The staging decision that determines treatment. In non-small cell lung cancer (NSCLC), whether cancer has spread to mediastinal lymph nodes - the nodes in the central chest between the lungs - is one of the most important determinants of treatment. Patients with N2 disease (cancer in ipsilateral mediastinal nodes) typically require neoadjuvant chemotherapy or chemoradiation before surgery, while those with N0-1 disease may proceed directly to surgical resection. Accurately classifying patients into these groups before any surgery begins is essential for delivering appropriate therapy.
Current diagnostic methods and their limitations. The gold standard for mediastinal lymph node assessment is histopathological examination through endobronchial ultrasound-guided biopsy (EBUS-TBNA) or video-assisted mediastinoscopy. However, these are invasive procedures with false-negative rates of 15-25%, patient tolerance limitations, and potential complications including bleeding, infection, and nerve injury. Standard PET-CT alone achieves pooled sensitivity of approximately 0.82 and specificity of 0.88 for nodal metastasis detection - useful but insufficient for definitive preoperative staging.
The gap in prior research. Most existing machine learning models for predicting lymph node metastasis in NSCLC have used only primary tumor imaging features or limited clinical variables, ignoring the PET-CT information from the lymph node stations themselves. The metabolic and anatomical characteristics of suspicious lymph nodes on PET-CT provide information distinct from the primary tumor and carry independent predictive value. No prior study had integrated tumor imaging, lymph node imaging, pathological features, and clinical data within a unified machine learning framework applied at the individual lymph node station level.
Study objective and approach. This study developed and validated machine learning models that combine PET-CT imaging features from both primary tumors and mediastinal lymph node stations with clinical and pathological data, comparing three model configurations against each other and against the standard PET-CT diagnostic threshold to identify the optimal approach for preoperative mediastinal staging.
Patient population. The retrospective cohort included 390 NSCLC patients from the First Hospital of Jilin University who underwent tumor resection with systematic mediastinal lymph node dissection between January 2017 and December 2023. All patients had 18F-FDG PET-CT within two weeks before surgery. Among the 390 patients, 241 (61.8%) had pN0-1 disease and 149 (38.2%) had pN2 mediastinal involvement. A total of 1,026 mediastinal lymph node stations were analyzed, of which 204 (19.9%) showed metastasis.
PET-CT feature extraction. PET-CT was performed on a Siemens Biograph 16HR scanner with standardized 18F-FDG injection and image acquisition protocols. For both primary tumors and each resected mediastinal lymph node station, regions of interest were manually delineated by two nuclear medicine physicians with approximately 15 years of experience each. Quantitative metabolic parameters automatically extracted included SUVmax, SUVmean, SUVpeak, SULmax, metabolic tumor volume (MTV), and total lesion glycolysis (TLG). Qualitative CT features assessed visually included spiculation, lobulation, bronchial cutoff, pleural indentation, and calcification.
Clinical and pathological features. Clinical variables included age, sex, smoking history, comorbidities (hypertension, diabetes, coronary heart disease), pulmonary function (FEV1), and 11 serum tumor markers including CEA, Cyfra21-1, CA125, CA19-9, NSE, SCC-Ag, AFP, and ProGRP. Pathological features included primary tumor histological type, microhistological subtype, maximum diameter, presence of necrosis, nerve invasion, vascular invasion, visceral pleural invasion, and TNM stage according to the 8th edition staging system.
Feature selection using random forest and SHAP. From an initial pool of 41 imaging features, 26 clinical and pathological variables, and 20 additional parameters, features were ranked by their contribution to Out-of-Bag error in a random forest model and by SHAP (SHapley Additive exPlanations) values. Features cumulatively contributing to 90% of total model variance were retained: 20 imaging features and 15 clinical-pathological features. SHAP analysis identified TLG, SUVmax, and MTV as the top imaging contributors, while CEA, CA19-9, and tumor histological type were the primary clinical-pathological predictors.
Eight algorithms compared. The study systematically evaluated Logistic Regression, Classification and Regression Tree (CART), Support Vector Machine (SVM), Gradient Boosting Decision Tree (GBDT), Random Forest, Multi-Layer Perceptron (MLP), XGBoost, and k-Nearest Neighbor (KNN) for the mediastinal lymph node metastasis prediction task. All models were trained with standardized pipelines, nested cross-validation for hyperparameter optimization, and probability calibration using Platt scaling or isotonic regression.
Three model configurations. Three distinct feature combinations were compared to quantify the contribution of each data domain: TPC (Tumor imaging + Pathological + Clinical features), LPC (Lymph node imaging + Pathological + Clinical features), and TLPC (Tumor imaging + Lymph node imaging + Pathological + Clinical features combined). This design allowed direct measurement of how much predictive value comes from lymph node imaging features alone versus primary tumor imaging versus the integrated combination.
Training, test, and validation sets. The dataset of 1,026 lymph node stations was split using stratified random sampling into a training set (n = 410), test set (n = 308), and validation set (n = 308), with the same proportion of positive and negative cases preserved in each subset. This three-way split provided separate held-out evaluation sets for initial model selection (test set) and confirmatory performance assessment (validation set), reducing the risk of overfitting to any single evaluation partition.
Statistical rigor in model comparison. AUC differences between models were formally tested using DeLong's method. Given that multiple models and feature combinations were compared, the Benjamini-Hochberg false discovery rate correction was applied to p-values to control for multiple comparisons, with an adjusted significance threshold of q less than 0.05. Decision curve analysis assessed clinical net benefit across the full range of decision thresholds, providing a more clinically relevant comparison than ROC statistics alone.
XGBoost outperforms all other algorithms. Across training, test, and validation sets, XGBoost consistently achieved the highest AUC values: 0.91 in training, 0.90 in the test set, and 0.87 in validation, with 87.2% accuracy. The next best performers were GBDT (0.90, 0.88, 0.85) and MLP (0.88, 0.87, 0.84), while CART and KNN performed worst (0.77-0.78 in test sets). Decision curve analysis confirmed that XGBoost generated positive clinical net benefit across all decision thresholds, clearly surpassing all alternative algorithms in the range of threshold probabilities most relevant to clinical practice.
Integrating all data sources is critical. The three model configurations showed a striking performance hierarchy. The combined TLPC model achieved AUC 0.90 (sensitivity 0.96, specificity 0.84) in the test set and AUC 0.87 (sensitivity 0.93, specificity 0.76) in validation. The LPC model (lymph node imaging only) reached AUC 0.78 in testing and 0.68 in validation - substantially lower. The TPC model (primary tumor imaging only) performed worst at AUC 0.67 in testing and 0.65 in validation, with sensitivity of only 0.56. This hierarchy confirms that lymph node imaging features provide more direct predictive information than primary tumor features, and combining both with clinical data further amplifies accuracy.
Outperforms standard clinical criteria. The conventional PET-CT diagnostic standard (SUVmax greater than 2.5 or lymph node short-axis diameter greater than 1 cm) achieved AUC 0.71 with sensitivity 0.74 and specificity 0.70. The TLPC-XGBoost model significantly outperformed this threshold on all three metrics: AUC improvement from 0.71 to 0.90, sensitivity improvement from 0.74 to 0.96, and specificity improvement from 0.70 to 0.84. The 96% sensitivity is particularly important clinically - it means the model misses very few true cases of mediastinal involvement that would require modified treatment planning.
Key predictive features identified by SHAP. SHAP analysis showed that metabolic features dominated the TLPC model: TLG, SUVmax, and MTV were the highest-importance imaging features, reflecting that total tumor metabolic burden and peak metabolic activity are the strongest imaging predictors of nodal spread. Among clinical and pathological variables, CEA level, CA19-9, and histological type (squamous versus adenocarcinoma) contributed most, with FEV1 as an additional independent factor reflecting systemic disease burden.
A first in lymph node station-level prediction. This study is the first to integrate 3D volumetric PET-CT features from individual mediastinal lymph node stations with primary tumor imaging, clinical laboratory values, and pathological characteristics within a unified machine learning framework. Prior studies modeled lymph node metastasis at the patient level or used only tumor-derived radiomic features, missing the complementary predictive information contained in the imaging characteristics of the lymph nodes themselves.
Clinical utility for treatment stratification. The TLPC model's 96% sensitivity means that nearly all patients with mediastinal involvement would be identified for neoadjuvant therapy or more intensive staging, potentially avoiding the harm of undertreatment. The 84% specificity means fewer patients without mediastinal involvement would be unnecessarily assigned to neoadjuvant protocols, avoiding treatment delays and toxicities. This combination supports more accurate preoperative N2 risk stratification and could guide decisions about whether invasive mediastinal staging is warranted in specific patients.
Why XGBoost excels for this data type. XGBoost's advantage in this study reflects its structural suitability for the data: its gradient boosting framework captures nonlinear interactions between features that simpler models assume are linear, its built-in regularization reduces overfitting when input features are correlated (as metabolic parameters are), and its native handling of missing values accommodates the incomplete clinical records common in retrospective studies. Random feature sampling at each tree also reduces multicollinearity effects among the highly correlated SUV parameters.
Key limitations and future work. The study used single-center data from Jilin University Hospital, and PET-CT was performed selectively (403 of 1,149 consecutive surgical patients) based on multidisciplinary team referral rather than routine acquisition. This enriched the cohort with higher-risk patients and may not reflect the distribution of a routine screening population. Manual 3D delineation of lymph node VOIs by nuclear medicine physicians is time-consuming and subject to inter-operator variability. External validation in multi-center cohorts and prospective evaluation of TLPC-guided treatment decisions are planned next steps.