Routine Clinical Clues Outperformed PET Quantitation in Preoperative NSCLC Nodal Staging
In non-small cell lung cancer, the difference between node-negative disease and occult lymph node metastasis can reshape every part of care, from surgical planning to adjuvant treatment. A new PLOS ONE study tested whether a machine-learning workflow built from routine clinical, serologic, and PET/CT-derived variables could improve preoperative identification of lymph node metastasis before the chest is opened. The answer was guarded rather than dramatic: a four-variable model reached moderate discrimination, but the strongest signal came from simple clinicopathologic features rather than the quantitative PET measures that often receive the most attention.
The study analyzed 255 patients with histologically confirmed adenocarcinoma or squamous cell carcinoma who underwent systematic lymph-node dissection. Pathologic lymph-node status served as the reference standard, which gives the analysis a firm surgical anchor. Patients were split 7:3 into a training cohort of 178 and an internal held-out test cohort of 77, with stratification by nodal status so the event distribution stayed balanced across the split.
Only 58 of 255 patients had lymph node metastasis.
Four Binary Predictors Carried the Signal
Candidate variables went through univariable and multivariable logistic regression in the training cohort before machine-learning development began. Four binary predictors survived that screening: central tumor location, elevated carcinoembryonic antigen, elevated squamous cell carcinoma antigen, and solid nodule status. Those features are biologically plausible on first principles. A centrally located tumor has closer access to hilar and mediastinal lymphatic drainage, while a solid morphology usually reflects more invasive tissue architecture than ground-glass–predominant disease. Elevated serum tumor markers may reflect a larger or more biologically active tumor burden, although they remain indirect surrogates rather than mechanistic proof of spread.
By contrast, none of the quantitative PET metabolic variables retained independent association with nodal metastasis after adjustment. SUVmax, SUVmean, metabolic tumor volume, and total lesion glycolysis were evaluated, but they did not remain in the final predictor set. For clinicians accustomed to treating PET-derived metabolism as a shorthand for aggressiveness, this is a useful corrective: metabolic intensity alone did not outperform the simpler morphologic and serologic signals in this dataset.
XGBoost Led the Field, but the Margin Was Small
Six classifiers were trained and tuned using repeated nested five-fold cross-validation: support vector machine, artificial neural network, XGBoost, extra trees, decision tree, and gradient boosting. Among them, XGBoost achieved the highest cross-validated AUC in the training cohort at 0.735, with a 95% confidence interval of 0.653 to 0.813, and was selected as the primary model.
In the internal held-out test cohort, XGBoost produced an AUC of 0.747 (95% CI, 0.631–0.846). A conventional multivariable logistic regression model using the same four predictors reached a test AUC of 0.733 (95% CI, 0.612–0.841). The difference was not statistically significant; the paired DeLong test yielded P = 0.301. In practical terms, the machine-learning approach did not clearly separate itself from a standard regression framework when both relied on the same limited feature set.
The model comparison is a reminder that algorithmic complexity does not guarantee clinical gain. When the predictor pool is small and dominated by binary variables, a well-specified regression model can keep pace with a more elaborate ensemble method.
Calibration, SHAP, and the Limits of Preoperative Certainty
Model performance was not judged by discrimination alone. The primary model was also assessed for calibration, clinical utility, and feature contribution using SHAP, the interpretability framework used to map each predictor’s contribution to the final output. That matters because nodal staging tools must do more than sort patients into higher and lower risk bins; they have to give surgeons and oncologists a probability estimate that can survive bedside scrutiny.
Calibration favored the reference logistic regression model over XGBoost. The Brier score was 0.152 for logistic regression versus 0.198 for XGBoost, indicating better overall probability accuracy for the simpler model. When a model is intended to inform preoperative decision-making, poor calibration can be as troublesome as weak discrimination, because an overconfident probability estimate may push clinicians toward unnecessary escalation or false reassurance.
The study’s final conclusion was measured and appropriate. XGBoost showed moderate discrimination and a high negative predictive value for pathological lymph node metastasis, but its positive predictive value remained limited. In a disease where the consequences of missed nodal spread are substantial, a tool that performs best at ruling disease out is useful only within a cautious framework. It may help identify some patients unlikely to harbor nodal disease, yet it cannot replace tissue-based staging when the pretest probability is concerning.
Why the PET Signal May Have Faded
The finding that quantitative PET metrics did not survive multivariable adjustment deserves attention. PET metabolism reflects glucose avidity, hypoxia, and cellular proliferation, but nodal spread depends on more than metabolic intensity. Tumor location, invasive architecture, and serum marker patterns may better capture the anatomic and biological conditions that allow malignant cells to access lymphatic channels. In a retrospective cohort with only 58 nodal events, the incremental value of imaging-derived quantitation may also have been blunted by limited event counts and the redundancy between related metabolic variables.
The study design adds both strength and restraint. All patients underwent systematic lymph-node dissection, so nodal status was not inferred from imaging alone. At the same time, the analysis came from a single center and used an internal held-out test cohort rather than an external multicenter validation set. The dataset also recorded CEA and SCC antigen only as normal or elevated, which reduces granularity and may hide dose-response information that could matter clinically.
What This Means for Surgical Staging
For thoracic teams, the practical message is not that machine learning has failed, but that the current formulation has not yet earned routine use. Central tumor location, elevated CEA, elevated SCC antigen, and solid nodule status form a concise risk profile that can be applied before surgery, and the resulting model may help identify patients at lower risk of occult nodal spread. Still, the positive predictive value was limited, so an elevated predicted risk should not be treated as proof of metastasis.
External multicenter validation is the next essential step. Until then, this model should be viewed as a research-stage tool rather than a preoperative standard, especially in a disease where nodal staging directly influences operative extent, adjuvant therapy, and prognosis.