The Clinical Challenge Determining whether lung cancer has spread to lymph nodes (N staging) or distant organs (M staging) is critical for treatment planning. Current methods - PET/CT, biopsy, and bronchoscopy - are expensive, invasive, and not universally accessible. A non-invasive, low-cost predictive model based on routine laboratory data would fill a meaningful clinical gap.
The ML Approach This study trained multiple machine learning models on clinical laboratory data from 1,629 lung cancer patients to predict both lymph node metastasis (N stage) and distant metastasis (M stage). The models used readily available blood test results - complete blood count, biochemistry, and tumor markers - as input features.
Best Model Performance Logistic regression achieved the best performance, with AUC of 0.888 for N-stage prediction and AUC of 0.875 for M-stage prediction. This demonstrates that patterns in routine clinical labs are substantially informative about metastatic status, even without imaging data.
Clinical Significance A model of this performance could serve as a clinical pre-screening tool to prioritize patients for staging workup, guide decisions about the extent of imaging required, or flag patients likely to have occult metastases missed on initial staging - all from data already collected as part of routine care.
Patient Population 1,629 lung cancer patients were included, spanning multiple histological subtypes (adenocarcinoma, squamous cell carcinoma, small cell lung cancer) and all disease stages. This diversity strengthens the model's potential generalizability but also introduces subtype-specific variation in laboratory patterns that models must learn to handle.
Input Features Input features included complete blood count parameters (WBC, neutrophil-to-lymphocyte ratio, platelet count), liver and kidney function tests, serum albumin, LDH, and lung cancer tumor markers (CEA, CYFRA21-1, NSE, SCC-Ag). These markers are routinely measured at diagnosis in most healthcare systems, requiring no additional tests.
ML Algorithms Compared The study compared logistic regression, random forest, XGBoost, AdaBoost, support vector machines, and neural networks. Each model was evaluated on both N-stage and M-stage prediction tasks using 5-fold cross-validation to ensure performance estimates were not inflated by overfitting to a single train-test split.
Feature Importance Analysis SHAP (SHapley Additive exPlanations) values were computed to identify which laboratory parameters contributed most to each model's predictions, providing interpretability and biological validation of the model's learned patterns.
N-Stage Prediction For lymph node metastasis prediction (N0 vs. N1-N3), logistic regression achieved AUC 0.888, outperforming random forest (AUC ~0.85), XGBoost, and AdaBoost. The simpler model's superiority over ensemble methods suggests the relationship between laboratory values and lymph node status is largely linear - complex nonlinear interactions add noise rather than signal with this feature set.
M-Stage Prediction For distant metastasis prediction, logistic regression again led with AUC 0.875. The slightly lower performance compared to N-stage prediction likely reflects the greater biological heterogeneity of distant metastasis patterns, which are more variable across different organ sites and tumor subtypes than lymph node involvement.
Key Predictive Features SHAP analysis identified neutrophil-to-lymphocyte ratio (NLR), CEA (carcinoembryonic antigen), CYFRA21-1, and serum albumin as the most influential predictors across both tasks. NLR captures systemic immune dysregulation associated with metastatic spread, while CEA and CYFRA21-1 are established lung cancer markers with known stage-dependent elevation patterns.
Subgroup Analysis Model performance was relatively consistent across histological subtypes for M-stage prediction but showed more variation for N-stage, with squamous cell carcinoma being somewhat better predicted than adenocarcinoma. This may reflect the more central location and predictable lymphatic spread patterns of squamous cell carcinoma.
Neutrophil-to-Lymphocyte Ratio NLR reflects the balance between innate immune activation and adaptive immune function. Elevated NLR indicates systemic inflammation and relative immunosuppression - conditions that facilitate cancer invasion and metastatic colonization. NLR is one of the most widely replicated prognostic markers in solid tumors and is consistently elevated in metastatic versus localized disease.
CEA and CYFRA21-1 These tumor-derived glycoproteins are shed into circulation by tumor cells in proportion to tumor burden and invasiveness. CEA elevation above 10 ng/mL and CYFRA21-1 above 3.3 ng/mL have established associations with lymph node involvement and distant spread in clinical guidelines. Their inclusion as top SHAP contributors validates that the model is capturing known biomarker biology rather than spurious correlations.
Serum Albumin Hypoalbuminemia is a marker of both poor nutritional status and systemic inflammation - both of which are more pronounced in metastatic disease. Low albumin also reflects the acute-phase response that accompanies more advanced cancer, making it an indirect systemic indicator of disease burden.
Platelet Count and LDH Thrombocytosis (elevated platelets) occurs in metastatic cancers partly because tumor cells stimulate platelet production and platelet activation facilitates metastatic seeding. LDH elevation is a marker of tissue destruction and anaerobic metabolism associated with rapidly growing or hypoxic tumor deposits, particularly in distant metastases.
Staging Workup Prioritization In settings where PET/CT or EBUS-guided bronchoscopy is capacity-constrained, a high N-stage probability from the model could fast-track patients to comprehensive staging workup. Conversely, very low probability patients could be managed with simpler imaging first, reducing unnecessary procedures.
Occult Metastasis Detection Some patients with apparently stage I or II disease on standard staging have occult metastases that are missed. A model predicting high M-stage probability despite negative conventional staging could prompt additional workup - whole-body MRI, additional CT coverage, or liquid biopsy - reducing upstaging surprises during treatment.
Low-Resource Settings In healthcare systems where advanced imaging is unavailable or unaffordable, a blood-test-based metastasis predictor provides actionable staging information at minimal additional cost. The required lab tests are available in even basic hospital settings globally.
Treatment Decision Support For borderline cases where the decision to add adjuvant chemotherapy hinges on the N2 lymph node status, a high-confidence ML prediction of nodal involvement could inform the discussion between the oncologist and patient - particularly when invasive confirmatory testing carries significant procedure risk.
Single-Center Limitation All 1,629 patients came from a single institution, creating potential for systematic biases in lab measurement protocols, patient demographics, and referral patterns that may not generalize. External validation on independent cohorts from different countries and healthcare systems is a prerequisite for broad clinical adoption.
Temporal Generalizability Laboratory reference ranges, assay platforms, and tumor marker assay calibrations change over time. A model trained on historical data may perform differently when deployed prospectively if systematic shifts in lab values occur due to assay changes or population-level health trends.
Missing Imaging Integration The model deliberately excludes imaging to demonstrate standalone value, but in practice, the highest-utility application would integrate lab predictions with imaging findings. A multimodal model combining blood markers with CT findings could substantially outperform either modality alone.
Future Directions Future work should validate the model prospectively in multi-institutional cohorts, explore whether dynamic changes in lab values over the treatment course predict treatment response or disease progression, and investigate whether the model can be extended to predict specific metastatic organ sites (bone, brain, liver, adrenal) which have different treatment implications.