Development of a machine learning model for preoperative prediction of spread through air spaces in resectable 
non‑small cell lung cancer: A single‑center retrospective study

Introduction

Lung cancer remains the leading cause of cancer-related mortality worldwide, accounting for 18.7% of all cancer deaths, with non-small cell lung cancer (NSCLC) comprising ~85% of cases (1). For patients with early-stage NSCLC, surgical resection offers the best chance for long-term survival (2). However, recurrence after curative surgery remains a major challenge (3). Emerging evidence highlights the prognostic importance of a pathological feature known as spread through air spaces (STAS), a pattern of invasion where tumor cells disseminate into alveolar spaces beyond the tumor margin, occurring in ~50% of resected NSCLC cases (4,5). The presence of STAS has been associated with early recurrence and poor survival, with 5-year recurrence-free and overall survival rates nearly halved compared with STAS-negative cases, particularly in patients undergoing limited resections (6). This suggests that lobectomy or wider surgical margins may be more appropriate for patients with STAS to reduce the risk of recurrence. However, its identification currently relies entirely on postoperative pathological examination of resected specimens, which remains the standard for STAS diagnosis (7). This creates a key gap in preoperative planning, as STAS status is unknown when deciding surgical extent or resection margin.

Due to this limitation, accurately predicting STAS preoperatively would aid thoracic surgeons in selecting surgical strategies, planning lymph node dissection and counseling patients about recurrence risks. Although several studies (8,9) have developed prediction models for STAS, most of them rely on radiomics features extracted from high-resolution CT images, which necessitate advanced image segmentation, texture analysis and specialized computational tools. These requirements limit the feasibility of such models in routine clinical workflows, particularly in community hospitals or resource-limited settings. Therefore, there is a clear need for accessible and generalizable models that use only routinely available clinical features.

Artificial intelligence, particularly in machine learning (ML), has created novel opportunities to improve predictive modeling in clinical research. Unlike traditional statistical methods, ML algorithms can capture complex and non-linear interactions among high-dimensional variables and have demonstrated notably increased performance in a range of medical applications, including pulmonary nodule characterization and lung cancer risk prediction (10–12). Despite these advantages, the limited interpretability of ML models has hindered their widespread clinical adoption. To address this concern, explainable artificial intelligence techniques such as SHapley Additive exPlanations (SHAP) have been introduced to improve model transparency. SHAP assigns interpretable contribution scores to individual features, offering both global insights into overall feature importance and local explanations for individual predictions (13,14).

The present study aimed to develop a practical and explainable ML model to predict STAS using only standard preoperative data, without relying on radiomics. The present study trained and compared five supervised ML algorithms based on routinely available clinical, radiological and serological variables to identify the optimal predictive model. Furthermore, SHAP analysis was applied to interpret feature contributions and visualize the decision-making process of the model. This explainable, data-driven model is intended to support personalized surgical planning and potentially improve preoperative risk stratification in patients with NSCLC in the future.

Patients and methods

Study population

The present study was a retrospective, single-center study that included patients aged ≥18 years who underwent curative-intent pulmonary resection for NSCLC at Anqing Municipal Hospital (Anqing, China) between January 2021 and December 2023. Patients were identified using the electronic medical records.

Inclusion criteria were as follows: i) Histopathologically confirmed NSCLC after surgical resection; and ii) resectable NSCLC staged retrospectively according to the 9th edition of the TNM classification (International Association for the Study of Lung Cancer, 2024) (15), which was applied during data analysis to ensure consistency with the most recent staging standards. The re-staging process was performed based on available pathological and imaging data by two thoracic surgeons to ensure accuracy and reproducibility. Exclusion criteria were as follows: i) Receipt of neoadjuvant chemotherapy, radiotherapy, immunotherapy or targeted therapy prior to surgery; ii) history of other malignancies; iii) multiple primary tumors in the same lobe of the lung; and iv) incomplete clinical, radiological or pathological data.

The present study was approved by the Ethics Committee of Anqing Municipal Hospital (approval no. 2025-152). The committee confirmed that the study design met ethical standards and that written informed consent was obtained from all participants prior to inclusion. No formal sample-size calculation was performed due to the retrospective nature of the present study and all eligible patients within the study period were included. Patients with missing data on any included variables were excluded from the analysis.

Data collection

Variables for STAS included: Age, sex, smoking history, lesion location and pathological tumor (T) stage (categorized as T0, T1 or T2 according to the 9th edition of the TNM classification).

Radiological features based on preoperative chest CT included nodule types [pure ground-glass opacity (pGGO), part-solid or solid], presence of lobulation, spiculation and vacuole sign. Due to the absence of universally standardized criteria for nodule classification, nodule types were defined according to widely accepted radiological principles to minimize inter-observer variability and ensure reproducibility (16). The nodule types were as follows: i) pGGO: A lesion indicating homogeneous increased attenuation of the lung parenchyma without any solid component obscuring the underlying lung architecture; ii) part-solid: A lesion containing both ground-glass and solid components, with the solid portion partially obscuring the lung parenchymal markings; and iii) solid: A lesion in which the underlying lung architecture is completely obscured by soft-tissue attenuation. All CT images were reviewed by two experienced radiologists blinded to the outcome.

Serum tumor markers included squamous cell carcinoma antigen (SCC), carcinoembryonic antigen (CEA), cytokeratin 19 fragment (CYFRA21-1), neuron-specific enolase (NSE) and ferritin. Continuous variables such as tumor marker levels and nodule diameter were transformed into categorical variables: CEA (≤5.0 ng/ml as normal), CYFRA21-1 (≤3.3 ng/ml), NSE (≤17.0 ng/ml), SCC (≤1.5 ng/ml) and ferritin (21.8-274.6 ng/ml for men, 13.0-150.0 ng/ml for women). All variable categorizations were predefined and applied consistently across training and test sets.

Outcome definition

The primary outcome of the present study was the presence of STAS. STAS was defined in accordance with the 2021 World Health Organization Classification of Thoracic Tumors (7). According to this definition, STAS is identified when free tumor cells are identified within alveolar spaces beyond the main tumor edge, with a minimum distance of at least two alveolar septa from the tumor margin.

All specimens were obtained from surgical resections and independently evaluated by experienced pulmonary pathologists blinded to clinical and imaging information. STAS+ status was assessed on hematoxylin and eosin (H&E)-stained sections, which were prepared according to routine standardized histopathological protocols (17), by two experienced pulmonary pathologists using light microscopy and was assigned only when detached tumor clusters (micropapillary, solid nests or single cells) were clearly visible beyond the tumor edge.

Statistical analysis and model development

Baseline characteristics were summarized as median and interquartile range (IQR) for continuous variables and as counts with percentages (n; %) for categorical variables. To develop a predictive model for STAS, five supervised ML algorithms were used: Logistic regression, random forest, eXtreme Gradient Boosting (XGBoost), support vector machine (SVM) and naïve Bayes (18–22).

The dataset was randomly divided into a training and a test set in an 8:2 ratio. Model training and evaluation were performed using 5-fold cross-validation on the training set to reduce random variability and prevent overfitting. Model performance was assessed using accuracy, sensitivity, specificity and area under the receiver operating characteristic (ROC) curve (AUC). Calibration performance was quantified using the Brier score (23), which measures the mean squared difference between predicted probabilities and actual outcomes, with lower values indicating improved probability calibration. Mean ROC curves were computed by averaging across folds. Decision curve analysis was conducted to assess the net clinical benefit across various threshold probabilities. To comprehensively assess model consistency, a Taylor diagram was generated to evaluate the correlation coefficient, standard deviation and root mean square error between predicted probabilities and observed outcomes across models.

To optimize model performance, grid search with 5-fold cross-validation was applied to tune hyperparameters for all five models on the training dataset. The complete parameter search spaces and optimal settings for each model are summarized in Table SI.

The optimal model was selected based on the highest AUC in the test set. No feature selection techniques were applied prior to model training. All categorical variables, including sex, smoking history and radiological features, were converted into dummy variables using one-hot encoding. Continuous variables such as tumor marker levels had already been transformed into categorical variables and therefore no additional normalization or scaling was required. The model outputs were treated as continuous probability scores and no fixed cut-off thresholds or predefined risk categories were used.

To enhance interpretability, global SHapley Additive exPlanations (SHAP) analysis was performed for the best-performing model, which was selected based on the highest AUC in the test set and further validated against using DeLong's test. Following established methodological practices reported in recent oncology and radiology research (24,25), the SHAP analysis quantified the overall contribution of each feature across the entire dataset. Summary and dependence plots were then generated to visualize feature importance, effect direction and potential interaction patterns.

All analyses were performed using Python (version 3.9), with implementation via the scikit-learn (https://scikit-learn.org/stable/index.html) and XGBoost libraries (https://xgboost.ai/).

Results

Baseline characteristics

A total of 584 patients with resected NSCLC were included, among whom 259 (44.4%) were classified as STAS+ and 325 (55.6%) as STAS−. The median age was 62 years (IQR, 55–69). Of the patients, 318 (54.5%) were women and 133 (22.8%) reported a history of smoking. Most patients were classified as T1 stage (n=467, 80.0%), with fewer at T2 (n=112, 19.1%) and T3 stages (n=5, 0.9%). Common radiological signs included lobulation (n=353, 60.5%), spicule sign (n=334, 57.2%) and vacuolar sign (n=84, 14.4%). Nodule types were primarily subsolid (n=326, 55.8%), followed by solid (n=199, 34.1%) and pure ground-glass nodules (n=59, 10.1%). The most frequent tumor locations were the right upper (n=207, 35.4%) and left upper lobes (n=153, 26.2%). Abnormal levels of tumor markers were present in a minority of patients: CYFRA21-1 (n=113, 19.4%), NSE (n=42, 7.2%), CEA (n=53, 9.1%), SCC (n=20, 3.4%) and ferritin (n=93, 15.9%). Detailed baseline characteristics are summarized in Table I.

Table I. Baseline characteristics of the present study population (n=584).

Table I.

Baseline characteristics of the present study population (n=584).

Variable	Patients
Age, years (range)	62 (55–69)
Sex
Female	318 (54.5)
Male	266 (45.5)
Smoking history (Yes)	133 (22.8)
T staging
T1	467 (80.0)
T2	112 (19.1)
T3	5 (0.9)
Lobulation (Yes)	353 (60.5)
Spicule sign (Yes)	334 (57.2)
Vacuolar sign (Yes)	84 (14.4)
Tumor location
Right lower lung	103 (17.6)
Right upper lung	207 (35.5)
Right middle lung	32 (5.5)
Left lower lung	89 (15.2)
Left upper lung	153 (26.2)
Nodule type
Subsolid	326 (55.8)
Solid	199 (34.1)
Ground-glass	59 (10.1)
SCC
Abnormal	20 (3.4)
Normal	564 (96.6)
CEA
Abnormal	53 (9.1)
Normal	531 (90.9)
CYFRA21-1
Abnormal	113 (19.4)
Normal	471 (80.6)
NSE
Abnormal	42 (7.2)
Normal	542 (92.8)
Ferritin
Abnormal	93 (15.9)
Normal	491 (84.1)
STAS
Positive	259 (44.4)
Negative	325 (55.6)

[i] Data are presented as n (%). SCC, squamous cell carcinoma; CYFRA21-1, cytokeratin fragment-19; NSE, neuron-specific enolase; STAS, spread through air spaces; CEA, carcinoembryonic antigen.

Model performance evaluation

In the training set, 5-fold cross-validation demonstrated stable performance (Fig. 1A) and the model generalized well to the test set (Fig. 1B). Among the five algorithms, XGBoost demonstrated the best overall performance, achieving the highest mean AUC across five cross-validation folds in the training set (0.868; 95% CI, 0.860-0.875; Fig. 1C) and maintaining robust performance in the test set (mean AUC =0.764; 95% CI, 0.718-0.815; Fig. 1D). In addition to the highest AUC, XGBoost exhibited balanced accuracy, sensitivity and specificity and achieved the lowest Brier scores (0.158 in the training set and 0.197 in the test set), indicating notably increased discrimination and calibration among the tested models (Table II).

Figure 1. Model evaluation curves. (A) 5-fold cross-validation AUC across folds on training set. (B) 5-fold cross-validation AUC across folds on test set. (C) ROC curves in the training set. (D) ROC curves on test set. (E) Calibration curves on training set. (F) Calibration curves on test set. (G) Pairwise DeLong's test P-values for AUC comparisons. Lower P-values (blue) indicate significant differences between models. (H) Taylor diagram summarizing correlation, standard deviation and root mean square error between predicted and observed values. Models closer to the reference point indicate improved overall agreement. AUC, area under the curve; ROC, receiver operating characteristic; XGBoost, eXtreme Gradient Boosting; SVM, support vector machine.

Table II. Performance of machine learning models.

Table II.

Performance of machine learning models.

Model	Dataset	Accuracy	Sensitivity	Specificity	AUC (95% CI)	Brier score
Logistic regression	Training	0.685	0.724	0.647	0.750 (0.729-0.770)	0.203
	Test	0.638	0.677	0.600	0.701 (0.582-0.783)	0.223
Random forest	Training	0.726	0.774	0.677	0.817 (0.802-0.835)	0.182
	Test	0.680	0.733	0.630	0.745 (0.667-0.802)	0.206
XGBoost	Training	0.775	0.743	0.808	0.868 (0.860-0.875)	0.157
	Test	0.680	0.657	0.700	0.764 (0.718-0.815)	0.197
Support vector machine	Training	0.751	0.819	0.682	0.842 (0.834-0.852)	0.173
	Test	0.669	0.751	0.588	0.730 (0.615-0.801)	0.212
Naïve Bayes	Training	0.677	0.636	0.718	0.743 (0.729-0.766)	0.240
	Test	0.660	0.621	0.700	0.695 (0.576-0.753)	0.269

[i] AUC, area under the curve; XGBoost, eXtreme gradient boosting.

DeLong's tests demonstrated that XGBoost yielded significantly higher AUCs compared with logistic regression, naïve Bayes and random forest (all P<0.05), whereas its difference with SVM did not reach statistical significance (P>0.05; Fig. 1G). Calibration curves (Fig. 1E and F) further demonstrated closer agreement between predicted probabilities and observed outcomes for XGBoost. The Taylor diagram demonstrated that the XGBoost model was positioned closest to the reference point, indicating the highest overall agreement with the observed data among all algorithms (Fig. 1H).

Feature importance analysis

To enhance model interpretability, the present study employed SHAP to analyze the contribution of individual features in the XGBoost model. The feature importance plot (Fig. 2A) demonstrated that the nodule type, lobulation and smoking history were the top three predictors contributing most to the decision-making of the model. The Beeswarm plot (Fig. 2B) further illustrated the direction and magnitude of the effect of each feature. For example, solid nodule types and presence of lobulation positively contributed to STAS prediction. The history of smoking was also associated with higher predicted risk. By contrast, features such as NSE and CYFRA21-1 exhibited minimal contribution.

Figure 2. SHAP analysis for the XGBoost model. (A) Bar plot showing the mean absolute SHAP value of each feature, indicating its overall importance in the model. (B) Beeswarm plot displaying the impact and directionality of each feature on model output across all cases. Red points represent higher feature values, blue points indicate lower values. XGBoost, eXtreme Gradient Boosting; SCC, squamous cell carcinoma; CYFRA21-1, cytokeratin fragment-19; SHAP, SHapley Additive exPlanations; NSE, neuron-specific enolase; CEA, carcinoembryonic antigen.

Discussion

In the present study, an XGBoost-based model was developed using real-world preoperative data to predict STAS in surgically resectable NSCLC. Unlike previous approaches that rely on postoperative pathology or advanced imaging techniques, the present study model enables early, non-invasive prediction of STAS and provides thoracic surgeons with a valuable decision-support tool to optimize preoperative surgical planning.

Methodologically, to enhance model robustness, reduce selection bias and determine the most appropriate algorithm for STAS prediction under real-world clinical constraints, the present study evaluated five distinct ML models. Among them, XGBoost consistently outperformed the others on both the training (AUC=0.868) and test (AUC=0.764) sets. As a gradient boosting framework, XGBoost is particularly adept at capturing complex non-linear relationships and feature interactions, and has demonstrated robust performance across a wide range of clinical diagnostic and prognostic tasks (26–28). These prior findings were consistent with the present study results, further supporting the suitability of XGBoost in the context of oncology. Therefore, the present study ultimately selected XGBoost as the final model for STAS prediction in this study.

In terms of predictive findings, the present study confirmed and extended existing knowledge regarding radiological and clinical associations of STAS, while offering improved generalizability and clinical practicality. Several prior studies have investigated the association between STAS and radiological or pathological features. For example, radiomics-based models have achieved high AUCs by incorporating handcrafted features and clinical data, but often suffer from limited generalizability due to complex processing pipelines (29). Large retrospective analyses have also associated STAS with programmed cell death-ligand 1 expression, spiculation and lobulated margins, yet these features alone provide inconsistent predictive power (30). A recent study using an XGBoost and SHAP framework in lung adenocarcinoma reported higher AUCs and identified similar predictive features, such as nodule type and lobulation, but relied heavily on radiomic features and was restricted to a single histological subtype (31). By contrast, the present study model offers broader applicability across NSCLC subtypes and requires only routinely collected variables, making it more feasible for real-world deployment.

Building on the performance of the model, the contribution of individual features to STAS prediction was further explored. The strong predictive value of nodule type and lobulation aligns with prior studies that have consistently associated these semantic imaging features with histological patterns associated with STAS. Specifically, previous studies have reported that solid and part-solid nodules, compared with pure ground-glass nodules, are more frequently associated with aggressive histological subtypes such as micropapillary and solid patterns (32,33). These lesions typically reflect higher tumor cellularity and invasiveness, making them more prone to tumor cell dissemination into peripheral airspaces (34). Lobulation, another key predictive feature, similarly corresponds with biological plausibility. Radiologically, lobulated tumor margins indicate an irregular and heterogeneous tumor-stromal interface, which may reflect underlying extracellular matrix remodeling, neoangiogenesis and disrupted cell adhesion at the invasive front. These alterations facilitate tumor cell detachment and alveolar spread, thereby elevating STAS risk (35–37). This interpretation is supported by multiple histopathological studies that have associated lobulated tumors with increased invasiveness and worse clinical outcomes (38,39).

Smoking history was also identified as a key predictor in the present study model. A recent predictive modeling study involving 1,212 patients with clinical stage IA lung adenocarcinoma (15) reported that smoking history was a consistent and independent predictor of STAS in both logistic regression models, with one achieving an AUC of 0.807 (40). Biologically, smoking contributes to NSCLC pathogenesis through mechanisms such as genomic instability (for example, TP53 mutations), chronic inflammation and epithelial-mesenchymal transition (41–43), all of which promote tumor invasiveness and may facilitate airspace spread. These findings reinforce the interpretability and clinical relevance of the present study model. Furthermore, smoking history is easily obtainable in routine preoperative evaluations, further supporting its utility as a practical feature in STAS risk prediction.

In contrast to the strong predictive signals from imaging and clinical factors, conventional serum tumor markers including CEA, CYFRA21-1 and NSE, contributed minimally to STAS prediction in the present study model. This finding aligns with previous studies, some of which included CEA in their models but reported non-significant associations with STAS status (P=0.359) or low odds ratios, suggesting limited discriminatory power (44,45). Although one study identified CEA as part of a combined clinical-radiological nomogram, its individual predictive value was relatively weak compared with radiological features such as density type and the distal ribbon sign (35). From a biological perspective, these serum markers primarily reflect systemic tumor burden rather than specific patterns of local invasion such as STAS (46). Furthermore, their serum levels are subject to variation due to non-tumoral factors and often lack specificity in early-stage disease. The present study findings thus reinforce the utility of readily accessible imaging features and clinical variables over traditional tumor markers for the preoperative risk assessment of STAS.

Despite these promising findings, several limitations should be acknowledged. First, this was a retrospective study based on a single-center cohort, which may limit the generalizability of the present study model. External validation using multicenter datasets with broader population diversity is necessary to confirm its robustness. Second, although the present study selected clinically accessible features, the interpretation of certain CT signs (for example, lobulation) may be subject to inter-observer variability. Future research incorporating automated image analysis or radiomics-based augmentation may further enhance reproducibility. Third, while SHAP plots improved model transparency, the current model still lacks direct histological validation of the features associated with STAS, which would be valuable for biological interpretability. Lastly, the present study dataset did not include genomic or molecular data, which might contribute additional predictive value and deserves exploration in future studies.

In conclusion, the present study developed an interpretable XGBoost model using routinely available preoperative clinical and semantic CT features to predict STAS in surgically resectable NSCLC. The model demonstrated favorable predictive performance and identified clinically notable features such as nodule type, lobulation and smoking history. By enabling early, non-invasive risk stratification, this approach may assist thoracic surgeons in tailoring surgical strategies prior to resection. Future prospective and multicenter studies are warranted to externally validate the model and facilitate its potential integration into clinical workflows.

Supplementary Material

Supporting Data

Acknowledgements

Not applicable.

Funding

Funding: No funding was received.

Availability of data and materials

The data generated in the present study may be requested from the corresponding author.

Authors' contributions

CY and YY conceived and designed the study. CY collected and analyzed the data. GD, BZ, LX and FC contributed to data interpretation and key revision of the manuscript. CY drafted the initial manuscript. YY supervised the study and revised the manuscript for key intellectual content. CY and YY confirm the authenticity of all the raw data. All authors read and approved the final version of the manuscript.

Ethics approval and consent to participate

The present study was approved by the Ethics Committee of Anqing Municipal Hospital (approval no. 2025-152; Anqing, China). Written informed consent was obtained from all participants prior to data collection and all patient data were anonymized before analysis to protect privacy.

Patient consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

References