This research is a retrospective observational study conducted at a single center. Between January 2022 and December 2024, we included 781 patients who had radical resection for EC at Shandong University’s Esophageal Group in the Department of Thoracic Surgery at Qilu Hospital.

The following standards were used to determine which patients were included:

Criteria for Inclusion: (1) Primary EC at postoperative assessment with pathological confirmation; (2) Elective esophageal resection with the goal of curing the disease.

Exclusion criteria include: (1) preoperative distant metastases; (2) emergency surgery; (3) a severe or active pulmonary infection before surgery; (4) a significant clinical documentation gap; and (5) numerous primary malignancies that are concurrent or metachronous.

The Ethics Review Committee of Qilu Hospital, which is affiliated with Shandong University, gave its approval to this research protocol. The Declaration of Helsinki’s ethical guidelines are closely followed in this investigation. The Qilu Ethics Committees approved the investigation (registration number: KYLL-202008-023-1). The Ethics Committee approved an exemption from seeking patients’ individual informed permission due to the study’s retrospective nature. Figure 1 depicts the comprehensive patient selection procedure.

Upon admission, all patients go through a thorough preoperative evaluation that includes imaging studies, cardiopulmonary function testing, and a thorough physical examination to rule out distant metastases and finally validate surgical indications. The hospital’s multidisciplinary team must jointly assess all cases that match surgical criteria in order to identify the best course of therapy. Individualized treatment plans will be created by oncology specialists for patients who need neoadjuvant therapy. These could consist of combination immunotherapy, chemoradiation, or chemotherapy.

Double-lumen endotracheal intubation was used for the procedure, which was conducted under general anesthesia. During preoperative planning, the surgical team decides the precise surgical approach depending on the patient’s unique situation and the location of the tumor. Surgical techniques include open and minimally invasive procedures, and access points include incisions in the left or right thorax. Standard two-field dissection, extended two-field dissection, and even three-field dissection are all possible for lymph nodes. The stomach is the preferred esophageal replacement organ during surgery, with the colon or jejunum coming in second or third. The esophageal bed or retrosternal approaches might be used to raise the organ. The neck or the thoracic cavity contain the anastomosis site. Either manual suturing or mechanical stapling are used to perform anastomosis. During surgery, a jejunostomy is performed if the placement of a nasojejunal feeding tube is not effective.

All patients were moved to the surgical critical care unit for one to three days of rigorous monitoring after surgery. On the first postoperative day, enteral nutrition and parenteral support were started, along with physical interventions such back and rotating percussion to encourage the removal of sputum. Patients were returned to the general ward once vital signs had stabilized. Oral feeding is usually started on the seventh postoperative day if anastomotic leaking from the upper gastrointestinal tract has been ruled out. Patients usually heal and are released 10–14 days after surgery if there are no problems such chylothorax, anastomotic leaking, or incisional infection. Patients diagnosed with intraoperative pleural metastasis undergo extended pleurectomy according to institutional guidelines, followed by adjuvant therapy (chemotherapy or immunotherapy), and are closely monitored during follow-up.

Through a retrospective analysis of the electronic medical record system, the following four categories of perioperative variables were methodically gathered for this study:

Pleural metastasis is defined as: macroscopically visible nodules or diffuse implantations on the parietal/visceral pleura observed by the primary surgeon during open thoracotomy or thoracoscopic surgery, confirmed by intraoperative frozen section pathology or definitive postoperative paraffin section pathology. Postoperative pulmonary complications include pneumonia, respiratory failure, acute respiratory distress syndrome (ARDS), clinically significant atelectasis, and massive pleural effusion. The clinical diagnosis of pneumonia is established when chest imaging shows new or increasing pulmonary infiltrates and at least two of the three criteria listed below are present: (1) body temperature >38 °C; (2) purulent airway secretions; (3) peripheral blood white blood cell count >10.0×10⁹/L or <4.0×10⁹/L (35). Respiratory failure is defined as a PaO₂/FiO₂ ratio ≤200 mmHg or the need for mechanical ventilation (36). Acute respiratory distress syndrome (ARDS) is diagnosed according to the Berlin criteria (37). Atelectasis is defined as the presence of unilateral or bilateral focal consolidation, indicating focal loss of ventilation. Atelectasis is not included in the case criteria for ARDS, pneumonia, or fluid overload (38). A massive pleural effusion is characterized by respiratory impairment that necessitates closed chest drainage or therapeutic thoracentesis (39).

The number of cigarettes smoked daily times the number of years smoked (cigarette-years) is the smoking index. When patients at nutritional risk get specialist nutritional care (such as enteral or parenteral nutrition or oral nutritional supplements) for at least seven days before surgery, this is referred to as preoperative nutritional intervention.

The following formula is used to determine the derived inflammatory markers:

This study’s statistical analyses were all conducted using R software (version 4.3.1). Depending on their normality, continuous variables were reported as mean ± standard deviation or median (interquartile range), and the Student’s t-test or Mann-Whitney U test were used for comparison. Categorical variables were presented as frequency (percentage) and compared using chi-square test or Fisher’s exact test.

Patients were divided into training and validation sets at random in a 7:3 ratio. The purpose of this split was to make sure that the risk factors that were found were robust; variables were found in the training set, and their stability was evaluated in the independent validation set. The primary objective was to identify factors independently associated with PPCs. Variables with p-values < 0.05 in univariate analysis were included in the multivariate logistic regression model. Forward stepwise selection was subsequently employed to obtain the final simplified model, thereby identifying independent factors. Results are presented as odds ratios (OR) with their 95% confidence intervals (CI).

Based on the final set of independent factors found in the primary study, a nomogram was created for exploratory reasons. The area under the receiver operating characteristic curve (AUC) and the optimism-corrected C-index (obtained via 1000 bootstrap resamples) were used to evaluate its discriminative capacity. The Hosmer-Lemeshow goodness-of-fit test and calibration plots were used to assess calibration. Decision curve analysis (DCA) was used to investigate clinical utility. SHapley Additive exPlanations (SHAP) analysis was used to display each factor’s contribution and direction of influence in the exploratory model.

In this study, 781 individuals who had radical EC resection were included. 91 patients (11.7%) experienced PPCs (PPCs group), whereas 690 individuals (88.3%) did not (non-PPCs group). Table 1 provides a detailed comparison of the baseline characteristics of the two groups. Overall, PPCs were associated with significantly lower nutritional immunological indicators (LMR, PNI, albumin-globulin ratio), and significantly higher systemic inflammatory markers (neutrophil count, NLR, dNLR, SIRI). In comparison to the group without PPCs, this group also showed noticeably greater rates of anastomotic fistula formation, postoperative ICU admission, and intraoperative pleural metastases. Age, gender, tumor stage, and histological type were among the baseline parameters that did not show statistically significant variations between the two groups.

This study used random sampling to split all 781 patients into a training set (n=549) and a validation set (n=232), with a sample size ratio of roughly 7:3 between the two groups, in order to ensure the robustness of the identified risk factors. The training set and validation set comparison is displayed in Table 2. The two sets were found to have excellent balance across all key variables, including demographic data, clinical features, tumor pathological parameters, surgery-related variables, laboratory indicators, and postoperative complication rates (all P > 0.05). This balanced distribution validates the suitability of the random split and offers a solid foundation for the factor identification and stability evaluation to follow.

Eleven variables were shown to be substantially linked to PPCs in patients with EC in the univariate logistic regression analysis (Table 3). Risk factors included neutrophil count, eosinophil count, monocyte count, dNLR, PLR, SII, extended lymph node dissection, intraoperative pleural metastasis detection, postoperative ICU admission, and anastomotic fistula formation; LMR showed protective effects. There was no statistical significance in the remaining clinical and pathological factors.

Five independent predictors of PPCs were finally found using forward stepwise multiple logistic regression analysis. The eosinophil count (OR = 5.924, p=0.027), intraoperative pleural metastasis (OR = 6.853, p=0.026), postoperative ICU admission(OR = 6.963, p=0.006), and postoperative anastomotic leakage (OR = 13.454, p=0.000) were found to be significant risk factors, while the LMR (OR = 0.791, p=0.021) was a protective factor. The five independent predictors found by the multivariate logistic regression are shown graphically in Figure 2’s forest plot. Table 4 displays the multivariate logistic regression analysis’s findings.

Based on the five independent factors found in the initial multivariable analysis, a nomogram was created for exploratory and illustrative reasons. The purpose of this visualization tool is to show how these elements work together, but it is not meant to be the main prediction model used in clinical settings. Among these, anastomotic fistula, postoperative ICU admission and pleural metastases are binary variables (0 denotes absence, 1 denotes presence). R statistical software was used to create a nomogram for PPCs in patients with EC based on the aforementioned algorithm. A simple nomogram is shown in Figure 3A, and a stylized nomogram is shown in Figure 3B.

Additionally, we performed multicollinearity diagnostics (Table 5). All of the variables in the final model had variance inflation factors (VIFs) of less than 1.02, which guaranteed the stability of the model’s findings and showed no substantial multicollinearity among the variables.

To evaluate the explanatory power and robustness of the identified risk factors, this study uses ROC curves, bias corrected C-index and radar charts to assess the predictive model’s performance. Figure 4A shows the ROC curve. The model exhibits moderate but stable discriminatory capability, as shown by the blue curve representing the training set (AUC = 0.665, 95% CI 0.585–0.745) and the pink curve representing the test set (AUC = 0.561, 95% CI 0.446–0.676). The model shows consistent performance across datasets, despite the AUC values reflecting the inherent difficulties in predicting PPCs simply based on preoperative and intraoperative parameters. The predictive model reaches its ideal cutoff value when specificity is 0.967 and sensitivity is 0.317 on the training set, as Figure 4B illustrates. The Youden index at this point is 0.185. The detailed results of the ROC curve are listed in Table 6 for reference. We used the Bootstrap approach (with 1000 resamples) to compute the bias corrected C-index. The model obtained a bias corrected C-index of 0.666 on the training set and 0.557 on the independent validation set, according to the results. A radar chart showing the model’s overall performance across several metrics on the training and test datasets is shown in Figure 5. The model retains strong and consistent classification capabilities across many datasets with no indications of overfitting, as shown by the remarkably identical graphical profiles covering large areas.

The model shows satisfactory calibration following correction, as seen in Figure 6, where the calibrated curve nearly resembles the ideal diagonal line. We used the Hosmer-Lemeshow goodness-of-fit test. The test did not produce any statistically significant results in either the validation set (p = 0.628) or the training set (p = 0.371). This shows that, over the whole risk spectrum, the estimated risks derived from these inflammatory and surgical factors show great consistency with actual observed probabilities, demonstrating that the model accurately reflects the correlation between the identified factors and clinical outcomes.

This study used decision curve analysis (DCA) and clinical impact curves (CIC) to assess the clinical relevance of the risk model. In the DCA of the training set (Figure 7A) and validation set (Figure 7B), the model’s net benefit curve showed positive clinical net benefit throughout a wide range of threshold probabilities (about 0.1 to 0.8). Both the number of people the model identified as high risk and the number of people who experienced actual events show a steady decline as the high-risk threshold varies in the CIC of Figures 8A, B. The proximity of the curves validates that incorporating Eosinophils, LMR, Intraoperative Pleural Metastasis, Postoperative ICU Admission and Anastomotic Leakage into risk stratification offers tangible clinical value, allowing clinicians to identify vulnerable patients who may benefit from intensified perioperative management.

The SHAP analysis variable importance plot in Figure 9 illustrates the magnitude of association between each risk factor and PPCs within the multivariable risk analysis. The five risk factors’ relative contributions to the risk probability are reflected in the SHAP values displayed in the figure. The most important predictor, according to analysis, is “LMR”. “Eosinophils count”, “Anastomotic Leakage”, “Postoperative ICU Admission” and “Pleural Metastasis” come next. The inflammatory marker “LMR” showed a notable protective effect among these, while “Eosinophils count” demonstrated a strong positive association with increased risk.

The SHAP dependency plot, which consists of five subplots examining five different indicators, is shown in Figure 10. The horizontal axis in each subplot represents the value range of the related indicator, and the vertical axis is the SHAP values. The biological and clinical relationships between risk factors and SHAP values are revealed using data distribution visualization, which makes it evident how each indicator modulates the risk of PPCs across a range of values. More significantly, it shows that the model views these inflammatory markers and surgical complications as interacting and amplifying one another rather than acting independently, ultimately determining the overall clinical risk profile.

The logistic curves for three continuous variables— eosinophils count, and LMR—are shown in (Figures 11A, B). In keeping with the typical features of a logistic function, these curves consistently show a large nonlinear effect of each variable on the response probability. They make evident the patterns of correlation between many factors and the likelihood of PPCs occurring.

To examine the relationship between continuous variables and PPCs, (Figures 12A, B) shows the RCS curves for eosinophil count and LMR, respectively. The findings show that eosinophil count show significant overall associations with outcomes (overall P < 0.05). There is no evidence of significant nonlinear trends (P−non−linear > 0.05), indicating effects closer to a linear pattern. However, the association between LMR and PPCs was not statistically significant (P−overall = 0.096, P−non−linear = 0.925).