A retrospective study design was adopted, and 245 patients who underwent surgery for spinal tumors at our hospital between January 2024 and March 2026 were selected as study participants. The inclusion criteria were as follows: 1) pathological diagnosis of primary or metastatic spinal tumor; 2) first-time resection of the tumor lesion with or without spinal internal fixation; 3) complete medical records, preoperative laboratory test results, surgical records, and postoperative follow-up data; 4) postoperative follow-up of ≥30 days, allowing definitive assessment of pulmonary complications; and 5) informed consent provided by the patients and their family members. This study was a single-center retrospective case study. All data were obtained from previously archived electronic medical records of the hospital, no intervention was performed on patients during the study, and all personal information was anonymized and de-identified, making it impossible to trace the data to specific individuals. This study was approved by the ethics committee of our hospital, with a waiver of informed consent. The exclusion criteria were as follows: 1) active pulmonary infection or acute exacerbation of chronic obstructive pulmonary disease before surgery; 2) concomitant end-stage multiple organ failure; and 3) preoperative requirement for mechanical ventilation. Patients were divided according to the occurrence of pulmonary complications within 30 days after surgery into a complication group (34 cases) and a non-complication group (211 cases). During the study period, patients who underwent surgery for spinal tumors at our hospital were screened for eligibility. After application of the exclusion criteria (active pulmonary infection or acute exacerbation of chronic obstructive pulmonary disease, concomitant end-stage multiple organ failure, or preoperative requirement for mechanical ventilation), and after exclusion of patients with incomplete medical records or loss to follow-up, 245 patients were ultimately included in the analysis. All 245 included patients had complete data for the variables analyzed. The number of patients excluded for each specific criterion was not systematically recorded in the retrospective database; this limitation, together with the requirement for complete records, is acknowledged as a potential source of selection bias.

With reference to relevant literature [3], postoperative pulmonary complications were defined as the occurrence of any of the following within 30 days after surgery: 1) pneumonia, characterized by fever, abnormal leukocyte count, and purulent airway secretions accompanied by newly developed infiltrates on imaging; 2) atelectasis, indicated by segmental or lobar collapse on imaging accompanied by hypoxemia; 3) respiratory failure, defined as an arterial oxygen partial pressure ≤ 60 mmHg or the need for reintubation; 4) moderate-to-large pleural effusion requiring puncture and drainage; and 5) pneumothorax requiring closed thoracic drainage. During hospitalization, PPCs were identified through daily review of medical records, nursing notes, chest radiography and computed tomography reports, and arterial blood gas results by the attending surgical team; after discharge, all patients were followed up at the outpatient clinic or by structured telephone interview at 30 days postoperatively. All suspected PPCs were independently adjudicated by two investigators (an anesthesiologist and an orthopedic surgeon) according to the predefined criteria; disagreements were resolved by consensus discussion with a senior physician. Patients who developed more than one type of PPC were counted as a single event in the composite outcome, and the most clinically severe complication was recorded for descriptive purposes. Nineteen preoperative and intraoperative clinical variables were collected as candidate predictors: preoperative variables included age, sex, BMI, albumin, hemoglobin, ADL, NRS2002, smoking history, pain, number of chronic comorbidities, tumor classification, preoperative interventional embolization, and history of radiotherapy or chemotherapy; intraoperative variables included surgical site, surgical approach, pathological fracture, operative duration, blood loss, and transfusion volume. Surgical procedure type (modus operandi) was categorized as conventional surgery (including posterior decompression, curettage, and intralesional/piecemeal resection; reference group) versus en bloc resection (including total spondylectomy and combined anterior-posterior radical resection). Intraoperative allogeneic red blood cell transfusion volume was categorized as no transfusion (0 mL; reference group; n = 73, 6 PPCs), limited transfusion (<800 mL; n = 121, 12 PPCs), and massive transfusion (≥800 mL; n = 51, 16 PPCs). Postoperative drainage status was recorded as a descriptive clinical variable but was not included as a candidate predictor because the model was intended for perioperative risk assessment before the postoperative course unfolded.

Data analysis was performed using SPSS 26.0 and R 4.3.1. Normally distributed continuous variables were described as means ± standard deviations, and between-group comparisons were conducted using independent-samples t-tests. Categorical variables were presented as counts and percentages, and between-group differences were assessed using the χ² test or Fisher’s exact test. Variables with P < 0.05 in univariate analysis were included in a multivariable binary logistic regression model, and model fit was evaluated using the likelihood ratio test. A random forest prediction model was constructed based on the variables selected through regression analysis, using the random Forest package in R. The model was set to grow 500 decision trees with mtry = 1 (the default for three candidate predictors); no additional hyper-parameter tuning was performed. Given the imbalance between complication and non-complication cases, no specific class-imbalance adjustment (e.g., class weighting or resampling) was applied. Model stability was assessed using stratified five-fold cross-validation. Within each fold, the random forest was trained on four folds and evaluated on the held-out fold; the optimal cutoff value was determined using the Youden index on the training folds and then applied to the held-out fold. Of note, variable selection based on univariate and multivariable logistic regression was performed on the entire cohort before cross-validation rather than being nested within each fold; this may introduce a degree of optimistic bias and is acknowledged as a limitation. Discriminative performance was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC). Calibration was assessed using the calibration intercept, calibration slope, and Brier score. For comparison, a multivariable logistic regression model using the same three predictors was also evaluated under the same cross-validation framework. From the optimal cutoff, sensitivity, specificity, accuracy, and the F1 score were calculated. A two-sided P < 0.05 was considered statistically significant.

Among the 245 patients, 34 developed postoperative pulmonary complications, yielding an incidence of 13.88% (34/245). Univariate analysis of 19 variables showed statistically significant differences for BMI (t-test, P = 0.045), operative duration (t-test, P = 0.0162), surgical approach (Fisher’s exact test, P < 0.001), and intraoperative transfusion volume (χ² test, P < 0.001). In contrast, age, albumin, hemoglobin, sex, ADL, NRS2002, smoking history, pain, number of chronic comorbidities, tumor classification, interventional embolization, chemoradiotherapy, surgical site, pathological fracture, and intraoperative blood loss all had P-values > 0.05. Because operative duration was collinear with surgical approach and lacked independent predictive value, it was excluded. Ultimately, BMI, surgical approach, and intraoperative transfusion volume were included in the multivariate analysis. See Table 1. Postoperative drainage status was not included as a candidate predictor, as specified in Section 2.2.

Three variables were included in the regression model, and the overall likelihood ratio test was statistically significant (χ² = 18.52, P = 0.00015, pseudo-R² = 0.1146). After adjustment, surgical approach was an independent risk factor (β = 1.3786, OR = 3.969, 95% CI: 1.165 - 13.526, P = 0.0275), indicating that patients undergoing en bloc resection had 3.969 times the odds of PPCs compared with those undergoing conventional surgery (reference group). The PPC event rates were 10/21 (47.6%) in the en bloc resection group and 24/224 (10.7%) in the conventional surgery group. BMI showed a protective trend (OR = 0.906, 95% CI: 0.815 - 1.008, P = 0.071), with each 1-unit increase in BMI associated with an approximately 9.4% reduction in risk. Massive transfusion showed a trend toward increased risk (OR = 2.774, 95% CI: 0.838 - 9.183, P = 0.095), whereas limited transfusion was not associated with risk (OR = 1.185, 95% CI: 0.421 - 3.341, P = 0.748), both compared with the no-transfusion reference group. See Table 2.

Table 1. Univariate analysis of postoperative pulmonary complications.

Table 2. Results of multivariable binary logistic regression analysis.

A random forest model was constructed using three variables. In the modeling cohort, the area under the ROC curve (AUC) was 0.6752, with an optimal cutoff value of 0.100, sensitivity of 100%, and specificity of 42.86%. After five-fold cross-validation, the AUC was 0.605. The optimal cutoff value determined by the Youden index was 0.360, at which the sensitivity was 57.14%, the specificity was 85.71%, the overall accuracy was 81.63%, and the F1 score was 0.4706. The cross-validated calibration intercept was −1.41 and the calibration slope was 0.08, with a Brier score of 0.152, indicating poor calibration with systematic overestimation of risk. For comparison, the multivariable logistic regression model built from the same three predictors achieved a cross-validated AUC of 0.663, a Brier score of 0.108, a calibration intercept of −0.03, and a calibration slope of 0.98, indicating substantially better calibration than the random forest model. The model showed better discrimination for patients without complications (high specificity), but had a relatively high missed-diagnosis rate; therefore, it is recommended as an auxiliary tool for clinical assessment. See Table 3.

Table 3. Diagnostic performance of the random forest prediction model.

Among the 245 patients in this cohort, 34 developed PPCs within 30 days after surgery, corresponding to an incidence of 13.88%. This rate was markedly higher than the 4.3% - 5.7% reported for surgery for common degenerative spinal diseases [5], but was comparable to data from studies specifically involving patients with spinal tumors (11.2% - 15.8%) [1]. This may be explained by the fact that patients with spinal tumors are often bedridden for prolonged periods because of tumor compression and therefore have reduced pulmonary functional reserve. In addition, surgery typically requires extensive soft-tissue dissection, resection of the affected vertebrae, and internal fixation and reconstruction, resulting in substantial trauma and prolonged anesthesia; massive intraoperative blood loss and transfusion are also common. Together, these factors increase the risk of pulmonary complications [4]. Moreover, thoracic spinal tumor surgery often requires transthoracic or extrapleural approaches, which directly disturb the intrathoracic environment, leading to higher incidences of postoperative atelectasis, pleural effusion, and pneumothorax than in lumbosacral surgery [6]. The present data indicate that the burden of PPCs after spinal tumor surgery should not be underestimated. Clinically, such patients should be prioritized for perioperative lung-protective management, including comprehensive preoperative pulmonary function assessment, optimized intraoperative ventilation, and enhanced postoperative airway care and early mobilization.

After adjustment for confounding factors, multivariable logistic regression showed that the surgical approach was an independent risk factor for PPCs (OR = 3.969, P = 0.0275). Similar conclusions were reached in the nomogram for PPCs after spinal tumor surgery developed by Zou et al. [1] and in the multicenter survey by Yan et al. [2]. From a pathophysiological perspective, more invasive procedures, such as combined anterior-posterior approaches and total spondylectomy, require extensive dissection of the paraspinal muscles and prolonged retraction of the pleura and diaphragm, which not only restricts intraoperative ventilation but also increases the likelihood of pulmonary contusion. Prolonged operative time may also lead to anesthetic accumulation and diaphragmatic inhibition, weakening postoperative cough and sputum clearance and thereby predisposing patients to atelectasis and pulmonary infection [3]. The anterior transthoracic approach to the thoracic spine may also directly injure and rupture the pleura, causing pulmonary contusion and pleural effusion; the reported incidence of PPCs for this type of procedure is as high as 14.1% - 29.4% [6]. Therefore, for patients scheduled to undergo highly invasive procedures, preoperative pulmonary function testing and respiratory training should be implemented, intraoperative ventilation strategies should be optimized and the duration of one-lung ventilation should be minimized, and postoperative airway management and early rehabilitation exercises should be intensified.

In this dataset, BMI showed a potential protective tendency (OR = 0.906, P = 0.071), indicating that each 1-unit increase in BMI was associated with an approximately 9.4% reduction in the risk of PPCs. Traditionally, obesity, reflected by a high BMI, is generally regarded as an adverse factor for postoperative pulmonary complications because obese patients have reduced chest wall compliance and decreased functional residual capacity, thereby increasing the likelihood of postoperative atelectasis [7]. However, among patients with spinal tumors, a low BMI often indicates malnutrition and tumor-related wasting; these patients may have insufficient respiratory muscle strength and impaired immunity, which may instead increase the risk of postoperative pulmonary infection [8]. The prediction model developed by Zou et al. [1] also incorporated BMI and identified low BMI as a risk factor. It should be noted that, in the present study, the P-value for BMI was close to 0.05, and the confidence interval slightly crossed 1; therefore, this conclusion requires confirmation in studies with larger sample sizes. In clinical practice, for patients with low BMI, preoperative nutritional intervention should be emphasized to improve respiratory muscle function and overall immune status.

Massive intraoperative transfusion showed a trend toward increased risk in this cohort (OR = 2.774, P = 0.095), whereas no clear association was observed between low-volume transfusion and complications (P = 0.748). Several previous studies have reported an association between intraoperative allogeneic blood transfusion and postoperative infectious complications, as well as pulmonary complications, after spinal surgery [9]. A study based on an interpretable machine learning model likewise identified transfusion as a key predictor of postoperative pneumonia [10]. Mechanistically, this association may be related to transfusion-related acute lung injury (TRALI), transfusion-associated circulatory overload (TACO), and immunosuppression induced by allogeneic blood products [11]. In the present cohort, massive transfusion did not reach statistical significance, which may be attributable to the limited number of positive events (only 34 cases) and the relatively low overall statistical power. Therefore, perioperative blood conservation should receive adequate attention in clinical practice. Strategies such as preoperative correction of anemia, intraoperative controlled hypotension, and cell salvage autologous transfusion may reduce the use of allogeneic blood and thereby lower the risk of pulmonary complications.

Based on the three variables identified by multivariable logistic regression, this study developed a random forest prediction model. The model achieved an AUC of 0.6752 in the development cohort, which decreased to 0.605 after internal cross-validation, indicating room for improvement in discriminatory performance. This performance was inferior to the C-index of 0.755 reported for the nomogram by Zou et al. [1] and lower than the AUC levels of high-quality PPC models summarized by Huang et al. [3]. Notably, a multivariable logistic regression model using the same three predictors achieved a slightly higher cross-validated AUC (0.663 vs. 0.605) and substantially better calibration (Brier score 0.108 vs. 0.152; calibration slope 0.98 vs. 0.08) than the random forest, suggesting that the logistic model may be preferable in this small-sample setting. This may be attributable to three main factors. First, the sample size was limited (245 patients), with only 34 positive events; constrained by the events-per-variable ratio, random forests are prone to overfitting in small samples. Second, the included predictors were routine clinical indicators and did not include more predictive specialty-specific variables such as pulmonary function measures (FEV1 and FVC), smoking pack-years, and ASA classification [3]. Third, spinal tumors are highly heterogeneous, and the mechanisms underlying PPCs may differ between primary and metastatic tumors and across tumor locations; the absence of subgroup analyses may also have reduced predictive accuracy. In recent years, machine learning has been increasingly applied to the prediction of surgical complications, and random forests have attracted attention because of their ability to capture nonlinear relationships and interaction effects [12]. However, in settings with small samples and low event rates, they may not necessarily outperform traditional logistic regression [13]. Future work should expand the sample size through multicenter collaboration, incorporate specialty-specific indicators such as pulmonary function, and introduce SHAP values to enhance interpretability, thereby further optimizing the prediction tool.