asia. Radical gastrectomy combined with standardized lymphadenectomy and gastrointestinal reconstruction remains an important treatment for resectable gastric cancer, and the quality of perioperative management directly affects short-term complications and long-term prognosis [1, 2]. Anastomotic leakage is one of the most serious complications after gastrectomy and may lead to intra-abdominal infection, sepsis, reintervention, prolonged hospitalization, and impaired long-term survival. Previous studies of gastric cancer surgery have suggested that anastomotic leakage is associated with patient baseline status, tumor stage, extent of resection, and operative factors [3, 4]. Preoperative nutritional status is closely related to tissue repair, collagen synthesis, immune function, and infection control. Perioperative nutrition guidelines emphasize that nutritional risk screening and nutritional therapy should be incorporated into the overall management pathway for patients undergoing oncologic surgery, particularly in those with malnutrition, reduced intake, or metabolic stress [5, 6]. The Controlling Nutritional Status (CONUT) score is calculated from serum albumin, total lymphocyte count, and total cholesterol and can be obtained from routine preoperative laboratory tests. It is objective, simple, and reproducible. In recent years, the value of the CONUT score and related nutritional indicators in postoperative complications and prognostic assessment of gastric cancer has attracted increasing attention, but evidence regarding its use for predicting clinically relevant anastomotic leakage and risk stratification remains insufficient [7, 8]. Therefore, this study aimed to evaluate the predictive value of the preoperative CONUT score for clinically relevant anastomotic leakage after gastric cancer surgery and to construct an anastomotic leakage risk prediction model incorporating the CONUT score and clinical operative factors, as well as a three-tier risk stratification system based on predicted probability. Materials and methods Study population Clinical data from patients who underwent radical gastrectomy at Qingdao Central Hospital, University of Health and Rehabilitation Sciences (Qingdao Central Hospital), from January 2021 to January 2025 were retrospectively analyzed. Patient treatment pathways followed the perioperative management procedures for gastric cancer at our center and referred to recent multidisciplinary guidelines for gastric cancer. For locally advanced gastric cancer, neoadjuvant therapy is generally determined by integrating clinical stage, resectability, symptom burden, comorbidities, patient preference, and treatment accessibility [9, 10]. The inclusion criteria were as follows: (1) postoperative pathological confirmation of gastric adenocarcinoma; (2) radical gastrectomy with R0 resection, defined as microscopically margin-negative resection, including distal gastrectomy, proximal gastrectomy, or total gastrectomy; (3) gastrointestinal reconstruction with completed anastomosis; (4) complete hematological and biochemical data obtained within 1 week before surgery and no later than 1 day before surgery; and (5) complete clinical records. The exclusion criteria were as follows: (1) preoperative neoadjuvant chemotherapy or radiotherapy; (2) other malignant tumors; (3) emergency surgery; (4) severe preoperative infection, active bleeding, or major organ failure; (5) preoperative therapeutic enteral nutrition/parenteral nutrition (EN/PN) support; and (6) incomplete clinical data. Patients who received preoperative therapeutic EN/PN support were excluded mainly because therapeutic nutritional support can alter serum albumin, peripheral blood lymphocyte count, total cholesterol, and the CONUT score in the short term, thereby confounding the association between unmodified preoperative nutritional status and postoperative anastomotic leakage. This exclusion criterion did not include routine dietary education, oral dietary advice, or preoperative nutritional risk screening. All patients received routine dietary education and nutritional risk assessment according to the perioperative procedures of our center. For patients with clear indications for nutritional support, the attending physician and nutrition team determined whether therapeutic EN/PN should be initiated according to clinical conditions. Among 348 initially screened patients, 27 were excluded because of preoperative neoadjuvant chemotherapy or radiotherapy, 21 because of preoperative therapeutic EN/PN support, 6 because of other malignant tumors, 4 because of emergency surgery, 4 because of severe preoperative infection, active bleeding, or major organ failure, and 3 because of incomplete clinical data. Finally, 283 patients were included. Among the 283 included patients, the reasons for not receiving therapeutic EN/PN were as follows: 246 patients did not meet the institutional criteria for initiating therapeutic EN/PN before surgery or required only routine dietary guidance; 21 patients had recommendations for nutritional support but declined additional therapeutic EN/PN after being fully informed; and 16 patients did not complete a standard course of therapeutic EN/PN because of tumor-related obstruction, bleeding tendency, or a short surgical waiting time that required prioritized surgery in a nonemergency setting. These circumstances were based on previous clinical pathway records. The tumor-node-metastasis (TNM) stage recorded in this study was the postoperative pathological stage. The 99 patients with pathological stage III disease in the final cohort all belonged to the direct-surgery cohort and had not received neoadjuvant therapy. The reasons included preoperative imaging and endoscopic assessments suggesting resectability and a clinical stage lower than the postoperative pathological stage, with postoperative upstaging to pathological stage III because of lymph node metastasis or local invasion (56 patients); pyloric obstruction, anemia, or bleeding tendency requiring prioritized surgery after multidisciplinary assessment (18 patients); advanced age, comorbidities, or general condition unsuitable for neoadjuvant therapy (13 patients); and patient refusal of neoadjuvant therapy after full explanation (12 patients) [11]. Therefore, the findings of this study are mainly applicable to patients who undergo radical gastrectomy directly without neoadjuvant therapy and cannot be directly extrapolated to patients with locally advanced gastric cancer receiving neoadjuvant therapy. This was a single-center retrospective study, and the sample size was determined by consecutive cases that met the inclusion and exclusion criteria during the study period. Because the number of primary outcome events was 39, the final risk prediction model included only five clinically meaningful variables retained in multivariable analysis. Bootstrap and temporal validation were used to assess model stability; however, the model remains at risk of overfitting and requires further external validation. This study was reviewed and approved by the Ethics Committee of Qingdao Central Hospital, University of Health and Rehabilitation Sciences (Qingdao Central Hospital), which served as the Institutional Review Board (IRB) (approval number: KY202601601). Because this retrospective observational study used only de-identified clinical data generated during previous diagnosis and treatment and did not increase participant risk, the IRB waived the requirement for individual written informed consent in accordance with the Measures for Ethical Review of Life Science and Medical Research Involving Humans and the institutional regulations for retrospective medical record studies. Data collection Data were collected from the hospital electronic medical record system and included the following categories: (1) general information: sex, age, body mass index (BMI), smoking history, drinking history, and comorbidities, including hypertension, diabetes mellitus, cardiovascular disease, and chronic obstructive pulmonary disease (COPD); (2) preoperative laboratory indicators: complete blood count within 1 week before surgery and no later than 1 day before surgery, including white blood cell (WBC) count, lymphocyte count, and hemoglobin; biochemical indicators, including serum albumin, total cholesterol, and total lymphocyte count; and tumor markers, including carcinoembryonic antigen (CEA) and carbohydrate antigen 19 − 9 (CA19-9); (3) tumor-related data: tumor location, tumor size, pathological type, differentiation degree, and TNM stage; (4) surgical data: surgical approach, extent of resection, reconstruction method, operative time, intraoperative blood loss, and number of lymph nodes dissected; and (5) postoperative complications: anastomotic leakage, pulmonary infection, intra-abdominal infection, wound infection, delayed gastric emptying, and other complications. TNM stage was recorded according to the eighth edition of the American Joint Committee on Cancer staging system, and all TNM stages reported in this study were postoperative pathological stages [12]. Diagnostic criteria for anastomotic leakage Anastomotic leakage was defined as digestive fluid drainage through an abdominal drain within 30 postoperative days, contrast extravasation detected by gastrointestinal radiography or computed tomography (CT), or anastomotic disruption identified during reoperation. According to the International Gastric Cancer Study Group grading criteria, anastomotic leakage is classified into three grades: grade I, requiring a change in management without invasive treatment; grade II, requiring invasive treatment without reoperation; and grade III, requiring surgical reintervention. The primary endpoint of this study was clinically relevant anastomotic leakage, defined as grade II or higher anastomotic leakage; the secondary endpoint was anastomotic leakage of any grade [13, 14]. Calculation of the CONUT score The CONUT score consists of serum albumin, peripheral blood total lymphocyte count, and total cholesterol and was calculated using the laboratory results closest to the date of surgery within 1 week before surgery [15]. Each component was scored according to its value, yielding a total score of 0–12 points: (1) serum albumin (g/L): ≥35, 0 points; 30-34.9, 2 points; 25-29.9, 4 points; and < 25, 6 points; (2) total lymphocyte count (×10⁹/L): ≥1.6, 0 points; 1.2–1.59, 1 point; 0.8–1.19, 2 points; and < 0.8, 3 points; and (3) total cholesterol (mmol/L): ≥4.65, 0 points; 3.62–4.64, 1 point; 2.59–3.61, 2 points; and < 2.59, 3 points. If a patient had multiple laboratory results, the data closest to the date of surgery were selected. Prediction model reporting and handling of missing data The development, validation, and reporting of the prediction model in this study followed recommendations for clinical prediction model reporting, with emphasis on model variables, β coefficients, discrimination, calibration, internal validation, and clinical net benefit [16, 17]. The key modeling variables and primary outcome had no missing values, and the main analysis was performed using complete cases. CEA and CA19-9 had small amounts of missing data and were analyzed using available cases only in baseline comparisons and univariable analysis. Multiple imputation by chained equations (MICE) was used for sensitivity analysis, and no substantial changes in variable direction or significance were observed after imputation (Supplementary Table 1). Statistical analysis SPSS version 26.0 was used for statistical analysis. Normality was assessed using the Shapiro-Wilk test, and homogeneity of variance was assessed using Levene’s test. Normally distributed continuous variables are presented as mean ± standard deviation (SD) and were compared between groups using independent-samples t tests. Nonnormally distributed continuous variables are presented as median (interquartile range) and were compared using the Mann-Whitney U test. Categorical variables are presented as number (percentage) [n (%)] and were compared using the χ² test or Fisher’s exact test. Trend analysis was performed using the trend χ² test. Logistic regression models were used to analyze risk factors for anastomotic leakage. All potential influencing factors were included in univariable analysis, and variables with P < 0.10 were entered into multivariable analysis. Multivariable analysis used backward stepwise selection with a removal criterion of P > 0.10, and collinearity was assessed using the variance inflation factor (VIF). The Hosmer-Lemeshow test was used to evaluate model goodness of fit. Receiver operating characteristic (ROC) curve analysis was used to evaluate the predictive value of the CONUT score and its components for anastomotic leakage. The area under the curve (AUC) and 95% confidence interval (CI) were calculated, and the sensitivity-prioritized screening threshold was determined according to the maximum Youden index. The DeLong test was used to compare AUCs between different indicators. Kaplan-Meier curves were used to evaluate the cumulative incidence of anastomotic leakage within 30 postoperative days in different CONUT score groups. The Log-rank test was used for between-group comparisons, and the Cox proportional hazards model was used to estimate the hazard ratio (HR) and 95% CI. The anastomotic leakage risk prediction model was constructed according to the results of multivariable analysis. Bootstrap resampling with 1000 repetitions was used for internal validation, and the corrected C-index, calibration slope, and calibration intercept were calculated. Exploratory temporal validation was performed by using patients enrolled from January 2021 to December 2023 as the modeling cohort and those enrolled from January 2024 to January 2025 as the temporal validation cohort, to describe model stability in patients treated during a later time period. Model discrimination in the validation cohort was expressed using the C-index and 95% CI, and calibration was assessed using the calibration slope, calibration intercept, Brier score, and Hosmer-Lemeshow test. Temporal validation is still a single-center internal validation method and cannot replace independent external validation. Calibration curves and decision curves were plotted to evaluate model calibration and clinical utility. A two-sided P < 0.05 was considered statistically significant. Results General characteristics A total of 283 patients who underwent radical gastrectomy were included in this study, including 189 men (66.8%) and 94 women (33.2%), with a mean age of 62.5 ± 10.8 years. Clinically relevant anastomotic leakage occurred in 39 patients (13.8%), including 0 grade I leaks, 28 grade II leaks (9.9%), and 11 grade III leaks (3.9%). In this study, the results for the primary endpoint were identical to those for anastomotic leakage of any grade. There were no statistically significant differences between the anastomotic leakage group and the non-anastomotic leakage group in age, sex, smoking history, drinking history, hypertension, cardiovascular disease, COPD, tumor location, pathological type, differentiation degree, surgical approach, reconstruction method, number of lymph nodes dissected, CEA, or CA19-9 (all P > 0.05). BMI, hemoglobin, serum albumin, total lymphocyte count, and total cholesterol were lower in the anastomotic leakage group than in the non-anastomotic leakage group, whereas WBC count, tumor size, proportion of TNM stage III disease, proportion of diabetes mellitus, proportion of total gastrectomy, operative time, intraoperative blood loss, and CONUT score were higher in the anastomotic leakage group; these differences were statistically significant (all P < 0.05) (Table 1). Table 1Comparison of clinical characteristics between the anastomotic leakage and non-anastomotic leakage groupsCharacteristicAnastomotic leakage group (n = 39)Non-anastomotic leakage group (n = 244)Statistic/P valueAge (years, mean ± SD)63.8 ± 11.262.2 ± 10.7t = 0.862, P = 0.390Sex [n (%)]χ² = 0.512, P = 0.474 Male28 (71.8)161 (66.0) Female11 (28.2)83 (34.0)BMI (kg/m², mean ± SD)21.3 ± 2.823.1 ± 3.2t = -3.315, P = 0.001Smoking history [n (%)]18 (46.2)102 (41.8)χ² = 0.261, P = 0.610Drinking history [n (%)]15 (38.5)86 (35.2)χ² = 0.151, P = 0.697Comorbidities [n (%)] Hypertension14 (35.9)78 (32.0)χ² = 0.237, P = 0.627 Diabetes mellitus12 (30.8)39 (16.0)χ² = 4.976, P = 0.026 Cardiovascular disease8 (20.5)42 (17.2)χ² = 0.252, P = 0.616 COPD5 (12.8)23 (9.4)χ² = 0.435, P = 0.510WBC count (×10⁹/L, mean ± SD)7.8 ± 2.36.9 ± 2.1t = 2.452, P = 0.015Hemoglobin (g/L, mean ± SD)118.6 ± 18.5128.3 ± 16.7t = -3.318, P = 0.001Tumor location [n (%)]χ² = 0.808, P = 0.668 Upper third16 (41.0)85 (34.8) Middle third12 (30.8)92 (37.7) Lower third11 (28.2)67 (27.5)Tumor size (cm, mean ± SD)4.8 ± 1.94.0 ± 1.6t = 2.822, P = 0.005Pathological type [n (%)]χ² = 0.293, P = 0.588 Intestinal type18 (46.2)124 (50.8) Diffuse type21 (53.8)120 (49.2)Differentiation degree [n (%)]χ² = 1.000, P = 0.317 Well/moderately differentiated13 (33.3)102 (41.8) Poorly differentiated26 (66.7)142 (58.2)TNM stage [n (%)]χ² = 7.369, P = 0.025 Stage I6 (15.4)68 (27.9) Stage II12 (30.8)98 (40.2) Stage III21 (53.8)78 (32.0)Surgical approach [n (%)]χ² = 0.726, P = 0.394 Open surgery23 (59.0)126 (51.6) Laparoscopic surgery16 (41.0)118 (48.4)Extent of resection [n (%)]χ² = 8.973, P = 0.011 Distal gastrectomy15 (38.5)135 (55.3) Proximal gastrectomy6 (15.4)52 (21.3) Total gastrectomy18 (46.2)57 (23.4)Reconstruction method [n (%)]χ² = 3.840, P = 0.147 Billroth I8 (20.5)72 (29.5) Billroth II7 (17.9)63 (25.8) Roux-en-Y24 (61.5)109 (44.7)Operative time (min, mean ± SD)248.6 ± 52.3215.7 ± 48.9t = 3.864, P < 0.001Intraoperative blood loss (mL, mean ± SD)285.4 ± 98.6218.3 ± 86.7t = 4.401, P < 0.001Number of lymph nodes dissected (mean ± SD)32.8 ± 12.530.6 ± 11.8t = 1.072, P = 0.285Serum albumin (g/L, mean ± SD)33.2 ± 4.838.5 ± 4.2t = -7.171, P < 0.001Total lymphocyte count (×10⁹/L, mean ± SD)1.18 ± 0.421.65 ± 0.58t = -4.858, P < 0.001Total cholesterol (mmol/L, mean ± SD)3.52 ± 0.864.18 ± 0.95t = -4.079, P < 0.001CEA (ng/mL, median [IQR])8.6 (3.2, 18.5)5.8 (2.4, 12.3)Z = -1.856, P = 0.063CA19-9 (U/mL, median [IQR])42.6 (15.3, 98.7)28.4 (12.6, 65.8)Z = -1.732, P = 0.083CONUT score (median [IQR])4 (3, 6)2 (1, 3)Z = -5.892, P < 0.001Normally distributed continuous variables are expressed as mean ± SD, nonnormally distributed continuous variables as median (IQR), and categorical variables as n (%). t values were obtained from independent-samples t tests, Z values from Mann-Whitney U tests, and χ² values from Pearson χ² tests or Fisher exact tests; P values are two-sidedBMI body mass index, COPD chronic obstructive pulmonary disease, WBC white blood cell, CEA carcinoembryonic antigen, CA19-9 carbohydrate antigen 19 − 9, TNM tumor-node-metastasis, CONUT Controlling Nutritional Status, SD standard deviation, IQR interquartile range, χ² chi-square statistic Table 1Comparison of clinical characteristics between the anastomotic leakage and non-anastomotic leakage groupsCharacteristicAnastomotic leakage group (n = 39)Non-anastomotic leakage group (n = 244)Statistic/P valueAge (years, mean ± SD)63.8 ± 11.262.2 ± 10.7t = 0.862, P = 0.390Sex [n (%)]χ² = 0.512, P = 0.474 Male28 (71.8)161 (66.0) Female11 (28.2)83 (34.0)BMI (kg/m², mean ± SD)21.3 ± 2.823.1 ± 3.2t = -3.315, P = 0.001Smoking history [n (%)]18 (46.2)102 (41.8)χ² = 0.261, P = 0.610Drinking history [n (%)]15 (38.5)86 (35.2)χ² = 0.151, P = 0.697Comorbidities [n (%)] Hypertension14 (35.9)78 (32.0)χ² = 0.237, P = 0.627 Diabetes mellitus12 (30.8)39 (16.0)χ² = 4.976, P = 0.026 Cardiovascular disease8 (20.5)42 (17.2)χ² = 0.252, P = 0.616 COPD5 (12.8)23 (9.4)χ² = 0.435, P = 0.510WBC count (×10⁹/L, mean ± SD)7.8 ± 2.36.9 ± 2.1t = 2.452, P = 0.015Hemoglobin (g/L, mean ± SD)118.6 ± 18.5128.3 ± 16.7t = -3.318, P = 0.001Tumor location [n (%)]χ² = 0.808, P = 0.668 Upper third16 (41.0)85 (34.8) Middle third12 (30.8)92 (37.7) Lower third11 (28.2)67 (27.5)Tumor size (cm, mean ± SD)4.8 ± 1.94.0 ± 1.6t = 2.822, P = 0.005Pathological type [n (%)]χ² = 0.293, P = 0.588 Intestinal type18 (46.2)124 (50.8) Diffuse type21 (53.8)120 (49.2)Differentiation degree [n (%)]χ² = 1.000, P = 0.317 Well/moderately differentiated13 (33.3)102 (41.8) Poorly differentiated26 (66.7)142 (58.2)TNM stage [n (%)]χ² = 7.369, P = 0.025 Stage I6 (15.4)68 (27.9) Stage II12 (30.8)98 (40.2) Stage III21 (53.8)78 (32.0)Surgical approach [n (%)]χ² = 0.726, P = 0.394 Open surgery23 (59.0)126 (51.6) Laparoscopic surgery16 (41.0)118 (48.4)Extent of resection [n (%)]χ² = 8.973, P = 0.011 Distal gastrectomy15 (38.5)135 (55.3) Proximal gastrectomy6 (15.4)52 (21.3) Total gastrectomy18 (46.2)57 (23.4)Reconstruction method [n (%)]χ² = 3.840, P = 0.147 Billroth I8 (20.5)72 (29.5) Billroth II7 (17.9)63 (25.8) Roux-en-Y24 (61.5)109 (44.7)Operative time (min, mean ± SD)248.6 ± 52.3215.7 ± 48.9t = 3.864, P < 0.001Intraoperative blood loss (mL, mean ± SD)285.4 ± 98.6218.3 ± 86.7t = 4.401, P < 0.001Number of lymph nodes dissected (mean ± SD)32.8 ± 12.530.6 ± 11.8t = 1.072, P = 0.285Serum albumin (g/L, mean ± SD)33.2 ± 4.838.5 ± 4.2t = -7.171, P < 0.001Total lymphocyte count (×10⁹/L, mean ± SD)1.18 ± 0.421.65 ± 0.58t = -4.858, P < 0.001Total cholesterol (mmol/L, mean ± SD)3.52 ± 0.864.18 ± 0.95t = -4.079, P < 0.001CEA (ng/mL, median [IQR])8.6 (3.2, 18.5)5.8 (2.4, 12.3)Z = -1.856, P = 0.063CA19-9 (U/mL, median [IQR])42.6 (15.3, 98.7)28.4 (12.6, 65.8)Z = -1.732, P = 0.083CONUT score (median [IQR])4 (3, 6)2 (1, 3)Z = -5.892, P < 0.001Normally distributed continuous variables are expressed as mean ± SD, nonnormally distributed continuous variables as median (IQR), and categorical variables as n (%). t values were obtained from independent-samples t tests, Z values from Mann-Whitney U tests, and χ² values from Pearson χ² tests or Fisher exact tests; P values are two-sidedBMI body mass index, COPD chronic obstructive pulmonary disease, WBC white blood cell, CEA carcinoembryonic antigen, CA19-9 carbohydrate antigen 19 − 9, TNM tumor-node-metastasis, CONUT Controlling Nutritional Status, SD standard deviation, IQR interquartile range, χ² chi-square statistic Comparison of clinical characteristics between the anastomotic leakage and non-anastomotic leakage groups Normally distributed continuous variables are expressed as mean ± SD, nonnormally distributed continuous variables as median (IQR), and categorical variables as n (%). t values were obtained from independent-samples t tests, Z values from Mann-Whitney U tests, and χ² values from Pearson χ² tests or Fisher exact tests; P values are two-sided BMI body mass index, COPD chronic obstructive pulmonary disease, WBC white blood cell, CEA carcinoembryonic antigen, CA19-9 carbohydrate antigen 19 − 9, TNM tumor-node-metastasis, CONUT Controlling Nutritional Status, SD standard deviation, IQR interquartile range, χ² chi-square statistic Univariable analysis of risk factors for anastomotic leakage Higher BMI, hemoglobin, serum albumin, total lymphocyte count, and total cholesterol were associated with a lower risk of anastomotic leakage. Diabetes mellitus, WBC count, tumor size, TNM stage III disease, total gastrectomy, operative time, intraoperative blood loss, and higher CONUT score were associated with a higher risk of anastomotic leakage (all P < 0.10) (Table 2). Table 2Univariable Logistic regression analysis of risk factors for anastomotic leakageVariableOR95% CIP valueAge1.0140.986–1.0430.327Sex (male vs. female)1.2990.614–2.7470.494BMI0.8250.726–0.9380.003Smoking history1.2020.606–2.3820.599Drinking history1.1530.558–2.3830.702Hypertension1.1920.587–2.4210.627Diabetes mellitus2.3161.094–4.9030.028Cardiovascular disease1.2410.543–2.8380.609COPD1.4060.512–3.8630.507WBC count1.1421.018–1.2820.023Hemoglobin0.9720.953–0.9910.004Pathological type (diffuse vs. intestinal)1.2050.612–2.3720.590Tumor size1.1821.023–1.3660.008Poor differentiation1.4320.701–2.9240.323TNM stage III2.6541.287–5.4710.008Laparoscopic surgery0.7400.375–1.4590.384Proximal gastrectomy1.1250.441–2.8710.805Total gastrectomy2.8251.411–5.6520.003Roux-en-Y reconstruction1.9820.996–3.9470.051Operative time1.0121.006–1.018< 0.001Intraoperative blood loss1.0061.003–1.009< 0.001Number of lymph nodes dissected1.0150.991–1.0400.224Serum albumin0.8020.736–0.874< 0.001Total lymphocyte count0.1860.084–0.413< 0.001Total cholesterol0.6580.456–0.9500.026CEA1.0120.996–1.0280.142CA19-91.0030.999–1.0070.158CONUT score1.6821.428–1.981< 0.001Univariable Logistic regression was used to evaluate the association between each variable and clinically relevant anastomotic leakage; continuous variables were entered into the model in their original units, and categorical variables used the reference levels listed in the table. P values are two-sidedOR odds ratio, CI confidence interval, BMI body mass index, COPD chronic obstructive pulmonary disease, WBC white blood cell, CEA carcinoembryonic antigen, CA19-9 carbohydrate antigen 19 − 9, TNM tumor-node-metastasis, CONUT Controlling Nutritional Status Table 2Univariable Logistic regression analysis of risk factors for anastomotic leakageVariableOR95% CIP valueAge1.0140.986–1.0430.327Sex (male vs. female)1.2990.614–2.7470.494BMI0.8250.726–0.9380.003Smoking history1.2020.606–2.3820.599Drinking history1.1530.558–2.3830.702Hypertension1.1920.587–2.4210.627Diabetes mellitus2.3161.094–4.9030.028Cardiovascular disease1.2410.543–2.8380.609COPD1.4060.512–3.8630.507WBC count1.1421.018–1.2820.023Hemoglobin0.9720.953–0.9910.004Pathological type (diffuse vs. intestinal)1.2050.612–2.3720.590Tumor size1.1821.023–1.3660.008Poor differentiation1.4320.701–2.9240.323TNM stage III2.6541.287–5.4710.008Laparoscopic surgery0.7400.375–1.4590.384Proximal gastrectomy1.1250.441–2.8710.805Total gastrectomy2.8251.411–5.6520.003Roux-en-Y reconstruction1.9820.996–3.9470.051Operative time1.0121.006–1.018< 0.001Intraoperative blood loss1.0061.003–1.009< 0.001Number of lymph nodes dissected1.0150.991–1.0400.224Serum albumin0.8020.736–0.874< 0.001Total lymphocyte count0.1860.084–0.413< 0.001Total cholesterol0.6580.456–0.9500.026CEA1.0120.996–1.0280.142CA19-91.0030.999–1.0070.158CONUT score1.6821.428–1.981< 0.001Univariable Logistic regression was used to evaluate the association between each variable and clinically relevant anastomotic leakage; continuous variables were entered into the model in their original units, and categorical variables used the reference levels listed in the table. P values are two-sidedOR odds ratio, CI confidence interval, BMI body mass index, COPD chronic obstructive pulmonary disease, WBC white blood cell, CEA carcinoembryonic antigen, CA19-9 carbohydrate antigen 19 − 9, TNM tumor-node-metastasis, CONUT Controlling Nutritional Status Univariable Logistic regression analysis of risk factors for anastomotic leakage Univariable Logistic regression was used to evaluate the association between each variable and clinically relevant anastomotic leakage; continuous variables were entered into the model in their original units, and categorical variables used the reference levels listed in the table. P values are two-sided OR odds ratio, CI confidence interval, BMI body mass index, COPD chronic obstructive pulmonary disease, WBC white blood cell, CEA carcinoembryonic antigen, CA19-9 carbohydrate antigen 19 − 9, TNM tumor-node-metastasis, CONUT Controlling Nutritional Status Multivariable analysis of independent risk factors for anastomotic leakage Variables with P < 0.10 in univariable analysis were entered into multivariable Logistic regression analysis. Because the CONUT score consists of serum albumin, total lymphocyte count, and total cholesterol, two models were used to avoid component overlap: Model 1 included the CONUT score but not its components, whereas Model 2 included the components of the CONUT score but not the total CONUT score. Collinearity diagnostics showed that all variables in Model 1 had VIF values < 5. CONUT score, diabetes mellitus, total gastrectomy, operative time, and intraoperative blood loss were independent risk factors for clinically relevant anastomotic leakage (all P < 0.05). The nutritional-component model showed that serum albumin remained significantly associated with anastomotic leakage risk, whereas total lymphocyte count and total cholesterol were not statistically significant after adjustment (both P > 0.05) (Table 3). Table 3Multivariable Logistic regression analysis of independent risk factors for anastomotic leakageVariableModel 1 βModel 1 SEModel 1 ORModel 1 95% CIModel 1 P valueModel 2 ORModel 2 95% CIModel 2 P valueCONUT score0.3540.1011.4251.168–1.739< 0.001———Serum albumin—————0.8150.748–0.889< 0.001Total lymphocyte count—————0.6240.245–1.5880.323Total cholesterol—————0.8620.583–1.2750.456Diabetes mellitus0.9110.4392.4861.052–5.8760.0382.3921.018–5.6190.045Total gastrectomy0.8410.4152.3181.028–5.2280.0432.2851.012–5.1590.047Operative time0.0080.0031.0081.002–1.0140.0121.0081.002–1.0140.011Intraoperative blood loss0.0040.0021.0041.001–1.0070.0181.0041.001–1.0070.016Model 1 included the CONUT score but not its components; Model 2 included the components of the CONUT score but not the total CONUT score. The C-index of Model 1 was 0.838 (95% CI: 0.779–0.897), and the Hosmer-Lemeshow test showed χ² = 6.324 and P = 0.611. The C-index of Model 2 was 0.802 (95% CI: 0.738–0.866), and the Hosmer-Lemeshow test showed χ² = 7.814 and P = 0.452. P values are two-sidedβ Logistic regression coefficient, SE standard error, OR odds ratio, CI confidence interval, CONUT Controlling Nutritional Status, C-index concordance index, χ² chi-square statistic Table 3Multivariable Logistic regression analysis of independent risk factors for anastomotic leakageVariableModel 1 βModel 1 SEModel 1 ORModel 1 95% CIModel 1 P valueModel 2 ORModel 2 95% CIModel 2 P valueCONUT score0.3540.1011.4251.168–1.739< 0.001———Serum albumin—————0.8150.748–0.889< 0.001Total lymphocyte count—————0.6240.245–1.5880.323Total cholesterol—————0.8620.583–1.2750.456Diabetes mellitus0.9110.4392.4861.052–5.8760.0382.3921.018–5.6190.045Total gastrectomy0.8410.4152.3181.028–5.2280.0432.2851.012–5.1590.047Operative time0.0080.0031.0081.002–1.0140.0121.0081.002–1.0140.011Intraoperative blood loss0.0040.0021.0041.001–1.0070.0181.0041.001–1.0070.016Model 1 included the CONUT score but not its components; Model 2 included the components of the CONUT score but not the total CONUT score. The C-index of Model 1 was 0.838 (95% CI: 0.779–0.897), and the Hosmer-Lemeshow test showed χ² = 6.324 and P = 0.611. The C-index of Model 2 was 0.802 (95% CI: 0.738–0.866), and the Hosmer-Lemeshow test showed χ² = 7.814 and P = 0.452. P values are two-sidedβ Logistic regression coefficient, SE standard error, OR odds ratio, CI confidence interval, CONUT Controlling Nutritional Status, C-index concordance index, χ² chi-square statistic Multivariable Logistic regression analysis of independent risk factors for anastomotic leakage Model 1 included the CONUT score but not its components; Model 2 included the components of the CONUT score but not the total CONUT score. The C-index of Model 1 was 0.838 (95% CI: 0.779–0.897), and the Hosmer-Lemeshow test showed χ² = 6.324 and P = 0.611. The C-index of Model 2 was 0.802 (95% CI: 0.738–0.866), and the Hosmer-Lemeshow test showed χ² = 7.814 and P = 0.452. P values are two-sided β Logistic regression coefficient, SE standard error, OR odds ratio, CI confidence interval, CONUT Controlling Nutritional Status, C-index concordance index, χ² chi-square statistic Predictive value of the CONUT score for anastomotic leakage The CONUT score showed moderate discrimination for clinically relevant anastomotic leakage, with a higher AUC than serum albumin, total lymphocyte count, and total cholesterol. Although the AUC difference versus serum albumin was statistically significant, it represented a modest incremental improvement (Table 4, Fig. 1). Table 4ROC curve analysis of different indicators for predicting anastomotic leakageIndicatorAUC95% CISensitivity (%)Specificity (%)P valueP value versus CONUT scoreCONUT score0.7820.716–0.84874.462.3< 0.001—Serum albumin0.7240.653–0.79669.268.4< 0.0010.042Total lymphocyte count0.6980.624–0.77264.167.2< 0.0010.008Total cholesterol0.6120.532–0.69156.462.70.012< 0.001P values indicate two-sided test results comparing the AUC of each indicator with 0.5; P values versus the CONUT score were calculated using the DeLong test, which compares differences in AUCs between correlated ROC curvesROC receiver operating characteristic, AUC area under the curve, CI confidence interval, CONUT Controlling Nutritional Status Table 4ROC curve analysis of different indicators for predicting anastomotic leakageIndicatorAUC95% CISensitivity (%)Specificity (%)P valueP value versus CONUT scoreCONUT score0.7820.716–0.84874.462.3< 0.001—Serum albumin0.7240.653–0.79669.268.4< 0.0010.042Total lymphocyte count0.6980.624–0.77264.167.2< 0.0010.008Total cholesterol0.6120.532–0.69156.462.70.012< 0.001P values indicate two-sided test results comparing the AUC of each indicator with 0.5; P values versus the CONUT score were calculated using the DeLong test, which compares differences in AUCs between correlated ROC curvesROC receiver operating characteristic, AUC area under the curve, CI confidence interval, CONUT Controlling Nutritional Status ROC curve analysis of different indicators for predicting anastomotic leakage P values indicate two-sided test results comparing the AUC of each indicator with 0.5; P values versus the CONUT score were calculated using the DeLong test, which compares differences in AUCs between correlated ROC curves ROC receiver operating characteristic, AUC area under the curve, CI confidence interval, CONUT Controlling Nutritional Status Fig. 1Receiver operating characteristic curves of the preoperative Controlling Nutritional Status score and its components for predicting clinically relevant anastomotic leakage. The curves show the discriminative ability of the CONUT score, serum albumin, total lymphocyte count, and total cholesterol for clinically relevant anastomotic leakage within 30 postoperative days. The AUC of the CONUT score was 0.782 (95% CI: 0.716–0.848), which was higher than that of each single nutritional indicator. The AUC difference compared with serum albumin was 0.058 (DeLong test P = 0.042), indicating modest incremental discrimination. AUC = area under the curve; CI = confidence interval; CONUT = Controlling Nutritional Status; ROC = receiver operating characteristic Receiver operating characteristic curves of the preoperative Controlling Nutritional Status score and its components for predicting clinically relevant anastomotic leakage. The curves show the discriminative ability of the CONUT score, serum albumin, total lymphocyte count, and total cholesterol for clinically relevant anastomotic leakage within 30 postoperative days. The AUC of the CONUT score was 0.782 (95% CI: 0.716–0.848), which was higher than that of each single nutritional indicator. The AUC difference compared with serum albumin was 0.058 (DeLong test P = 0.042), indicating modest incremental discrimination. AUC = area under the curve; CI = confidence interval; CONUT = Controlling Nutritional Status; ROC = receiver operating characteristic The AUC of the CONUT score for predicting anastomotic leakage was 0.782 (95% CI: 0.716–0.848, P < 0.001). For clinical screening purposes, this study used a CONUT score ≥ 3 as a sensitivity-prioritized screening threshold. This threshold corresponded to a sensitivity of 74.4% (29/39), specificity of 62.3% (152/244), positive predictive value of 24.0% (29/121), negative predictive value of 93.8% (152/162), and false-positive rate of 37.7% (92/244). Further exploration of a higher-specificity threshold showed that a CONUT score ≥ 4 corresponded to a sensitivity of 59.0% (23/39), specificity of 78.3% (191/244), positive predictive value of 30.3% (23/76), and negative predictive value of 92.3% (191/207). The number of false positives decreased from 92 to 53, whereas the number of missed cases increased from 10 to 16. Therefore, a threshold of ≥ 3 is more suitable for screening and enhanced monitoring, whereas ≥ 4 may be used as an alternative threshold when reducing false-positive interventions is prioritized (Table 4, Supplementary Table 2). Comparison of anastomotic leakage incidence across CONUT score groups The incidence of anastomotic leakage increased progressively with higher CONUT score categories (P < 0.001). The cumulative risk of clinically relevant anastomotic leakage within 30 postoperative days was higher in the high-CONUT group than in the low-CONUT group (Log-rank P < 0.001; HR = 4.28, 95% CI: 2.09–8.76) (Table 5, Fig. 2). Table 5Incidence of anastomotic leakage across CONUT score groupsCONUT scoreNo. of patientsAnastomotic leakage [n (%)]Crude OR (95% CI)Statistic/P value0–216210 (6.2)1.00 (reference)3456 (13.3)2.34 (0.80–6.83)4–55514 (25.5)5.19 (2.15–12.54)≥ 6219 (42.9)11.40 (3.89–33.41)Pearson χ² = 29.150, P < 0.001Trend χ² testTrend χ² = 28.195, P < 0.001Crude ORs were estimated by unadjusted Logistic regression using the CONUT score 0–2 group as the reference. The Pearson χ² test was used to compare differences in categorical distributions across groups, and the trend χ² test was used to evaluate the linear trend between score category and anastomotic leakage incidence. P values are two-sidedCONUT Controlling Nutritional Status, OR odds ratio, CI confidence interval, χ² chi-square statistic Table 5Incidence of anastomotic leakage across CONUT score groupsCONUT scoreNo. of patientsAnastomotic leakage [n (%)]Crude OR (95% CI)Statistic/P value0–216210 (6.2)1.00 (reference)3456 (13.3)2.34 (0.80–6.83)4–55514 (25.5)5.19 (2.15–12.54)≥ 6219 (42.9)11.40 (3.89–33.41)Pearson χ² = 29.150, P < 0.001Trend χ² testTrend χ² = 28.195, P < 0.001Crude ORs were estimated by unadjusted Logistic regression using the CONUT score 0–2 group as the reference. The Pearson χ² test was used to compare differences in categorical distributions across groups, and the trend χ² test was used to evaluate the linear trend between score category and anastomotic leakage incidence. P values are two-sidedCONUT Controlling Nutritional Status, OR odds ratio, CI confidence interval, χ² chi-square statistic Incidence of anastomotic leakage across CONUT score groups Crude ORs were estimated by unadjusted Logistic regression using the CONUT score 0–2 group as the reference. The Pearson χ² test was used to compare differences in categorical distributions across groups, and the trend χ² test was used to evaluate the linear trend between score category and anastomotic leakage incidence. P values are two-sided CONUT Controlling Nutritional Status, OR odds ratio, CI confidence interval, χ² chi-square statistic Fig. 2Cumulative incidence curves of clinically relevant anastomotic leakage within 30 postoperative days by Controlling Nutritional Status score category. The Kaplan-Meier curves show that the cumulative incidence of clinically relevant anastomotic leakage within 30 postoperative days was higher in the high-CONUT score group than in the low-CONUT score group, with a statistically significant between-group difference (Log-rank P < 0.001; Cox proportional hazards model HR = 4.28, 95% CI: 2.09–8.76). CI = confidence interval; CONUT = Controlling Nutritional Status; HR = hazard ratio Cumulative incidence curves of clinically relevant anastomotic leakage within 30 postoperative days by Controlling Nutritional Status score category. The Kaplan-Meier curves show that the cumulative incidence of clinically relevant anastomotic leakage within 30 postoperative days was higher in the high-CONUT score group than in the low-CONUT score group, with a statistically significant between-group difference (Log-rank P < 0.001; Cox proportional hazards model HR = 4.28, 95% CI: 2.09–8.76). CI = confidence interval; CONUT = Controlling Nutritional Status; HR = hazard ratio According to the sensitivity-prioritized threshold of CONUT score ≥ 3, patients were divided into a low-CONUT group (CONUT score 0–2, n = 162) and a high-CONUT group (CONUT score ≥ 3, n = 121). The incidence of anastomotic leakage was 6.2% (10/162) in the low-CONUT group and 24.0% (29/121) in the high-CONUT group, with a statistically significant difference (χ² = 17.386, P < 0.001). Further classification into four groups according to CONUT scores of 0–2, 3, 4–5, and ≥ 6 showed a stepwise increase in the incidence of anastomotic leakage as the score increased (Pearson χ² = 29.150, P < 0.001; trend χ² = 28.195, P < 0.001) (Table 5). Construction and validation of the anastomotic leakage risk prediction model Based on the results of multivariable analysis (Model 1), the CONUT score, diabetes mellitus, total gastrectomy, operative time, and intraoperative blood loss were included to construct the anastomotic leakage risk prediction model. The Logistic regression equation was as follows: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} &\mathrm{Logit}(P)\;=-8.652\;+0.354\;\times\mathrm{CONUT}\;\mathrm{score}\;+0.911\;\\&\times\mathrm{diabetes}\;\mathrm{mellitus}\;(\mathrm{yes}\;=\;1,\;\mathrm{no}\;=\;0)\;+\;0.841\;\\&\times\;\mathrm{total}\;\mathrm{gastrectomy}\;(\mathrm{yes}\;=\;1,\;\mathrm{no}\;=\;0)\;+\;0.008\;\\&\times\;\mathrm{operative}\;\mathrm{time}\;(\min)\;+\;0.004\;\\&\times\;\mathrm{intraoperative}\;\mathrm{blood}\;\mathrm{loss}\;(\mathrm{mL}) \end{aligned}$$\end{document}, and the predicted probability was \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P=\exp\left[\mathrm{Logit}\left(P\right)\right]/\left\{1+\exp\left[\mathrm{Logit}\left(P\right)\right]\right\}$$\end{document}. For bedside estimation, the simplified score used the β coefficient of the CONUT score (0.354) as one weight unit. Other β coefficients were divided by 0.354 and rounded to clinically convenient units: diabetes mellitus, 0.911/0.354 = 2.57, assigned 2.5 points; total gastrectomy, 0.841/0.354 = 2.38, assigned 2.5 points; operative time per 45 min, 0.008 × 45/0.354 = 1.02, assigned 1 point; and intraoperative blood loss per 90 mL, 0.004 × 90/0.354 = 1.02, assigned 1 point. Therefore, the bedside simplified score was calculated as CONUT score × 1 + diabetes mellitus × 2.5 + total gastrectomy × 2.5 + operative time/45 + intraoperative blood loss/90. The main statistical inference and risk stratification were based on exact Logistic predicted probabilities, and the simplified score was used only as an auxiliary bedside estimation tool (Supplementary Table 4). The apparent C-index of this model was 0.838 (95% CI: 0.779–0.897). Bootstrap internal validation with 1000 resamples showed a corrected C-index of 0.826 (95% CI: 0.768–0.884). Exploratory temporal validation showed that the modeling cohort included 206 patients, of whom 28 developed anastomotic leakage, and the temporal validation cohort included 77 patients, of whom 11 developed anastomotic leakage. The event rates did not differ significantly between the two cohorts (χ² = 0.023, P = 0.880). In the temporal validation cohort, the C-index was 0.803 (95% CI: 0.701–0.905), the Brier score was 0.104, the calibration slope was 0.912, the calibration intercept was 0.036, and the Hosmer-Lemeshow test showed χ² = 5.821 and P = 0.667. The risk prediction model maintained relatively stable discrimination in the full cohort, bootstrap internal validation, and temporal validation. Decision curve analysis showed positive net benefit in the low-to-moderate threshold probability range (Table 6, Figs. 3 and 4, Supplementary Table 3). Table 6Performance evaluation of the anastomotic leakage risk prediction modelPerformance measureApparent model/full cohortBootstrap internal validationTemporal validationDescriptionSample size283 patients1000 resamples77 patientsThe temporal validation cohort included patients treated from January 2024 to January 2025Anastomotic leakage events39 eventsEstimated according to resampling distribution11 eventsThe modeling cohort had 28 events; 28 + 11 = 39C-index0.838 (95% CI: 0.779–0.897)0.826 (95% CI: 0.768–0.884)0.803 (95% CI: 0.701–0.905)Temporal validation was exploratory internal validationCalibration slope1.000 (reference)0.968 (SE = 0.042)0.912A slope < 1 indicates possible mild overfittingCalibration intercept0.000 (reference)-0.018 (SE = 0.036)0.036A value close to 0 indicates limited systematic biasHosmer-Lemeshow testχ² = 6.324, P = 0.611Not applicableχ² = 5.821, P = 0.667P > 0.05 suggests no obvious lack of fitBrier score0.092 (95% CI: 0.068–0.116)0.0960.104Lower values indicate smaller overall prediction errorIDI/NRIIDI = 0.186; NRI = 0.412Not applicableNot applicableReference model included only the CONUT scoreThe C-index evaluates model discrimination; the calibration slope and intercept evaluate agreement between predicted and observed probabilities; the Brier score evaluates overall prediction error; the Hosmer-Lemeshow test evaluates goodness of fit; IDI and NRI used the model including only the CONUT score as the reference model. P values are probability valuesCI confidence interval, SE standard error, IDI integrated discrimination improvement, NRI net reclassification improvement, C-index concordance index, χ² chi-square statistic Table 6Performance evaluation of the anastomotic leakage risk prediction modelPerformance measureApparent model/full cohortBootstrap internal validationTemporal validationDescriptionSample size283 patients1000 resamples77 patientsThe temporal validation cohort included patients treated from January 2024 to January 2025Anastomotic leakage events39 eventsEstimated according to resampling distribution11 eventsThe modeling cohort had 28 events; 28 + 11 = 39C-index0.838 (95% CI: 0.779–0.897)0.826 (95% CI: 0.768–0.884)0.803 (95% CI: 0.701–0.905)Temporal validation was exploratory internal validationCalibration slope1.000 (reference)0.968 (SE = 0.042)0.912A slope < 1 indicates possible mild overfittingCalibration intercept0.000 (reference)-0.018 (SE = 0.036)0.036A value close to 0 indicates limited systematic biasHosmer-Lemeshow testχ² = 6.324, P = 0.611Not applicableχ² = 5.821, P = 0.667P > 0.05 suggests no obvious lack of fitBrier score0.092 (95% CI: 0.068–0.116)0.0960.104Lower values indicate smaller overall prediction errorIDI/NRIIDI = 0.186; NRI = 0.412Not applicableNot applicableReference model included only the CONUT scoreThe C-index evaluates model discrimination; the calibration slope and intercept evaluate agreement between predicted and observed probabilities; the Brier score evaluates overall prediction error; the Hosmer-Lemeshow test evaluates goodness of fit; IDI and NRI used the model including only the CONUT score as the reference model. P values are probability valuesCI confidence interval, SE standard error, IDI integrated discrimination improvement, NRI net reclassification improvement, C-index concordance index, χ² chi-square statistic Performance evaluation of the anastomotic leakage risk prediction model The C-index evaluates model discrimination; the calibration slope and intercept evaluate agreement between predicted and observed probabilities; the Brier score evaluates overall prediction error; the Hosmer-Lemeshow test evaluates goodness of fit; IDI and NRI used the model including only the CONUT score as the reference model. P values are probability values CI confidence interval, SE standard error, IDI integrated discrimination improvement, NRI net reclassification improvement, C-index concordance index, χ² chi-square statistic Fig. 3Calibration curve of the anastomotic leakage risk prediction model. The calibration curve compares model-predicted probabilities with observed anastomotic leakage probabilities. After bootstrap internal validation, the calibration slope was 0.968 and the calibration intercept was − 0.018, indicating acceptable overall calibration. Bootstrap is an internal validation method based on repeated resampling; a calibration slope below 1 suggests possible mild overfitting, and a calibration intercept close to 0 suggests limited systematic bias Calibration curve of the anastomotic leakage risk prediction model. The calibration curve compares model-predicted probabilities with observed anastomotic leakage probabilities. After bootstrap internal validation, the calibration slope was 0.968 and the calibration intercept was − 0.018, indicating acceptable overall calibration. Bootstrap is an internal validation method based on repeated resampling; a calibration slope below 1 suggests possible mild overfitting, and a calibration intercept close to 0 suggests limited systematic bias Fig. 4Decision curve analysis of the anastomotic leakage risk prediction model. The decision curve evaluates the clinical net benefit of the model at different threshold probabilities. The model had positive net benefit within the 5%-20% threshold probability range and generally provided greater net benefit than the default strategies of intensified management for all patients or no intensified management for any patient. Net benefit at higher threshold probabilities should be interpreted cautiously. DCA = decision curve analysis Decision curve analysis of the anastomotic leakage risk prediction model. The decision curve evaluates the clinical net benefit of the model at different threshold probabilities. The model had positive net benefit within the 5%-20% threshold probability range and generally provided greater net benefit than the default strategies of intensified management for all patients or no intensified management for any patient. Net benefit at higher threshold probabilities should be interpreted cautiously. DCA = decision curve analysis Decision curve analysis showed that the model had positive net benefit within the 5%-20% threshold probability range and generally provided greater net benefit than the default strategies of intensified management for all patients or no intensified management for any patient. When the threshold probability exceeded 20%, the net benefit gradually approached 0, with only a small positive net benefit around 45%. Therefore, clinical application of the model should mainly focus on the low-to-moderate threshold probability range, and net benefit at higher thresholds should be interpreted cautiously (Fig. 4, Supplementary Table 3). Three-tier risk stratification based on predicted probability After classification according to model-predicted probability, the incidence of anastomotic leakage increased stepwise in the low-, intermediate-, and high-risk groups, and both the between-group difference and trend test were statistically significant (all P < 0.001) (Table 7). Table 7Incidence of anastomotic leakage across risk strataRisk stratumNo. of patientsAnastomotic leakage [n (%)]Crude OR (95% CI)Adjusted ORᵃ (95% CI)Statistic/P valueLow-risk group1384 (2.9)1.00 (reference)1.00 (reference)Intermediate-risk group9813 (13.3)5.1 (1.6–16.2)4.6 (1.4–15.1)High-risk group4722 (46.8)29.5 (9.4–92.9)24.5 (7.8–76.9)Pearson χ² = 56.925, P < 0.001Trend χ² testTrend χ² = 50.406, P < 0.001OR odds ratio, CI confidence interval, BMI body mass index, TNM tumor-node-metastasis, χ² chi-square statisticᵃ Adjusted ORs were obtained from multivariable Logistic regression adjusted for age, sex, BMI, and TNM stage. Risk stratification was based on exact Logistic predicted probabilities. The Pearson χ² test was used to compare incidence differences among the three groups, and the trend χ² test was used to evaluate the linear trend between risk category and anastomotic leakage incidence. P values are two-sided Table 7Incidence of anastomotic leakage across risk strataRisk stratumNo. of patientsAnastomotic leakage [n (%)]Crude OR (95% CI)Adjusted ORᵃ (95% CI)Statistic/P valueLow-risk group1384 (2.9)1.00 (reference)1.00 (reference)Intermediate-risk group9813 (13.3)5.1 (1.6–16.2)4.6 (1.4–15.1)High-risk group4722 (46.8)29.5 (9.4–92.9)24.5 (7.8–76.9)Pearson χ² = 56.925, P < 0.001Trend χ² testTrend χ² = 50.406, P < 0.001OR odds ratio, CI confidence interval, BMI body mass index, TNM tumor-node-metastasis, χ² chi-square statisticᵃ Adjusted ORs were obtained from multivariable Logistic regression adjusted for age, sex, BMI, and TNM stage. Risk stratification was based on exact Logistic predicted probabilities. The Pearson χ² test was used to compare incidence differences among the three groups, and the trend χ² test was used to evaluate the linear trend between risk category and anastomotic leakage incidence. P values are two-sided Incidence of anastomotic leakage across risk strata OR odds ratio, CI confidence interval, BMI body mass index, TNM tumor-node-metastasis, χ² chi-square statistic ᵃ Adjusted ORs were obtained from multivariable Logistic regression adjusted for age, sex, BMI, and TNM stage. Risk stratification was based on exact Logistic predicted probabilities. The Pearson χ² test was used to compare incidence differences among the three groups, and the trend χ² test was used to evaluate the linear trend between risk category and anastomotic leakage incidence. P values are two-sided According to model-predicted probability, patients were divided into the low-risk group (n = 138), intermediate-risk group (n = 98), and high-risk group (n = 47). The incidences of anastomotic leakage in the three groups were 2.9% (4/138), 13.3% (13/98), and 46.8% (22/47), respectively, showing a clear risk gradient, with a statistically significant between-group difference (Pearson χ² = 56.925, P < 0.001). The trend χ² test showed a significant positive association between risk stratification and anastomotic leakage incidence (trend χ² = 50.406, P < 0.001). Compared with the low-risk group, the crude OR was 29.5 (95% CI: 9.4–92.9, P < 0.001) in the high-risk group and 5.1 (95% CI: 1.6–16.2, P = 0.005) in the intermediate-risk group. After multivariable adjustment, both the high-risk and intermediate-risk groups still showed higher risks of anastomotic leakage (Table 7). Discussion This study showed that the preoperative CONUT score was associated with an increased risk of clinically relevant anastomotic leakage after radical gastrectomy for gastric cancer. The risk prediction model incorporating CONUT score, diabetes mellitus, total gastrectomy, operative time, and intraoperative blood loss had moderate-to-good discrimination and remained relatively stable in bootstrap internal validation and exploratory temporal validation. The three-tier risk stratification based on predicted probability showed a clear risk gradient, but the model requires external validation before use in broader populations. Nutritional Risk Screening 2002 (NRS-2002) integrates nutritional impairment, disease severity, and age and is suitable for nutritional risk screening in hospitalized patients. Selection of nutritional screening tools for patients with cancer should also consider tumor location, treatment stage, symptom burden, and resource availability [18, 19]. The Mini Nutritional Assessment (MNA) is more commonly used to assess malnutrition risk in older patients, whereas the Patient-Generated Subjective Global Assessment (PG-SGA) integrates weight change, dietary intake, symptom burden, functional status, and physical examination and is widely used for comprehensive nutritional assessment in patients with cancer [20, 21]. Patients with gastric cancer often experience nutritional risk because of tumor-related catabolism, reduced intake, obstruction, perioperative fasting, and treatment-related reactions. Systematic nutritional assessment and individualized nutritional management may improve the quality of perioperative recovery [22, 23]. The advantage of the CONUT score lies in its objectivity, automation, and suitability for retrospective data analysis; however, it cannot directly capture reduced intake, weight loss, gastrointestinal symptoms, muscle mass, or functional status. Therefore, it is more suitable as a screening indicator to trigger further nutritional assessment and should not replace comprehensive nutritional assessment tools [24]. This study positioned a CONUT score ≥ 3 as a sensitivity-prioritized screening threshold rather than a high-risk standard that should directly trigger invasive intervention. This threshold had relatively high sensitivity and negative predictive value and is suitable for preoperative risk screening and excluding low-risk patients. However, its specificity was 62.3%, meaning that approximately one-third of patients without anastomotic leakage could be classified as screening positive. For screening-positive patients, a more appropriate application is intensified nutritional assessment, glycemic management, postoperative monitoring, and early imaging evaluation, rather than direct initiation of high-cost or invasive interventions. If the clinical goal is to reduce false-positive interventions, a CONUT score ≥ 4 may be used as a higher-specificity alternative threshold, at the cost of an increased number of missed cases. Decision curve analysis (DCA) integrates the benefit of true positives and the cost of false positives into net benefit and helps determine whether a prediction model has potential clinical utility at different threshold probabilities [25, 26]. The decision curve in this study showed that the model had positive net benefit within the low-to-moderate threshold probability range of 5%-20%, suggesting that it is more suitable for decision support related to enhanced monitoring and noninvasive preventive measures. Net benefit was limited at higher threshold probabilities, and the clinical value of the model should not be overstated. Single-center retrospective models are susceptible to selection bias, temporal drift, and overfitting. Recent multicenter validation studies and anastomotic leakage cohort studies have suggested that discrimination, calibration, and net benefit should be validated across different regions, surgical teams, and perioperative pathways before model extrapolation [27, 28]. Temporal validation can partially reflect model stability in patients treated during a later time period, but it remains an internal validation method based on single-center data and cannot replace true external validation. Imaging, intraoperative images, and machine learning models provide new technical approaches for predicting anastomotic leakage, but clinical translation still requires evidence regarding sample representativeness, external validation, interpretability, and implementation cost [29, 30]. Compared with complex models, the model in this study consists of routine preoperative laboratory indicators and clinical operative factors and has favorable interpretability and availability; however, model updating and assessment of calibration drift in independent cohorts are still required before clinical implementation. Postoperative complications after gastric cancer surgery are jointly influenced by age, comorbidities, nutritional status, extent of surgery, and perioperative management. Risk assessment models should avoid interpreting nutritional indicators and surgical indicators in isolation [31, 32]. This study found that diabetes mellitus, total gastrectomy, operative time, and intraoperative blood loss were associated with anastomotic leakage, suggesting that nutritional risk should be integrated with metabolic status, procedure complexity, and surgical trauma in perioperative risk assessment. Management of high-risk patients should focus on preoperative nutritional and metabolic optimization, meticulous intraoperative technique, postoperative monitoring of infection and bleeding, and timely intervention; it should not rely solely on a single screening threshold to determine delayed surgery or invasive treatment [33, 34]. Early postoperative bleeding, infection, and nutrition-related complications may all affect anastomotic healing. Perioperative nutritional intervention for patients at nutritional risk may be part of an integrated management strategy, but implementation should be individualized according to patient risk and clinical context [35, 36]. This study has several limitations. First, it was a single-center retrospective study and may have selection and information biases. Second, this study excluded patients who received neoadjuvant therapy and those who received therapeutic enteral or parenteral nutritional support before surgery, with the aim of reducing confounding effects of treatment interventions on the preoperative CONUT score and anastomotic healing risk. Therefore, the findings are mainly applicable to patients who directly undergo radical gastrectomy without preoperative therapeutic nutritional support intervention. In patients receiving neoadjuvant therapy or systematic nutritional optimization, the cutoff values, effect sizes, and calibration of the CONUT score may differ and require further validation in independent cohorts. Third, although bootstrap resampling and temporal splitting were used to assess model stability, both were based on single-center retrospective data and cannot replace independent external validation. The generalizability of the model across different regions, surgical teams, perioperative management strategies, and populations receiving neoadjuvant therapy remains uncertain. Future studies should establish prospective multicenter validation cohorts in collaboration with regional medical centers and evaluate model updating, calibration drift, and clinical net benefit during validation. Conclusions The preoperative CONUT score was associated with an increased risk of clinically relevant anastomotic leakage after gastric cancer surgery. An anastomotic leakage risk prediction model incorporating CONUT score, diabetes mellitus, total gastrectomy, operative time, and intraoperative blood loss may provide a reference for perioperative risk assessment, although multicenter external validation is still required before extrapolated application. Supplementary Information Supplementary Material 1. Supplementary Material 1.