Preoperative prediction of lymph node metastasis in lung adenocarcinoma based on tumor size and carcinoembryonic antigen Authors: Tao Lin, Jian Zhao, Feng Zhu, Tengfei Ge, Lei Tang, Dan Liu, Lu Wang, Hua Guo, Peng Qian, Wenglong Zhang, Xianbao Cao, Dongchun Ma Published: 2026 ABSTRACT BackgroundLymph node metastasis (LNM) is a critical determinant of staging and treatment decisions in lung adenocarcinoma. We aimed to develop and temporally validate a simple, clinically accessible preoperative model to predict LNM.MethodsA consecutive cohort of 6,406 patients with resected lung adenocarcinoma was chronologically divided into a training set (n=4,484) and a validation set (n=1,922). Independent predictors were identified using multivariable logistic regression. Model performance was assessed using the area under the curve (AUC), calibration analysis, and decision curve analysis (DCA). A nomogram was constructed.ResultsThe overall prevalence of lymph node (LN) positivity was 3.6% (233/6,406), with rates of 3.4% and 4.2% in the training and validation cohorts, respectively. In the multivariable analysis, only tumor size [odds ratio (OR) 1.09, 95% confidence interval (CI): 1.07–1.11, P<0.001] and carcinoembryonic antigen (CEA; OR 1.04, 95% CI: 1.02–1.06, P<0.001) were identified as independent predictors, whereas inflammatory and nutritional indices [the neutrophil-to-lymphocyte ratio (NLR), systemic immune-inflammation index (SII), prognostic nutritional index (PNI), and C-reactive protein-to-albumin ratio (CAR)] were not significant. The resulting two-variable model achieved strong discrimination, with AUC values of 0.873 (95% CI: 0.844–0.899) in the training cohort and 0.848 (95% CI: 0.806–0.886) in the validation cohort. At the Youden-derived threshold, the model yielded sensitivities of 81.0% and 83.8%, specificities of 80.1% and 71.2%, and negative predictive values (NPVs) of 99.2% and 99.0% in the training and validation cohorts, respectively. Calibration plots and the DCA demonstrated good agreement between predicted probabilities and observed outcomes and supported the clinical utility of the model.ConclusionsA two-variable model (incorporating tumor size and CEA) provides robust preoperative estimation of LN risk and may help reduce unnecessary invasive staging procedures. Highlight box Key findings • A simple logistic regression model based on tumor size and carcinoembryonic antigen (CEA) accurately predicts lymph node (LN) positivity with excellent negative predictive value and calibration. What is known and what is new? • Tumor size and CEA are recognized as potential risk factors for lymph node metastasis in lung adenocarcinoma. • In a large temporally validated cohort, we confirmed that tumor size and CEA outperform systemic inflammatory and nutritional indices for LN prediction, achieving an excellent negative predictive value. What is the implication, and what should change now? • The model can rule out lymph node metastasis preoperatively and may serve as a supportive low-risk stratification tool to aid clinical decision-making regarding invasive staging Introduction Lung cancer is the leading cause of cancer-related death worldwide, with approximately 2.5 million new cases and 1.8 million deaths annually, according to GLOBOCAN 2022 (1). Adenocarcinoma is the most common histological subtype of non-small cell lung cancer (NSCLC), accounting for approximately 40–50% of cases (2). Accurate staging of lung adenocarcinoma is critical for therapeutic decision-making and prognostication (3). Among staging parameters, lymph node metastasis (LNM) plays a central role, as the presence of nodal involvement often necessitates more extensive surgical resection, influences the use of adjuvant therapy, and is closely associated with long-term survival outcomes (4). Despite advances in imaging and invasive staging modalities, the preoperative prediction of LNM remains a major clinical challenge. Conventional chest computed tomography (CT) and positron emission tomography (PET) are valuable tools, but often fail to detect micrometastases, leading to under-staging (5). Conversely, false-positive findings may result in overtreatment. Endobronchial ultrasound-guided transbronchial needle aspiration (EBUS-TBNA) improves the diagnostic accuracy for lymph node metastasis but is invasive, operator-dependent, and not universally available (6). Thus, non-invasive, cost-effective, and accurate prediction tools urgently need to be developed. In recent years, the role of serum biomarkers in the prediction of LNM has been investigated. Carcinoembryonic antigen (CEA), one of the most widely used tumor markers, has been consistently associated with tumor aggressiveness, recurrence risk, and lymph node (LN) involvement in early-stage NSCLC (6,7). Similarly, systemic inflammatory and nutritional indices, including the neutrophil-to-lymphocyte ratio (NLR), systemic immune-inflammation index (SII), prognostic nutritional index (PNI), and C-reactive protein-to-albumin ratio (CAR), have emerged as prognostic indicators in various malignancies (8-10). Theoretically, systemic inflammation can foster an environment conducive to tumor dissemination by promoting angiogenesis and lymphangiogenesis, while nutritional impairment may reflect weakened host immune surveillance, potentially facilitating early nodal spread. These indices reflect the host immune and nutritional status, which are increasingly recognized as important contributors to tumor progression. Nevertheless, the incremental predictive value of inflammatory and nutritional indices over traditional tumor markers remains controversial (11-13). Several studies have reported associations between these indices and LNM, but findings have been inconsistent, often due to small cohorts, heterogeneous populations, and limited rigorous validation. Further, the integration of these indices into practical preoperative models remains limited. To address these gaps, we conducted a large-scale retrospective study of 6,406 patients with surgically resected lung adenocarcinoma. By systematically evaluating tumor-specific, inflammatory, and nutritional indices, and applying temporal validation, we aimed to identify robust predictors of LN positivity. Based on these findings, we developed a simple, clinically accessible logistic regression model and constructed a nomogram for bedside application. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1323/rc). Methods Patient selection We retrospectively reviewed consecutive patients with lung adenocarcinoma who underwent curative-intent surgical resection at Anhui Chest Hospital between January 2019 and February 2025. All data were derived from a prospectively maintained institutional clinical database, enabling rapid extraction and cleaning of eligible cases up to February 2025. The inclusion criteria were as follows: (I) pathologically confirmed lung adenocarcinoma; (II) availability of complete surgical pathology data, including lymph node (LN) status; and (III) preoperative laboratory examinations performed within one week before surgery. The exclusion criteria were as follows: a history of prior malignancies, receipt of neoadjuvant therapy, or missing essential clinical data. A total of 6,406 eligible patients were included in the final analysis. To facilitate temporal validation, the patients were ordered chronologically by admission date, with the first 70% (n=4,484) assigned to the training cohort and the remaining 30% (n=1,922) assigned to the validation cohort. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Institutional Review Board of Anhui Chest Hospital (No. KJ2025-041), and the requirement for individual informed consent was waived due to the retrospective nature of the study. Data collection Demographic and clinical data, including age, sex, and tumor size, were extracted from the electronic medical record system. Laboratory values collected within one week prior to surgery included complete blood counts, serum albumin, hypersensitive C-reactive protein (hs-CRP), and CEA. Pathological LN status was determined by postoperative histological examination and was defined as positive if at least one regional LN contained metastatic carcinoma. Variable definition All inflammatory and nutritional indices were derived from preoperative peripheral blood tests. The NLR was calculated as the absolute neutrophil count divided by the absolute lymphocyte count. The SII was defined as the platelet count multiplied by the neutrophil count, and divided by the lymphocyte count. The PNI was determined using the following formula: PNI = serum albumin (g/L) + 5 × lymphocyte count (109/L). The CAR was calculated as hs-CRP (mg/L) divided by serum albumin (g/L). Tumor size was defined as the longest diameter of the lesion on preoperative CT imaging, expressed in millimeters. Serum CEA levels were measured in ng/mL. LN positivity was defined as pathological examination confirming metastasis in one or more regional lymph nodes. Statistical analysis Continuous variables are presented as mean ± standard deviation (SD) or median with interquartile range (IQR) and were compared between groups using the Student’s t-test or Mann-Whitney U test, as appropriate. Categorical variables are expressed as counts and percentages and were compared using the Chi-squared test. Univariable logistic regression was performed to assess the association between each variable and LN positivity. Variables with P<0.05 in the univariable analysis were entered into multivariable logistic regression to identify independent predictors. Odds ratios (ORs) with 95% confidence intervals (CIs) were reported. A logistic regression prediction model was constructed based on independent predictors. Model performance was evaluated in terms of discrimination [area under the curve (AUC) of the receiver operating characteristic (ROC) curve], calibration (calibration plots with bootstrap resampling), and clinical utility [decision curve analysis (DCA)]. A nomogram was constructed to facilitate bedside application. Statistical significance was defined as two-sided P<0.05. Analyses were performed using R (version 4.2.1, https://www.r-project.org) and Python 3.6, including the scikit-learn and statsmodels packages. Results Baseline characteristics A total of 6,406 patients with surgically resected lung adenocarcinoma were included in the study, comprising 4,484 in the training cohort and 1,922 in the validation cohort. The overall prevalence of LN positivity was 3.6% (3.4% in the training cohort, and 4.2% in validation cohort). The baseline characteristics of the patients are summarized in Table 1. The two cohorts were generally comparable in terms of sex distribution, tumor size, and most laboratory indices. However, statistically significant differences were observed in age and several hematological parameters, likely reflecting the temporal split design. Table 1Baseline characteristics of patients in the training and validation cohortsVariablesTotal (n=6,406)Training cohort (n=4,484)Validation cohort (n=1,922)PAge (years)56.1±12.055.7±11.957.1±12.3<0.001Sex0.14 Male2,344 (36.6)1,614 (36.0)730 (38.0) Female4,062 (63.4)2,870 (64.0)1,192 (62.0)Tumor size (mm)11.0 (7.0–18.0)11.0 (7.0–17.0)11.0 (7.0–18.0)<0.001LN positive233 (3.6)153 (3.4)80 (4.2)0.16CEA (ng/mL)1.72 (1.11–2.67)1.79 (1.17–2.79)1.53 (0.97–2.39)<0.001NLR1.77 (1.36–2.39)1.76 (1.36–2.37)1.79 (1.37–2.42)<0.001SII351.7 (254.5–497.7)351.1 (253.6–495.4)352.5 (255.1–503.1)<0.001PNI49.2 (45.9–52.6)48.9 (45.5–52.2)50.2 (46.9–53.6)<0.001CAR0.02 (0.01–0.05)0.03 (0.01–0.05)0.02 (0.01–0.04)<0.001Data are presented as mean ± standard deviation, median (interquartile range) or n (%), as appropriate. CAR, C-reactive protein-to-albumin ratio; CEA, carcinoembryonic antigen; LN, lymph node; NLR, neutrophil-to-lymphocyte ratio; PNI, prognostic nutritional index; SII, systemic immune-inflammation index. Baseline characteristics of patients in the training and validation cohorts Data are presented as mean ± standard deviation, median (interquartile range) or n (%), as appropriate. CAR, C-reactive protein-to-albumin ratio; CEA, carcinoembryonic antigen; LN, lymph node; NLR, neutrophil-to-lymphocyte ratio; PNI, prognostic nutritional index; SII, systemic immune-inflammation index. Univariable and multivariable analysis Univariable logistic regression identified tumor size, CEA, NLR, SII, PNI, and CAR as significant predictors of LN positivity. In the multivariable analysis, only tumor size (OR 1.09, 95% CI: 1.07–1.11, P<0.001) and CEA (OR 1.04, 95% CI: 1.02–1.06, P<0.001) remained independent predictors of LN positivity (Table 2). These two parameters were thus selected for model construction. Table 2Univariable and multivariable logistic regression analysis for lymph node metastasisVariablesUnivariable MultivariableOR (95% CI)P valueOR (95% CI)P valueTumor size1.09 (1.07–1.11)<0.0011.09 (1.07–1.11)<0.001CEA1.04 (1.02–1.06)<0.0011.04 (1.02–1.06)<0.001NLR1.12 (1.01–1.24)0.031.05 (0.92–1.18)0.47SII1.09 (1.00–1.19)0.0481.02 (0.93–1.13)0.67PNI0.95 (0.91–0.99)0.020.98 (0.93–1.03)0.38CAR1.10 (1.01–1.21)0.031.03 (0.93–1.14)0.57CAR, C-reactive protein-to-albumin ratio; CEA, carcinoembryonic antigen; CI, confidence interval; NLR, neutrophil-to-lymphocyte ratio; OR, odds ratio; PNI, prognostic nutritional index; SII, systemic immune-inflammation index. Univariable and multivariable logistic regression analysis for lymph node metastasis CAR, C-reactive protein-to-albumin ratio; CEA, carcinoembryonic antigen; CI, confidence interval; NLR, neutrophil-to-lymphocyte ratio; OR, odds ratio; PNI, prognostic nutritional index; SII, systemic immune-inflammation index. Model performance The logistic regression model incorporating tumor size and CEA demonstrated robust discriminatory ability, yielding an AUC of 0.873 (95% CI: 0.844–0.899) in the training cohort and 0.848 (95% CI: 0.806–0.886) in the validation cohort (Figure 1). Calibration analysis showed close agreement between predicted probabilities and observed outcomes in both cohorts (Figure 2). When applying the Youden-derived threshold (0.325), the model achieved a sensitivity of 81.0%, specificity of 80.1%, and negative predictive value (NPV) of 99.2% in the training cohort, and a sensitivity of 83.8%, specificity of 71.2%, and NPV of 99.0% in the validation cohort. Detailed diagnostic performance metrics are summarized in Table 3. ROC curves of the prediction model in the training and validation cohorts. ROC curves demonstrating the discrimination ability of the model. The AUC was 0.873 in the training cohort and 0.848 in the validation cohort, indicating good predictive performance across both datasets. AUC, area under the curve; ROC, receiver operating characteristic. Calibration plots of the prediction model in the training and validation cohorts. Calibration plots comparing predicted and observed probabilities of lymph node metastasis. The dashed diagonal line represents perfect calibration. Both training and validation cohorts showed good agreement between predicted and observed outcomes, particularly in the clinically relevant low-risk range (predicted probability <0.3). Table 3Model performance and operating characteristics CohortAUC (95% CI)Sensitivity (%)Specificity (%)PPV (%)NPV (%) Training0.873 (0.844–0.899)81.080.112.699.2 Validation0.848 (0.806–0.886)83.871.211.299.0AUC, area under the curve; CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value. Model performance and operating characteristics AUC, area under the curve; CI, confidence interval; NPV, negative predictive value; PPV, positive predictive value. Clinical utility The DCA demonstrated that the model provided a greater net benefit compared with “treat-all” or “treat-none” strategies across a wide range of threshold probabilities (1–10%) (Figure 3). This finding indicates that the model is particularly suitable for rule-out screening in low-prevalence settings, helping to minimize unnecessary invasive staging procedures. DCA of the prediction model in the training and validation cohorts. DCA illustrating the net benefit of the prediction model compared with “treat-all” and “treat-none” strategies across a range of threshold probabilities (0.01–0.20). The model demonstrated superior net clinical benefit, supporting its potential utility for preoperative decision-making. DCA, decision curve analysis. Nomogram presentation A nomogram integrating tumor size and CEA was constructed to enable individualized risk estimation at the bedside (Figure 4). This tool provides clinicians with a straightforward graphical interface for estimating the probability of LN positivity and may aid in clinical decision-making regarding mediastinal staging and surgical planning. Nomogram based on tumor size and serum CEA (ng/mL) for preoperative prediction of lymph node metastasis in lung adenocarcinoma. Points for each predictor are summed to obtain a total score, which corresponds to the predicted probability on the bottom scale. The nomogram reflects the final two-variable logistic model (tumor size and CEA only). The probability axis is calibrated to the training cohort, so that the predicted risk at the cohort medians (tumor size, 11 mm; CEA, 1.79 ng/mL) equals the observed prevalence (3.4%). CEA, carcinoembryonic antigen; LN, lymph node. Discussion This comprehensive analysis of 6,406 patients with surgically resected lung adenocarcinoma, comprising a training cohort of 4,484 and a validation cohort of 1,922, identified tumor size and serum CEA as the only independent preoperative predictors of LNM. Although several inflammatory and nutritional indices, including the NLR, SII, PNI, and CAR, were significantly associated with LN positivity in the univariable analyses, these associations did not persist after multivariable adjustment. This finding underscores the dominant influence of tumor burden and tumor-derived biomarkers compared with systemic indices, and highlights the importance of multivariable modeling to avoid spurious conclusions (14). Our results are consistent with previous studies demonstrating the predictive role of tumor size and CEA in occult LNM. Bao et al. demonstrated that tumor size and serum markers were significant predictors of LN involvement in small-sized NSCLC (15). Similarly, Yu et al. reported that tumor size and CEA were independently associated with mediastinal nodal metastasis (16). Maeda et al. further showed that CEA levels >5 ng/mL were predictive of nodal spread in early-stage NSCLC (17). More recently, Wang et al. developed a PET/CT-based radiomics model incorporating CEA, confirming the value of this biomarker in predicting occult LNM (18). Compared with these earlier studies, the present work extends prior evidence by validating the predictive value of tumor size and CEA in a much larger cohort and by systematically evaluating the additional role of inflammatory and nutritional indices. We have clarified in the Discussion that while CEA is not a universal standard, maximizing its utility in regions where it is routinely collected is highly beneficial for risk stratification. Although systemic inflammatory indices such as the NLR and SII have been associated with poor outcomes in lung cancer (19,20), their predictive role for LNM remains uncertain. Our findings suggest that these indices may reflect a pro-tumor inflammatory milieu but are not independent drivers of nodal spread once tumor size and CEA are considered. Similar observations have been reported in other cancers, where the PNI and inflammation-based scores are associated with long-term survival but are less directly linked to local invasion or LNM (21,22). These results indicate that systemic inflammation may act as a background factor rather than a standalone determinant of lymph node metastasis and caution against relying on such indices for individualized LN risk stratification. However, the highly homogeneous and low-risk profile of our surgical cohort constrains the variability of systemic inflammatory and nutritional biomarkers, thereby attenuating their predictive efficacy. Such indices only deliver limited incremental predictive value within low-risk surgically resected lung adenocarcinoma cohorts, rather than lacking predictive utility for lymph node metastasis across all patient populations. The predictive model we constructed based on tumor size and CEA demonstrated strong performance, with an AUC of 0.873 in the training cohort and an AUC of 0.848 in the validation cohort. Calibration analysis showed close agreement between the predicted and observed risks. Notably, the model achieved consistently high NPVs of 99.2% in the training cohort and 99.0% in the validation cohort. These high NPVs are clinically meaningful, as patients identified as very low risk for LNM may safely avoid unnecessary invasive staging procedures such as mediastinoscopy or EBUS-TBNA, thereby reducing costs, procedural risks, and patient burden. While model sensitivity remained stable across cohorts (81.0% and 83.8%, respectively), specificity was lower in the validation cohort (71.2% vs. 80.1%). This variability suggests that external validation across broader populations is warranted before the widespread clinical adoption of the model. Notably, due to the temporal split design, several baseline characteristics differed significantly between the training and validation cohorts, including age, tumor size, and CEA levels (all P<0.001, Table 1). Such differences likely reflect real-world changes over time, including the increasing adoption of early detection and screening programs, which may have led to the diagnosis of smaller tumors, as well as evolving diagnostic and treatment practices. This phenomenon, often referred to as “cohort shift”, is important to acknowledge, as it may affect model performance when applied to later patient populations. Indeed, in our study, specificity decreased from 80.1% in the training cohort to 71.2% in the validation cohort, suggesting that changes in baseline distributions may have contributed to this reduction. Far from being a limitation, this finding underscores the necessity of rigorous temporal validation. By evaluating the model on a future dataset with shifted characteristics, we obtain a more realistic assessment of its robustness. At the same time, these results highlight the need for external multicenter validation before broad application, as further cohort differences across institutions and healthcare systems may similarly influence model calibration and clinical utility. The strengths of this study include its large sample size, use of routinely collected clinical and laboratory data, and temporal validation, which collectively enhance the credibility and practical relevance of the findings. By directly comparing systemic indices with tumor-specific markers, we were able to clarify their relative contributions to LN risk prediction. Moreover, the development of a nomogram enhances clinical usability by enabling individualized risk estimation in real-world settings. In real-world clinical settings and resource-limited regions where PET-CT and EBUS are not universally available, this nomogram serves as a cost-effective first-line risk-stratification tool with substantial clinical value. Nevertheless, several limitations merit consideration. First, the retrospective single-center design may limit the generalizability of the findings, and multicenter prospective validation is required to confirm these findings. Second, the overall prevalence of LN positivity was relatively low (3.6%). This reflects the highly selected nature of our surgical cohort, which predominantly consisted of early-stage tumors (median size 11 mm) and a high proportion of likely screen-detected ground-glass opacities. Consequently, the excellent negative predictive value of our model is partly driven by this low baseline prevalence. Unexpected bulky lymph node metastasis identified incidentally during surgery was extremely uncommon within this rigorously selected study cohort. Third, radiological features (e.g., consolidation-to-tumor ratio and radiomics-based parameters) and molecular biomarkers [e.g., epidermal growth factor receptor (EGFR) and Kirsten rat sarcoma viral oncogene homolog (KRAS) mutations] were not included, although previous studies suggest these may provide additional prognostic information. Furthermore, due to the retrospective nature of the database, exact data regarding the utilization of preoperative PET-CT and EBUS were not systematically captured. In our real-world practice, these modalities are highly selectively applied. Therefore, our nomogram is not intended to replace PET-CT or EBUS, but rather to serve as a cost-effective, upfront triage adjunct. Finally, while calibration was generally good, the model tended to overestimate risk at higher predicted probabilities, likely due to the limited number of events in this subgroup, underscoring the need for larger datasets to refine model performance in high-risk patients. The prevalence of LNM in our cohort was low (3.6%). This likely reflects the clinical characteristics of our study population, which included a high proportion of screen-detected adenocarcinomas with ground-glass opacity components. While this yielded a model with an excellent NPV (≥99%), it is important to emphasize that NPV is highly dependent on disease prevalence. In populations with higher baseline rates of LN positivity (e.g., patients with larger or more solid tumors), the NPV is expected to decrease, while the predictive value may improve (~11–12%). Clinicians should thus interpret and apply the model according to the pre-test probability of LNM in their target population. Conclusions In conclusion, this study demonstrates that a simple preoperative model based on tumor size and CEA provides robust and clinically relevant prediction of LNM in lung adenocarcinoma. The model shows excellent rule-out performance, enabling clinicians to identify low-risk patients and potentially serve as a supportive low-risk stratification tool to aid clinical decision-making regarding invasive staging. These findings establish a benchmark model based on widely available clinical variables, which may serve as a foundation for future refinements incorporating advanced imaging and molecular biomarkers. Prospective multicenter studies are warranted to further validate and extend its clinical utility. Supplementary The article’s supplementary files as Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1323/prf