Predicting Anastomotic Leak After Neoadjuvant Chemoimmunotherapy and Esophagectomy: A Dynamic Inflammatory and Nutritional Biomarker Model Authors: Zhenyuan Yang, Yizhou Huang, Maohui Chen, Bingqiang Cai, Hongmu Li, Taidui Zeng, Shuliang Zhang, Chun Chen, Bin Zheng Published: 2026 ABSTRACT AbstractPurposeTo determine the incidence of anastomotic leak (AL) following esophagectomy after neoadjuvant immunochemotherapy (NICT) for esophageal squamous cell carcinoma (ESCC) and to develop a novel prediction model integrating dynamic inflammatory and nutritional biomarkers.Patients and MethodsThis two-center, retrospective study enrolled 550 patients (350 in training cohort, 200 in external validation cohort) with locally advanced ESCC who underwent NICT followed by esophagectomy. For each inflammatory and nutritional index, we calculated the change (Δ) from before NICT to before surgery. Multivariable logistic regression was used to identify independent predictors and construct the prediction model for AL. Model performance was assessed by area under the curve (AUC), calibration, and decision curve analysis.ResultsThe incidence of AL was 10.57% (training) and 11.50% (validation). Most leaks were cervical (93.33%), of Grade II/III severity (78.33%), and nearly half led to anastomotic stenosis (46.67%). The final prediction model incorporated ΔNPR_cat, ΔPAR_cat, ΔHALP_cat and two clinical variables (BMI < 18.5 kg/m2, ECOG performance status ≥ 1). The model demonstrated good discrimination, with an AUC of 0.797 (95% CI: 0.714–0.881) in the training cohort, which was confirmed in the external validation cohort (AUC=0.730, 95% CI: 0.629–0.832). Calibration plots showed good agreement between predicted and observed probabilities.ConclusionAL remains a significant complication after NICT and esophagectomy. We established and validated a practical prediction model that dynamically integrates biomarker trajectories and clinical factors. This tool may help identify high-risk patients preoperatively and thereby support the implementation of tailored perioperative strategies, such as intensified monitoring, proactive nutritional support, or selective anastomotic protection measures.Keywords: esophageal carcinoma, esophagectomy, anastomotic leakage, neoadjuvant immunochemotherapy, prediction model Introduction Esophageal cancer remains a considerable public health burden worldwide, according to the International Agency for Research on Cancer (IARC).1 Anastomotic leak (AL), a severe complication after esophagectomy, is strongly associated with increased perioperative mortality, prolonged hospital stays, compromised quality of life, and poorer long-term outcomes.2–5 Despite continuous refinements in surgical technique and perioperative care, the incidence of AL has not seen a commensurate decline, remaining a persistent and critical issue in clinical practice.6,7 The advent of multimodality therapy has established neoadjuvant therapy as the standard of care for locally advanced esophageal cancer. Nevertheless, evidence suggests that conventional neoadjuvant chemotherapy or chemoradiotherapy may elevate the risk of anastomotic healing-related complications.8–10 Meanwhile, neoadjuvant immunotherapy has recently emerged, yielding remarkable rates of pathological response and improved survival outcomes.11–14 A critical consideration, however, is that the distinct mechanism of action of immune checkpoint inhibitors may perturb the physiological processes of tissue repair through elusive pathways.15 As the population of patients proceeding to surgery after neoadjuvant immunochemotherapy (NICT) grows rapidly, the specific risk profile and true incidence of AL in this context constitute a critical yet uncharacterized clinical challenge, creating a knowledge gap that demands immediate clarification. Current evidence suggests that the pathogenesis of AL involves a complex interplay of patient-related factors, systemic immune-inflammatory responses, and surgical variables.16,17 Early detection of AL is critical for mitigating associated morbidity and mortality. Timely identification enables clinicians to suspend enhanced recovery protocols, initiate further diagnostic investigations, and adjust therapeutic interventions promptly, thereby positively influencing postoperative recovery. In the realm of risk prediction, previous studies have developed models for AL based on various inflammatory and nutritional biomarkers, including the neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), prognostic nutritional index (PNI), combined indices such as PNI-NLR, as well as prealbumin and C-reactive protein (CRP) levels.18–22 However, the majority of existing models rely on static, single-timepoint measurements obtained during the perioperative period. This approach is inherently limited in its capacity to reflect the dynamic physiological evolution of a patient’s status, consequently constraining predictive accuracy. Although individual biomarkers are inherently imperfect, the integration of multiple parameters into composite scoring systems represents a viable strategy to enhance diagnostic performance by improving both sensitivity and specificity.23 Presently, however, predictive models incorporating multiple dynamic biomarkers specifically for patients receiving NICT remain limited. Guided by this context and the identified limitations, our study aims to fulfil two core objectives: firstly, to determine the precise incidence and delineate the associated clinical features of AL in patients with esophageal cancer treated with NICT, and second, to develop and validate a novel, dynamically integrated prediction model for this serious postoperative outcome. This novel model integrates the dynamic trajectories of inflammatory and nutritional biomarkers from pre- to post-treatment with established clinical risk factors, thereby constructing a feasible risk-stratification tool. It is designed to preoperatively identify high-risk patient subgroups warranting targeted optimisation, ultimately providing a reliable foundation for individualised surgical and perioperative management. Materials and Methods Patients and Study Design This was a two-center, retrospective cohort study. We retrospectively reviewed patients with locally advanced esophageal squamous cell carcinoma (ESCC) who received NICT followed by radical esophagectomy at Fujian Medical University Union Hospital and The First Affiliated Hospital of Xiamen University between June 2023 and April 2025. A standard interval of 6 to 8 weeks was observed between the completion of neoadjuvant therapy and surgery for all patients. Patients were included according to the following criteria: (1) clinical diagnosis of locally advanced ESCC (cT2-4a and/or N+ M0, Stage II-IVA) and having received 2–3 cycles of NICT (paclitaxel or nab-paclitaxel combined with a programmed cell death protein 1inhibitor); (2) subsequent underwent radical esophagectomy; and (3) availability of complete perioperative clinical and laboratory data. The exclusion criteria were as follows: (1) any pre-existing condition known to significantly alter systemic inflammatory or nutritional metrics—such as active infections, autoimmune disorders, or severe hepatic/renal dysfunction—prior to the initiation of NICT, to prevent confounding; and (2) incomplete key perioperative clinical or laboratory data, which precluded inclusion in the final analysis. This retrospective study was approved by the Institutional Review Board of Fujian Medical University Union Hospital (No. 2025KY660), which granted a waiver of informed consent for the collection and analysis of medical records due to the retrospective and observational nature of the study. Data Collection and Definitions Clinical data were retrieved from the hospital electronic medical record system. Baseline information included demographics, body mass index (BMI), smoking and alcohol history, Eastern Cooperative Oncology Group (ECOG) performance status, and clinicopathological tumor characteristics, etc. Perioperative data encompassed the surgical approach, operative duration, and lymphadenectomy fields, etc. The calculation of inflammatory and nutritional indicators collected in this study was based on two sets of blood counts and biochemical results obtained prior to neoadjuvant therapy and within one week before surgery. The specific calculation formulae are as follows: Neutrophil-to-Lymphocyte Ratio (NLR) = Neutrophil Count (×109/L)/Lymphocyte Count (×109/L) Neutrophil-to-Lymphocyte Ratio (NLR) = Neutrophil Count (×109/L)/Lymphocyte Count (×109/L) Lymphocyte-to-Monocyte Ratio (LMR) = Lymphocyte Count (×109/L)/Monocyte Count (×109/L) Lymphocyte-to-Monocyte Ratio (LMR) = Lymphocyte Count (×109/L)/Monocyte Count (×109/L) Platelet-to-Lymphocyte Ratio (PLR) = Platelet Count (×109/L)/Lymphocyte Count (×109/L) Platelet-to-Lymphocyte Ratio (PLR) = Platelet Count (×109/L)/Lymphocyte Count (×109/L) Neutrophil-to-Platelet Ratio (NPR) = Neutrophil Count (×109/L)/Platelet Count (×109/L) Neutrophil-to-Platelet Ratio (NPR) = Neutrophil Count (×109/L)/Platelet Count (×109/L) Platelet-to-Albumin Ratio (PAR) = Platelet Count (×109/L)/Albumin (g/L) Platelet-to-Albumin Ratio (PAR) = Platelet Count (×109/L)/Albumin (g/L) Systemic Immune-inflammation Index (SII) = [Platelet Count (×109/L) × Neutrophil Count (×109/L)]/Lymphocyte Count (×109/L) Systemic Immune-inflammation Index (SII) = [Platelet Count (×109/L) × Neutrophil Count (×109/L)]/Lymphocyte Count (×109/L) Systemic Inflammation Response Index (SIRI) = [Neutrophil Count (×109/L) × Monocyte Count (×109/L)]/Lymphocyte Count (×109/L) Systemic Inflammation Response Index (SIRI) = [Neutrophil Count (×109/L) × Monocyte Count (×109/L)]/Lymphocyte Count (×109/L) Prognostic Nutritional Index (PNI) = Serum Albumin (g/L) + 5 × Lymphocyte Count (×109/L) Prognostic Nutritional Index (PNI) = Serum Albumin (g/L) + 5 × Lymphocyte Count (×109/L) Hemoglobin–Albumin–Lymphocyte–Platelet (HALP) Score = [Hemoglobin (g/L) × Albumin (g/L) × Lymphocyte Count (×109/L)]/Platelet Count (×109/L) Hemoglobin–Albumin–Lymphocyte–Platelet (HALP) Score = [Hemoglobin (g/L) × Albumin (g/L) × Lymphocyte Count (×109/L)]/Platelet Count (×109/L) The dynamic change value (Δ) of each indicator, calculated as the relative change from baseline (pre-NICT) to preoperative assessment, was determined by the formula: Δ = (Post-value - Pre-value)/Pre-value With Pre-value and Post-value referring to the measurements obtained before NICT and within one week before surgery, respectively. Data from Fujian Medical University Union Hospital were designated as the training cohort, and those from The First Affiliated Hospital of Xiamen University formed the external validation cohort. The primary study outcome was the incidence of postoperative anastomotic leakage. Accordingly, the study population was dichotomized into an Anastomotic Leakage Group (ALG) and a Non-Leakage Group (NLG) for comparative analysis. For patients with AL, further details on the leak location, healing duration, and the grading of AL were collected. Additionally, comprehensive details on anastomotic strictures—including their incidence, median time to development, dysphagia score—were also documented in both the training and validation cohorts. The grading of AL was performed according to established criteria from the relevant literature.24,25 Briefly, AL were classified as follows: Grade I, subclinical leak identified only on imaging, requiring no therapeutic intervention; Grade II, clinically evident leak manageable with conservative measures such as wound drainage or percutaneous catheter drainage, without the need for surgical revision; Grade III, major anastomotic disruption necessitating surgical revision; and Grade IV, severe conduit necrosis requiring surgical diversion. Dysphagia is scored according to the Mellow scale (also known as Atkinson, Ogilvie, Mellow-Pinkas or European Organisation for Research and Treatment of Cancer and/or Radiation Therapy Oncology Group scale), ranging from no dysphagia (grade 0) to no passage and unable to swallow anything (grade 4).26 Of note, this grading scale is sometimes alternatively presented as a 5-point scale (Grade 1 to 5) in the literature. Statistical Analysis Continuous variables were evaluated for normality; those conforming to a normal distribution are reported as mean ± standard deviation and compared between groups with the Student’s t-test, while non-normally distributed variables are presented as median (interquartile range) and compared using the Mann–Whitney U-test. All categorical data are presented as number (percentage), and intergroup comparisons were performed using the Chi-square test or Fisher’s exact test, whichever was applicable. Statistical analyses were conducted employing R software (version 4.3.1). Statistical significance was defined as a two-tailed p-value < 0.05. The modeling strategy for the training cohort proceeded as follows: First, the diagnostic performance of each Δ indicator was evaluated by generating receiver operating characteristic (ROC) curves. The optimal cutoff value for each indicator was identified by maximizing the Youden index (sensitivity + specificity - 1). Each continuous Δ variable was subsequently dichotomized according to its respective cutoff to create a new categorical variable (Δ_cat). After excluding indicators with high multicollinearity (variance inflation factor ≥ 5), the eligible Δ_cat variables were carried forward. These categorized Δbiomarkers, together with a set of predefined clinical candidate variables—namely age, sex, BMI, smoking and alcohol history, ECOG performance status, and clinical T/N stage—were then entered into univariate logistic regression analysis. Variables with p < 0.1 in the univariate analysis were subsequently included in a multivariable logistic regression model with backward stepwise selection, and those maintaining a stricter threshold of p < 0.05 were retained to construct the final prediction model. The model’s performance was evaluated through ROC curves, calibration plots, and nomograms, with internal validation performed via bootstrapping resampling with 1000 resamples. Finally, the model was applied to the external validation cohort to assess its generalizability and clinical applicability. Results Baseline Patient Characteristics The baseline demographic and clinical profiles of the patients in the training and validation cohorts are summarized in Table 1. The final study population comprised 350 patients in the training cohort and 200 in the validation cohort. Based on the occurrence of postoperative anastomotic leakage, patients were stratified into an anastomotic leak group (ALG; training: n=37, validation: n=23) and a non-leak group (NLG; training: n=313, validation: n=177). Comparative analysis revealed that, in both cohorts, the two groups differed significantly with respect to ECOG status ≥1, BMI < 18.5 kg/m2, and post-NICT levels of neutrophils and hemoglobin. All other recorded baseline characteristics were identified as comparable between the groups. Detailed laboratory values for the training and validation cohorts are presented in Supplementary Table 1. Table 1Baseline Demographic and Clinical Characteristics of Patients in the Training and Validation CohortsVariableTraining CohortsValidation CohortsALG (n=37)NLG (n=313)pALG (n=23)NLG (n=177)pSex, n (%)0.6970.857 Male31 (83.78)254 (81.15)19 (82.61)139 (78.53) Female6 (16.22)59 (18.85)4 (17.39)38 (21.47)Age, years, median (IQR)61.00 (57.00, 70.00)60.00 (56.00, 66.00)0.26666.00 (61.00, 68.00)62.00 (56.00, 67.00)0.117BMI, kg/m2, median (IQR)20.98 (18.49, 23.51)21.48 (19.38, 23.26)0.57922.05 (18.45, 23.58)21.72 (19.53, 24.14)0.768BMI < 18.5 kg/m2, n (%)0.0110.046 No27 (72.97)279 (89.14)16 (69.57)155 (87.57) Yes10 (27.03)34 (10.86)7 (30.43)22 (12.43)ECOG performance status, n (%)0.0020.025 023 (62.16)260 (83.07)14 (60.87)147 (83.05) ≥114 (37.84)53 (16.93)9 (39.13)30 (16.95)Smoking history, n (%)0.0930.372 Never8 (21.62)111 (35.46)12 (52.17)75 (42.37) Current or former29 (78.38)202 (64.54)11 (47.83)102 (57.63)Drinking history, n (%)0.4540.474 Never13 (35.14)130 (41.53)13 (56.52)86 (48.59) Current or former24 (64.86)183 (58.47)10 (43.48)91 (51.41)Hypertension, n (%)0.8980.554 Never30 (81.08)251 (80.19)22 (95.65)158 (89.27) Current or former7 (18.92)62 (19.81)1 (4.35)19 (10.73)Diabetes, n (%)0.5541.000 Never36 (97.30)292 (93.29)22 (95.65)170 (96.05) Current or former1 (2.70)21 (6.71)1 (4.35)7 (3.95)Tumor Location, n (%)0.9500.279 Upper thoracic5 (13.51)42 (13.42)5 (21.74)19 (10.73) Middle thoracic21 (56.76)170 (54.31)13 (56.52)105 (59.32) Lower thoracic11 (29.73)101 (32.27)5 (21.74)53 (29.94)Surgical procedure, n (%)0.8120.601 VATS30 (81.08)239 (76.36)18 (78.26)148 (83.62) RATS6 (16.22)64 (20.45)5 (21.74)28 (15.82) Open1 (2.70)10 (3.19)0 (0.00)1 (0.56)Operation duration, M (Q1, Q3)342.00 (306.00, 400.00)336.00 (306.00, 374.00)0.360347.00 (314.00, 394.50)329.50 (295.00, 368.75)0.172Lymphadenectomy fields, n (%)0.8081.000 2-field32 (86.49)266 (84.98)18 (78.26)139 (78.53) 3-field5 (13.51)47 (15.02)5 (21.74)38 (21.47)ICU, n(%)1.0000.996 No35 (94.59)300 (95.85)21 (91.30)166 (93.79) Yes2 (5.41)13 (4.15)2 (8.70)11 (6.21)cT, n(%)0.5480.704 26 (16.22)72 (23.00)7 (30.43)51 (28.81) 322 (59.46)182 (58.15)13 (56.52)90 (50.85) 49 (24.32)59 (18.85)3 (13.04)36 (20.34)cN, n(%)0.5150.634 06 (16.22)47 (15.02)6 (26.09)30 (16.95) 111 (29.73)120 (38.34)7 (30.43)47 (26.55) 214 (37.84)117 (37.38)6 (26.09)63 (35.59) 36 (16.22)29 (9.27)4 (17.39)37 (20.90)Lymphpositive, n(%)0.9630.093 No19 (51.35)162 (51.76)14 (60.87)75 (42.37) Yes18 (48.65)151 (48.24)9 (39.13)102 (57.63)TRG, n(%)0.5810.233 010 (27.03)61 (19.49)3 (13.04)9 (5.08) 18 (21.62)67 (21.41)3 (13.04)51 (28.81) 210 (27.03)78 (24.92)11 (47.83)72 (40.68) 39 (24.32)107 (34.19)6 (26.09)45 (25.42) Baseline Demographic and Clinical Characteristics of Patients in the Training and Validation Cohorts Characteristics of AL in Patients with NICT The detailed characteristics of the 60 patients who developed AL are summarized in Table 2. Among the total cohort, the vast majority of leaks (56/60, 93.33%) occurred at the cervical anastomosis. According to the established grading criteria, most leaks were classified as Grade II (41/60, 68.33%), followed by Grade I (13/60, 21.67%) and Grade III (6/60, 10.00%); no Grade IV leaks were observed. The median time to AL diagnosis was 7.00 days postoperatively, with a protracted median healing duration of 28.00 days. Subsequently, anastomotic stenosis developed in nearly half of the AL patients (28/60, 46.67%), at a median time of 72.00 days after surgery. Among all AL patients, the dysphagia score at the time of assessment was Grade 0 in 50.00% (30/60), Grade I in 26.67% (16/60), Grade II in 15.00% (9/60), and Grade III in 8.33% (5/60). No statistically significant differences were observed in any of these characteristics between the training and validation cohorts (all p > 0.05). Table 2Characteristics of Anastomotic Leakage in the Training and Validation CohortsVariableTotal (n=60)Training Cohorts (n=37)Validation Cohorts (n=23)pLocation, n (%)1.000 Cervical56 (93.33)35 (94.59)21 (91.30) Intrathoracic4 (6.67)2 (5.41)2 (8.70)Grade, n (%)null I13 (21.67)9 (24.32)4 (17.39) II41 (68.33)25 (67.57)16 (69.57) III6 (10.00)3 (8.11)3 (13.04) IV0 (0.00)0 (0.00)0 (0.00)Time to leak, days, median (range)7.00 (3.00, 24.00)7.00 (3.00, 23.00)7.00 (3.00, 24.00)0.759Healing duration, days, median (range)28.00 (7.00, 70.00)27.00 (8.00, 70.00)28.00 (7.00, 56.00)0.939Anastomotic stenosis, n (%)0.887 No32 (53.33)20 (54.05)12 (52.17) Yes28 (46.67)17 (45.95)11 (47.83)Time to anastomotic stenosis, days, median (range)72.00 (25.00, 108.00)74.00 (25.00, 134.00)71.00 (29.00, 132.00)0.867Dysphagia severity, n (%)Null Grade 0 (No dysphagia)30 (50.00)19 (51.35)11 (47.83) Grade I (Dysphagia to solids)16 (26.67)10 (27.03)6 (26.09) Grade II (Dysphagia to soft solids)9 (15.00)5 (13.51)4 (17.39) Grade III (Dysphagia to liquids)5 (8.33)3 (8.11)2 (8.70) Grade IV (Cannot even swallow saliva)0 (0.00)0 (0.00)0 (0.00) Characteristics of Anastomotic Leakage in the Training and Validation Cohorts Model Construction and Performance Assessment in the Training Cohort The diagnostic characteristics of the dynamic biomarkers are detailed in Table 3, which includes the area under the curve (AUC) for each Δ indicator, the optimal cutoff value identified via the maximum Youden index, and the performance metrics for the derived categorical (Δ_cat) variables. Assessment of multicollinearity (Supplementary Figure 1) revealed that all variables had VIF values below the accepted threshold of 5, indicating no substantial collinearity. Consequently, all candidate predictors were deemed eligible and incorporated into the logistic regression model. As summarized in Table 4, univariate and multivariate logistic regression analyses were performed on the Δ_cat indicators and clinical factors. In the multivariate model, ΔNPR_cat, ΔPAR_cat, ΔHALP_cat, BMI < 18.5, and ECOG ≥ 1 emerged as independent predictors of anastomotic leak (P < 0.05) and were thus retained for the final model construction. The model took the following form: Table 3Diagnostic Performance of Dynamic Inflammatory and Nutritional Indicators for Predicting Anastomotic Leak in the Training CohortIndexAUC (95% CI)Cut OffSensitivitySpecificityContinuous Δ variablesΔNLR0.617 (0.519–0.715)0.6580.3240.882ΔLMR0.505 (0.398–0.612)−0.2170.5140.374ΔPLR0.533 (0.435–0.630)−0.0780.6220.524ΔNPR0.662 (0.574–0.751)0.0750.7840.530ΔPAR0.580 (0.493–0.667)−0.1190.2700.508ΔSII0.541 (0.439–0.643)0.0640.4590.665ΔSIRI0.536 (0.437–0.634)−0.0290.6760.470ΔPNI0.656 (0.568–0.745)−0.0210.1890.492ΔHALP0.606 (0.508–0.704)−0.0950.3780.364Categorized Δ variablesΔNLR_cat0.603 (0.525–0.682)0.3240.882ΔLMR_cat0.556 (0.470–0.642)0.4860.626ΔPLR_cat0.545 (0.460–0.629)0.4050.684ΔNPR_cat0.657 (0.584–0.730)0.7840.530ΔPAR_cat0.611 (0.533–0.689)0.7300.492ΔSII_cat0.562 (0.476–0.648)0.4590.665ΔSIRI_cat0.573 (0.491–0.654)0.6760.470ΔPNI_cat0.659 (0.590–0.729)0.8110.508ΔHALP_cat0.629 (0.545–0.712)0.6220.636 Diagnostic Performance of Dynamic Inflammatory and Nutritional Indicators for Predicting Anastomotic Leak in the Training Cohort Table 4Univariate and Multivariate Analysis of Preoperative Factors in the Final Prediction ModelVariableUnivariableMultivariable 1Multivariable 2pOR (95% CI)pOR (95% CI)pΔNPR ≤0.075 >0.075<0.0013.265 (1.330–8.012)0.0103.655 (1.516–8.813)0.004ΔPAR ≤ −0.119 > −0.1190.0130.256 (0.100–0.655)0.0040.213 (0.084–0.537)0.001ΔHALP ≤ −0.095 > −0.0950.0030.272 (0.105–0.705)0.0070.154 (0.065–0.367)<0.001BMI ≥18.5 <18.50.0072.602 (1.005–6.738)0.0492.433 (0.944–6.270)0.066ECOG <1 ≥10.0033.142 (1.333–7.405)0.0092.646 (1.166–6.004)0.020ΔPNI ≤ −0.021 > −0.021<0.0010.419 (0.154–1.136)0.087Age≥70 <70 ≥700.0062.195 (0.870–5.535)0.096Smoking Never Current or former0.0982.356 (0.916–6.060)0.075ΔNLR ≤0.658 >0.6580.001ΔLMR ≤ −0.217 > −0.2170.186ΔPLR ≤ −0.078 > −0.0780.276ΔSII ≤0.064 >0.0640.138ΔSIRI ≤ −0.029 > −0.0290.097Sex Male Female0.697Drinking Never Current or former0.455Hypertension Never Current or former0.898Diabetes Never Current or former0.360Tumor Location Upper thoracic Middle thoracic0.944 Lower thoracic0.876cT 2 30.440 40.276cN 0 10.537 20.900 30.439Operation duration0.244 Univariate and Multivariate Analysis of Preoperative Factors in the Final Prediction Model Logit(P) = −1.814 + (1.296 × ΔNPR_cat) + (−1.548 × ΔPAR_cat) + (−1.868 × ΔHALP_cat) + (0.889 × [BMI<18.5]) + (0.973 × [ECOG≥1]) Model Validation with Internal and External Testing The model demonstrated strong discriminative ability, achieving an AUC of 0.797 (95% confidence interval: 0.714–0.881). The corresponding calibration plot and clinical prediction nomogram are presented in Figure 1. Internal validation was performed using the bootstrapping method (1000 iterations) to assess the model’s stability. The corresponding Harrell’s concordance index was 0.769. To evaluate its generalizability, the model was tested on the external validation cohort. The optimal cutoff values for the Δ, as previously established in the training cohort (Table 2), were used to generate the corresponding Δ_cat variables in the validation cohort. The ROC curve for the validation cohort is presented in Figure 2. The model demonstrated satisfactory discriminatory power in the external validation set, achieving an AUC of 0.730 (95% CI: 0.629–0.832). Model Performance in Training Cohorts (n = 350). (A) Nomogram. (B) ROC curve. (C) Calibration plot. Variable definitions: ΔNPR_cat: Categorized based on the Δ in neutrophil-to-platelet ratio (NPR) from pre-NICT to before surgery. ΔNPR = (NPR_post – NPR_pre)/NPR_pre. The cutoff value is 0.075. ΔPAR_cat: Categorized based on the Δ in platelet-to-albumin ratio (PAR) from pre-NICT to before surgery. ΔPAR = (PAR_post – PAR_pre)/PAR_pre. The cutoff value is −0.119. ΔHALP_cat: Categorized based on the Δ in the hemoglobin–albumin–lymphocyte–platelet (HALP) score from pre-NICT to before surgery. ΔHALP = (HALP_post – HALP_pre)/HALP_pre. The cutoff value is −0.095. BMI: Body mass index, stratified as < 18.5 kg/m2 versus ≥ 18.5 kg/m2. ECOG PS: Eastern Cooperative Oncology Group performance status, stratified as 0 versus ≥ 1. ROC curve of the Validation cohort (n = 200). Discussion Based on a large-scale multicenter retrospective analysis, this study systematically characterized the spectrum of AL following esophagectomy after NICT. Leveraging the Δ changes in systemic inflammatory and nutritional biomarkers observed during neoadjuvant therapy, we constructed and validated a predictive model for AL risk after esophagectomy. The model exhibited moderate-to-good discriminatory ability and calibration in the training cohort, whereas it preserved acceptable predictive performance upon external validation. This study attempts to address an unmet need within the evolving NICT landscape by offering a novel predictive solution for postoperative AL risk. Our study delineates the clinical profile of AL following NICT, with incidence rates of 10.57% (37/350) and 11.50% (23/200) in the training and validation cohorts, confirming that AL remains a persistent challenge. Over 90% of leaks were cervical, consistent with the prevalent McKeown approach. Most cases were clinically significant: 75.68% and 82.61% required active intervention (Grade II/III). Specifically, Grade II leaks (67.57%; 69.57%) necessitated invasive drainage, while Grade III leaks (8.11%; 13.04%) mandated reoperation. AL healing was protracted (median 28 days), resulting in repeated interventions, prolonged hospitalization, impaired quality of life, and substantial resource utilization. Notably, 45.95% and 47.83% of AL patients developed symptomatic strictures, of whom 47.06% (8/17) and 54.55% (6/11) experienced severe dysphagia (Mellow Grade II/III), drastically limiting oral intake. This high incidence corroborates previous reports that AL is strongly associated with an elevated risk of stricture,27 highlighting a major determinant of long-term functional impairment. These findings establish AL as a key driver of poor short- and long-term outcomes in the NICT era, underscoring the paramount importance of preoperative risk identification for proactive management. Our model integrates Δ biomarker with two key clinical variables: dichotomized BMI (< 18.5 kg/m2) and ECOG performance status (≥ 1). This combination highlights the sustained importance of nutritional and functional status in predicting surgical outcomes after intensive neoadjuvant therapy. The association between ECOG ≥ 1 and AL indicates that a patient’s functional status, even when mildly impaired, is a critical determinant of surgical outcome. In the NICT setting, this metric may capture the net impact of therapy-related decline on physiological reserve. Our findings reaffirm that baseline nutritional status remains a cornerstone of surgical recovery, even in the era of potent modern systemic therapy. Paradoxically, while NICT effectively activates anti-tumor immunity, its associated toxicities may concurrently erode a patient’s physiological reserve, thereby diminishing their tolerance to subsequent surgical trauma. In contrast to the traditional focus on technical and oncological factors, our model broadens the perspective to encompass systemic patient status. This affirms that evaluating overall physiological condition and tolerance provides an essential, complementary dimension to risk stratification in the context of NICT. Departing from reliance on static assessments, our model’s key contribution is its focus on the dynamic physiological narrative unfolding during NICT. A central premise of our model is that NICT profoundly alters the patient’s physiological landscape. The dynamic response to NICT—reflecting the net change in inflammatory and nutritional reserves—may provide a superior gauge of resilience to the stress of esophagectomy and subsequent healing potential compared to any single preoperative assessment. Based on the analytic findings, Δindicators including ΔNPR_cat, ΔPAR_cat, and ΔHALP_cat were independently associated with AL risk, thereby validating our initial hypothesis. For instance, a rise in ΔNPR_cat suggests aggravated systemic inflammation, whereas a drop in ΔHALP_cat integrates the decline of hemoglobin, albumin, and lymphocytes, indicative of compromised oxygen delivery, nutritional depletion, and weakened immune defense. The final model integrated three Δbiomarkers with the clinical parameters of BMI and ECOG performance status, effectively characterizing a high-risk patient phenotype following neoadjuvant therapy. The model demonstrated moderate-to-good discriminatory ability in the training cohort, with an AUC of 0.797 (95% CI: 0.714–0.881), and the calibration curve showed reasonable agreement between predicted and observed probabilities. External validation yielded an AUC of 0.730 (95% CI: 0.629–0.832), confirming stable performance characteristics across cohorts. The moderate decrease in AUC in the validation set is consistent with the expected attenuation during model generalization. This model offers immediate clinical utility in the pre-operative window following NICT. It enables rapid risk stratification via a nomogram, identifying high-risk patients who may benefit from tailored interventions such as scheduled treatment delays for prehabilitation, refined surgical techniques, and enhanced post-operative surveillance, ultimately aiming to proactively reduce AL incidence and optimize resource use. Several prediction models AL have been developed based on inflammatory and nutritional biomarkers; however, direct quantitative comparison is challenging owing to substantial heterogeneity in patient populations, treatment backgrounds, and predictor selection. Wu et al reported an AUC of 0.737 for postoperative NLR and PLR.19 Gao et al focused on postoperative drain amylase and prealbumin for early diagnosis rather than preoperative risk prediction.21 He et al found that the combination of NLR and PNI on postoperative day 3 achieved an AUC of 0.826.18 Azer et al demonstrated a high negative predictive value for postoperative day 8 CRP, yet similarly did not address the neoadjuvant immunotherapy population.22 In contrast to these prior efforts, our model was specifically developed for patients with ESCC receiving NICT. It integrates the Δ biomarkers from pre- to post-neoadjuvant therapy, thereby capturing the evolving trajectory of inflammatory and nutritional status. Furthermore, our model was developed in a large, two-center cohort (n = 550) with an AUC of 0.797 in the training set and was externally validated (AUC 0.730). Regarding our decision to dichotomize continuous variables, we acknowledge that this approach may reduce statistical power and increase the risk of overfitting. The primary rationale for this choice was to maximize clinical accessibility and point-of-care usability: our goal was to provide a bedside risk stratification tool that requires no complex computation and can be directly integrated into a nomogram for individualized risk visualization. This study has several limitations that should be acknowledged. First, the retrospective design inherently carries risks of selection bias and unmeasured confounding. Second, although external validation was performed, both the training and validation cohorts were derived from two institutions within the same geographic region, which may limit the generalizability of our findings to broader populations with different demographic, dietary, or clinical practice patterns. Multi-regional or prospective validation studies are warranted to further establish the model’s applicability. Third, the limited number of AL events, attributable to the inherently low incidence of anastomotic leak, may constrain model robustness in the absence of multicenter data. Furthermore, the dynamic biomarkers included here represent only a fraction of possible predictors. Future investigations could explore integrating multidimensional data—such as radiomic features or specific inflammatory cytokines—to further improve the model’s predictive accuracy and clinical applicability. Conclusion Anastomotic leak remains a formidable challenge in patients undergoing esophagectomy after neoadjuvant chemoimmunotherapy, profoundly compromising both short- and long-term recovery. We have established and validated a predictive model incorporating dynamic biomarker profiles and clinical factors that effectively stratifies anastomotic leakage risk in patients undergoing esophagectomy after neoadjuvant chemoimmunotherapy. Following independent external validation, this model indicates promising calibration and potential clinical applicability, although further prospective evaluation is warranted.