AI Model Predicts Lung Cancer Spread Using Routine PET Scans and Blood Tests

Lung cancer remains the deadliest malignancy worldwide, and one of the most consequential questions a surgeon faces before operating on a patient with non-small cell lung cancer is deceptively simple: has the tumor already spread to the lymph nodes? The answer shapes everything that follows, from the extent of surgical dissection to the choice of adjuvant therapy and the patient’s overall prognosis. A new study published in PLOS One by Yunchao Xu, Jiajun Hao, Yujiao Dou, and Biao Zeng tackles this question with an interpretable machine-learning workflow built entirely from data that clinicians already collect in routine practice, offering a glimpse of how artificial intelligence might streamline preoperative staging without expensive new tests.

The research team assembled a retrospective cohort of 255 patients with histologically confirmed non-small cell lung cancer, every one of whom had undergone systematic lymph node dissection, meaning that the true pathological status of their lymph nodes was known with certainty. This gold-standard ground truth is critical, because many previous prediction studies relied on sampling rather than complete dissection, leaving room for missed metastases to distort the results. The patients were randomly divided, with stratification by their pathological lymph node status to preserve the balance of positive and negative cases, into a training cohort of 178 patients and an internal held-out test cohort of 77 patients.

Before any machine learning began, the investigators screened candidate predictors using univariable and multivariable logistic regression within the training cohort. This filtering step matters because feeding an algorithm hundreds of weakly informative variables invites overfitting, the statistical equivalent of memorizing noise. What survived the screen was strikingly parsimonious: central tumor location, elevated carcinoembryonic antigen, elevated squamous cell carcinoma antigen, and solid nodule status on imaging. Perhaps the most surprising finding was what did not survive. None of the quantitative PET metabolic parameters, such as standardized uptake values that measure how avidly a tumor consumes radioactive glucose, remained independently associated with lymph node metastasis.

That negative result challenges a widespread intuition. Positron emission tomography combined with computed tomography is a cornerstone of modern lung cancer staging, and many clinicians assume that the metabolic numbers extracted from these scans carry independent predictive power. In this dataset, once clinical location, serological tumor markers, and the morphological character of the nodule were accounted for, the quantitative metabolic measurements added nothing. The study’s title emphasizes morphological features for good reason: whether a nodule appears solid, as opposed to partly ground-glass in texture, turned out to be one of the four pillars of prediction, a visual quality that radiologists assess routinely.

With four predictors in hand, the team trained six different machine-learning algorithms, optimizing each through repeated nested five-fold cross-validation. Nested cross-validation is the rigorous way to tune hyperparameters while keeping an honest estimate of performance, since the inner loop selects settings and the outer loop evaluates them on data not used for that selection. Among the six candidates, XGBoost, a gradient-boosted decision tree method that builds an ensemble of weak learners sequentially, each correcting the errors of its predecessors, achieved the highest cross-validated area under the receiver operating characteristic curve in the training cohort: 0.735, with a 95 percent confidence interval of 0.653 to 0.813. It was therefore designated the primary model.

The decisive test came when all six models, frozen and unmodified, were applied to the held-out test cohort of 77 patients. XGBoost delivered an AUC of 0.747, with a 95 percent confidence interval of 0.631 to 0.846, confirming that its training performance was not an artifact of overfitting. An AUC in this range represents moderate discrimination, meaning the model separates patients with and without nodal spread considerably better than a coin flip but far short of diagnostic certainty. The authors also compared the test-set AUCs of all six algorithms using paired DeLong tests with Holm correction for multiple comparisons, and found no statistically significant differences among them, a humbling result for anyone hoping that a sophisticated algorithm would dramatically outperform its rivals.

To probe that point further, the researchers added a conventional multivariable logistic regression model using the same four predictors as a simple reference. On the test cohort, this old-fashioned statistical workhorse achieved an AUC of 0.733, with a confidence interval of 0.612 to 0.841, and a DeLong test found no significant difference from XGBoost, with a P value of 0.301. Even more telling, the logistic regression model posted a Brier score of 0.152, compared with 0.198 for XGBoost, indicating that the simpler model’s predicted probabilities were actually better calibrated to observed outcomes. In plain terms, a transparent equation with four terms matched the flagship machine-learning algorithm in discrimination and beat it in calibration.

The primary XGBoost model was additionally evaluated for calibration, clinical utility, and feature contribution using SHAP, or SHapley Additive exPlanations, a technique borrowed from cooperative game theory that assigns each feature a quantified contribution to every individual prediction. This interpretability layer is increasingly considered essential for clinical machine learning, because a black-box risk score that physicians cannot interrogate is unlikely to earn trust or survive regulatory scrutiny. The study also reported a high negative predictive value, meaning that when the model predicted an absence of lymph node metastasis, that reassurance was usually correct, a property with real surgical relevance for selecting patients who might safely avoid extended nodal dissection.

Yet the authors are admirably candid about the limits. The model’s positive predictive value was limited, so a positive prediction could not reliably confirm metastasis, and invasive pathological confirmation remains indispensable. The absence of significant differences among the six algorithms, combined with the near-identical performance of plain logistic regression, suggests that the predictive ceiling here is set by the information content of the four features themselves, not by the sophistication of the learner. Moderate discrimination, a modest sample of 255 patients from a single center, and an internal test cohort rather than a truly external one all argue for caution.

The study’s real contribution may be methodological discipline rather than algorithmic novelty. It demonstrates a standardized, transparent workflow, from predictor screening through nested cross-validation to held-out evaluation, calibration analysis, and SHAP-based explanation, using only parameters available before surgery. The authors conclude that external multicenter evaluation is required before any clinical implementation, and that the results support cautious interpretation. If future multi-institutional validation confirms these findings, the payoff could be meaningful: a simple, interpretable risk score, computable from a CT report and two blood tests, that helps surgeons stratify lymph node metastasis risk in non-small cell lung cancer and decide, with greater confidence, who needs the most aggressive nodal surgery and who does not.

Subject of Research: Predicting lymph node metastasis in non-small cell lung cancer using clinical, serological, and PET/CT-derived features with interpretable machine learning

Article Title: Predicting lymph node metastasis in non-small cell lung cancer using clinical and PET/CT-derived morphological features: An interpretable machine-learning study

Article References: Xu, Y., Hao, J., Dou, Y., & Zeng, B. (2026). Predicting lymph node metastasis in non-small cell lung cancer using clinical and PET/CT-derived morphological features: An interpretable machine-learning study. PLOS One, 21(10), e0361043. https://doi.org/10.1371/journal.pone.0361043

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0361043

Keywords: non-small cell lung cancer, lymph node metastasis, machine learning, XGBoost, PET/CT, logistic regression, SHAP, tumor markers, carcinoembryonic antigen, cancer staging, predictive modeling, cross-validation