Development and validation of a nomogram integrating multi-dimensional clinical factors for predicting lung cancer-related mediastinal/hilar lymph node metastasis before endobronchial ultrasound-guided transbronchial needle aspiration

Abstract

Background

Accurate preoperative identification of lung cancer-related mediastinal/hilar lymph node (LN) metastasis in patients with chest computed tomography (CT)-detected lymphadenopathy is critical for endobronchial ultrasound-guided transbronchial needle aspiration (EBUS-TBNA) and lung cancer treatment planning. Single clinical indicators lack sufficient accuracy for distinguishing benign from malignant lymphadenopathy, thus creating an urgent need for a comprehensive predictive tool integrating multidimensional routine clinical data. This study aimed to develop and validate a nomogram prediction model based on patient clinical characteristics, imaging features, serological markers, and endoscopic findings to predict the probability that mediastinal/hilar lymphadenopathy is attributable to lung cancer metastasis before EBUS-TBNA.

Methods

Clinical data of patients with CT-detected mediastinal/hilar lymphadenopathy who underwent EBUS-TBNA at the Affiliated Hospital of Southwest Medical University (January 2024–June 2025) were retrospectively collected. Following inclusion and exclusion screening, 298 patients were enrolled and categorized into the malignant group (n=173, pathologically confirmed as lung cancer metastasis) and the benign group (n=125, no evidence of malignant tumors on pathology and during ≥6 months of follow-up). The total cohort was stratified by LN pathology (benign vs. lung cancer metastasis) and randomly split into a training set (70%, n=208) and an internal validation set (30%, n=90). Univariate and multivariate logistic regressions were performed in the training set to identify independent predictors, and a nomogram was constructed. Model performance was evaluated by area under the curve (AUC), sensitivity, specificity, calibration (calibration curves, Hosmer-Lemeshow test), and decision curve analysis (DCA) in both sets.

Results

Seven independent predictors were identified: smoking history, CT-suspected malignancy, bronchoscopy-suspected mucosal invasion, LN short-axis diameter, carcinoembryonic antigen (CEA), neuron-specific enolase (NSE), and carbohydrate antigen 125 (CA125). The nomogram showed excellent discrimination (training set AUC =0.952, validation set AUC =0.932) and good calibration (Hosmer-Lemeshow test: training set, P=0.69; validation set, P=0.26). At optimal cut-offs, the sensitivity (81.0%/82.7%) and specificity (95.4%/94.7%) were high in both sets. DCA showed superior net clinical benefits over the extreme strategies.

Conclusions

We developed a clinically applicable nomogram that integrates routine clinical, imaging, serological, and endoscopic indicators to predict lung cancer-related mediastinal/hilar LN metastasis in patients with CT-detected lymphadenopathy. Using readily available clinical data, this model achieves excellent performance and clinical utility without reliance on advanced imaging [e.g., positron emission tomography/computed tomography (PET/CT)] or molecular testing, supporting precise preoperative risk stratification to guide EBUS-TBNA decision-making. It is especially suitable for patients who refuse or are ineligible for high-cost, high-radiation advanced examinations due to economic burden, radiation concerns, or medical contraindications. Additionally, a simplified nomogram was constructed as a complementary screening tool for resource-limited settings, further extending the overall clinical applicability of our predictive approach.

Highlight box.

Key findings

• We identified seven independent predictors of lung cancer-related mediastinal/hilar lymph node (LN) metastasis: smoking history, computed tomography-suspected malignancy, bronchoscopy-suspected mucosal invasion, LN short-axis diameter, and serum tumor markers (carcinoembryonic antigen, neuron-specific enolase, and carbohydrate antigen 125).

• The nomogram showed excellent discrimination (area under the curve =0.952 in the training set, 0.932 in the validation set) and good calibration (Hosmer-Lemeshow test: P=0.69 and 0.26, respectively), with superior net clinical benefit in decision curve analysis.

What is known and what is new?

• Endobronchial ultrasound-guided transbronchial needle aspiration (EBUS-TBNA) is the first-line staging method for lung cancer lymphadenopathy, but single indicator has poor accuracy for distinguishing benign/malignant lesions.

• This study innovatively integrates routine clinical, imaging, serological and endoscopic data to develop a nomogram that realizes individualized preoperative risk stratification without advanced testing.

What is the implication, and what should change now?

• This nomogram can act as a decision-support tool to optimize EBUS-TBNA use and improve invasive staging efficiency. Future multicenter prospective validation is required to confirm the generalizability of our findings.

Introduction

Lung cancer is one of the most common malignant tumors with the highest incidence and mortality worldwide, and its prognosis is closely related to the stage at diagnosis (1). Accurate assessment of mediastinal and hilar lymph node (LN) status is crucial for the clinical staging of lung cancer and formulation of treatment plans, which directly affects surgical decision-making, radiotherapy target delineation, and selection of systemic treatment strategies (2). Endobronchial ultrasound-guided transbronchial needle aspiration (EBUS-TBNA) has become the first-line minimally invasive method for obtaining mediastinal and hilar LN tissues for pathological evaluation. With the advantages of real-time performance, accuracy, and safety, its clinical application has become increasingly widespread (3).

However, in clinical practice, for patients with lymphadenopathy indicated by chest computed tomography (CT), it remains a great challenge to efficiently and accurately screen out those most likely to have LN metastasis (i.e., those who most require EBUS-TBNA for confirmation and staging) from a large number of suspected cases. Notably, lymphadenopathy in these patients may be caused by various etiologies, including lung cancer metastasis and benign inflammatory lesions. Although preoperative patient characteristics such as chest CT findings, serum tumor markers, and bronchoscopic manifestations have certain predictive value, their single use is insufficient in terms of predictive sensitivity or specificity, making it difficult to accurately differentiate metastatic lymphadenopathy from benign inflammatory lymphadenopathy (4-6). Several predictive models for lung cancer-related LN metastasis have been reported, such as those focusing on early-stage non-small cell lung cancer (NSCLC) patients (7) or integrating molecular markers (8), but most are designed for confirmed lung cancer cases rather than patients with unexplained lymphadenopathy, and some rely on advanced imaging [e.g., positron emission tomography/computed tomography (PET/CT)] (9) that may be inaccessible to some patients due to economic burden, radiation concerns, or medical contraindications. There is an urgent clinical need for a tool that can systematically integrate multidimensional and fragmented preoperative data and convert them into an intuitive and quantifiable individual risk assessment.

To address this gap, this study aimed to develop and validate a nomogram prediction model based on patient clinical characteristics, imaging features, serological markers, and endoscopic findings through a retrospective analysis. This model not only identifies key predictive factors, but also strives to provide a visualized quantitative tool for individually predicting the probability that mediastinal/hilar lymphadenopathy is attributable to lung cancer metastasis before EBUS-TBNA. Ultimately, it is expected to promote the transformation of clinical decision-making towards a precise risk assessment model based on multidimensional evidence. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1-0447/rc).

Methods

Participants

The clinical data of patients who underwent EBUS-TBNA at the Bronchoscopy Room of the Affiliated Hospital of Southwest Medical University from January 2024 to June 2025 were retrospectively collected. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Institutional Review Board of the Affiliated Hospital of Southwest Medical University (No. KY2026045). The requirement for informed consent was waived because of the retrospective nature of the study.

The inclusion criteria were as follows: (I) mediastinal or hilar LN enlargement on chest CT, with EBUS-TBNA targeted at these LNs; (II) serum tumor marker measurements including squamous cell carcinoma antigen (SCCA), pro-gastrin-releasing peptide (ProGRP), carcinoembryonic antigen (CEA), neuron-specific enolase (NSE), and carbohydrate antigen 125 (CA125) were completed prior to EBUS-TBNA.

The exclusion criteria were as follows: (I) a final diagnosis of malignant tumors other than primary lung cancer (e.g., lymphoma, metastatic lung tumors); (II) confirmed primary lung cancer with negative EBUS-TBNA LN findings, without further examinations to determine the mediastinal/hilar LN metastasis status; and (III) loss to follow-up during the study period.

Preoperative variable collection

All variables for analysis were clinically obtained prior to EBUS-TBNA, with relevant data retrospectively extracted from the electronic medical record system. A complete case analysis was adopted, with no imputed missing values. The variables were divided into four categories:

1. Demographic and clinical variables included age, sex, smoking history, comorbidities [diabetes mellitus, hypertension, coronary heart disease, chronic obstructive pulmonary disease (COPD)], and CT-suspected malignancy. Based on electronic medical record data, smoking history was categorized into “ever smoked” (including current and former smokers) and “never smoked”. All the patients underwent preoperative non-contrast or contrast-enhanced chest CT scans. Board-certified radiologists formulated preliminary judgments of “benign” or “malignant” in the reports, based on comprehensive assessment of imaging features of pulmonary lesions (including morphology, density, margin, pleural indentation, spiculation, and mediastinal/hilar LN characteristics, etc.) (10). Accordingly, we classified the patients into two categories: “CT-suspected malignancy” and “CT-not-suspected malignancy”.

2. Radiological variables: LN short-axis diameter was included as a variable. Specifically, the target LN was identified per the puncture site in the EBUS-TBNA report, and its LN short-axis diameter was independently measured on preoperative chest CT images by two observers blinded to the patients’ pathological results and prognostic outcomes (a radiologist and a respiratory physician). The mean of the two measurements was used for subsequent analysis.

3. Laboratory variables: included neutrophil-to-lymphocyte ratio (NLR) and serum tumor markers (CEA, NSE, CA125, ProGRP, and SCCA). NLR was calculated from the complete blood count (CBC). Positive thresholds for serum tumor markers were defined as the upper limits of the hospital’s reference ranges: CEA >6.00 ng/mL, NSE >15.7 ng/mL, CA125 >35 IU/mL, ProGRP >67.42 pg/mL, and SCCA >2.5 ng/mL. These serum tumor markers were then dichotomized into binary variables (positive vs. negative) based on the aforementioned thresholds for inclusion in the regression analysis.

4. Endoscopic variable: included bronchoscopy-suspected mucosal invasion. This variable was defined as visual evidence of malignant-related mucosal involvement observed under bronchoscopy (e.g., intraluminal growth, mucosal irregularities, or intraluminal bulge induced by extraluminal growth) (6) and was extracted from routine EBUS-TBNA operation reports.

EBUS-TBNA procedure and gold standard for outcome variable definition

EBUS-TBNA was performed by experienced bronchoscopists using Olympus equipment (electronic bronchoscope BF-260, endobronchial ultrasound bronchoscope BF-UC-260-FW, ultrasound host EU-ME1, and puncture needle NA-201SX-4021). All procedures were performed in accordance with the American College of Chest Physicians (ACCP) clinical guidelines (3,11,12). Based on preoperative chest imaging findings, the operators identified the target regions in advance, and all sonographically detectable mediastinal and hilar LNs were assessed. After assessing procedural safety and feasibility, suspicious LNs were selected for biopsy. The obtained specimens were sent for histopathological and cytopathological examinations; immunohistochemistry or tuberculosis-related tests were performed if necessary to confirm the diagnosis.

The gold standard for determining the outcome variable (mediastinal/hilar LN metastasis status) was defined as follows based on EBUS-TBNA pathological findings combined with long-term follow-up results:

1. Malignant group: patients with positive EBUS-TBNA pathology for lung cancer-related LN metastasis, or patients with initially negative EBUS-TBNA who were subsequently confirmed by other invasive diagnostic procedures (including surgical operation or repeat EBUS-TBNA).

2. Benign group: patients with no malignant lesions identified by EBUS-TBNA pathological examinations, and no evidence of malignancy detected after at least 6 months of clinical follow-up.

Statistical analysis

Statistical analyses were performed using SPSS (version 25.0; IBM Corp., Armonk, NY, USA) and R 4.5.2 (The R Foundation for Statistical Computing). Before model development, all eligible patients were stratified according to LN pathology (benign vs. lung cancer metastasis) and randomly divided into a training set (70%) for model construction and an internal validation set (30%) to maintain a consistent baseline distribution between the two subsets. The normality of continuous variables was tested using the Shapiro-Wilk test within each subgroup. Because individual continuous variables exhibited inconsistent normality across subgroups (i.e., normally distributed in some subgroups but non-normally distributed in others), all continuous variables were uniformly analyzed using nonparametric methods to ensure analytical consistency. Continuous variables are expressed as median (interquartile range) and categorical variables as numbers (percentages). Between-group comparisons were performed using the Mann-Whitney U test for continuous variables and the chi-square test or Fisher’s exact test for categorical variables.

In the training set, a univariate logistic regression analysis was initially performed to screen for candidate predictive variables. Variables with P<0.05 in the univariate analysis were included in the multivariate logistic regression using the enter method. Only independent predictive factors (P<0.05) in the multivariate logistic regression were used to construct the final nomogram using the rms package in R software. For improved clinical accessibility in resource-limited settings, a simplified nomogram was subsequently developed by manually selecting the most readily available and clinically practical variables from the independent predictors of the original model. The difference in the area under the receiver operating characteristic (ROC) curve between the original and simplified nomograms was compared using the DeLong test within the validation set. Model performance was evaluated in both the training and validation sets, including discrimination [ROC curve, area under the curve (AUC), optimal cut-off value determined by the Youden index, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV)], calibration (calibration curve, Hosmer-Lemeshow test, calibration regression intercept, and slope), and clinical utility [decision curve analysis (DCA)]. All statistical tests were two-sided, and statistical significance was set at P<0.05.

Results

Study population and baseline characteristics

A total of 488 patients underwent EBUS-TBNA at the bronchoscopy room of our center between January 1, 2024, and June 30, 2025. Among them, 348 patients met the inclusion criteria (111 patients were excluded due to missing serum tumor marker measurements, and 29 patients were excluded because EBUS-TBNA was targeted at lung masses rather than mediastinal/hilar LNs). According to the exclusion criteria, 10 patients were excluded due to a final diagnosis of malignant tumors other than primary lung cancer; 33 were excluded because, despite a confirmed diagnosis of primary lung cancer, their mediastinal/hilar LN metastasis status could not be definitively determined; and 7 were lost to follow-up during the study period. Ultimately, 298 patients were included in the study cohort (Figure 1).

Figure 1.

Flowchart of patient inclusion and exclusion. CA125, carbohydrate antigen 125; CEA, carcinoembryonic antigen; CT, computed tomography; EBUS-TBNA, endobronchial ultrasound-guided transbronchial needle aspiration; NSE, neuron-specific enolase; ProGRP, pro-gastrin-releasing peptide; SCCA, squamous cell carcinoma antigen.

Within this cohort, patients were stratified into malignant and benign groups based on pathological diagnosis and long-term follow-up outcomes. In the malignant group (n=173), adenocarcinoma was the most prevalent histological subtype (73 cases, 42.2%), followed by small cell carcinoma (65 cases, 37.6%) and squamous cell carcinoma (26 cases, 15.0%). All patients in the benign group (n=125) showed no evidence of malignancy on subsequent invasive diagnostic procedures or during a minimum follow-up period of 6 months.

To develop and validate our predictive model, the entire cohort was randomly divided into a training set (n=208, 70%) and a validation set (n=90, 30%) using stratified random sampling, stratified by LN pathology (benign vs. metastatic from lung cancer), to ensure balanced distribution of the key outcome variable between the two sets. The training set included 121 patients with pathologically confirmed mediastinal/hilar LN metastasis from primary lung cancer, whereas the validation set included 52 such patients.

Baseline demographic, clinical, and laboratory characteristics of the training and validation sets are presented in Table 1. No statistically significant differences were observed between the two sets in terms of sex, age, smoking history, comorbidities (diabetes mellitus, hypertension, coronary heart disease, and COPD), CT-suspected malignancy, LN short-axis diameter, bronchoscopy-suspected mucosal invasion, and levels of CEA, NSE, CA125, ProGRP, and NLR (all P>0.05). Only the serum SCCA levels showed a slight but statistically significant difference (P=0.03). Overall, these findings indicate that the baseline characteristics were well balanced between the training and validation sets, ensuring the reliability of subsequent model development and validation.

Table 1. Characteristics of patients in the two datasets.

Characteristic	Training set					Validation set				P
	Total (n=208)	Benign group (n=87)	Malignant group (n=121)	P		Total (n=90)	Benign group (n=38)	Malignant group (n=52)	P
Sex				<0.001*					0.16	0.86
Female	58 (27.9)	36 (41.4)	22 (18.2)			26 (28.9)	14 (36.8)	12 (23.1)
Male	150 (72.1)	51 (58.6)	99 (81.8)			64 (71.1)	24 (63.2)	40 (76.9)
Age (years)	61.0 [56.0–69.0]	60.0 [54.0–69.0]	62.0 [58.0–70.0]	0.12		62.0 [53.0–70.3]	58.5 [50.0–68.8]	64.5 [55.0–70.8]	0.21	0.52
Smoking history				<0.001*					0.27	0.69
No	101 (48.6)	55 (63.2)	46 (38.0)			46 (51.1)	22 (57.9)	24 (46.2)
Yes	107 (51.4)	32 (36.8)	75 (62.0)			44 (48.9)	16 (42.1)	28 (53.8)
Comorbidities
Diabetes mellitus	28 (13.5)	14 (16.1)	14 (11.6)	0.35		15 (16.7)	6 (15.8)	9 (17.3)	0.85	0.47
Hypertension	51 (24.5)	17 (19.5)	34 (28.1)	0.18		19 (21.1)	10 (26.3)	9 (17.3)	0.30	0.52
Coronary heart disease	9 (4.3)	4 (4.6)	5 (4.1)	>0.99		6 (6.7)	1 (2.6)	5 (9.6)	0.40	0.40
COPD	25 (12.0)	13 (14.9)	12 (9.9)	0.27		15 (16.7)	8 (21.1)	7 (13.5)	0.34	0.28
CT-suspected malignancy				<0.001*					<0.001*	0.87
No	49 (23.6)	45 (51.7)	4 (3.3)			22 (24.4)	20 (52.6)	2 (3.8)
Yes	159 (76.4)	42 (48.3)	117 (96.7)			68 (75.6)	18 (47.4)	50 (96.2)
LN short-axis diameter (mm)	19.5 [15.0–27.0]	16.0 [13.0–20.0]	22.0 [18.0–30.0]	<0.001*		20.0 [16.0–26.5]	17.0 [13.8–20.3]	24.0 [18.3–32.0]	<0.001*	0.48
Bronchoscopy-suspected mucosal invasion				<0.001*					<0.001*	0.40
No	121 (58.2)	72 (82.8)	49 (40.5)			57 (63.3)	33 (86.8)	24 (46.2)
Yes	87 (41.8)	15 (17.2)	72 (59.5)			33 (36.7)	5 (13.2)	28 (53.8)
CEA (ng/mL)				<0.001*					<0.001*	0.40
≤6	123 (59.1)	77 (88.5)	46 (38.0)			58 (64.4)	33 (86.8)	25 (48.1)
>6	85 (40.9)	10 (11.5)	75 (62.0)			32 (35.6)	5 (13.2)	27 (51.9)
NSE (ng/mL)				<0.001*					0.006*	0.14
≤15.7	105 (50.5)	64 (73.6)	41 (33.9)			37 (41.1)	22 (57.9)	15 (28.8)
>15.7	103 (49.5)	23 (26.4)	80 (66.1)			53 (58.9)	16 (42.1)	37 (71.2)
CA125 (IU/mL)				<0.001*					0.001*	0.32
≤35	126 (60.6)	71 (81.6)	55 (45.5)			60 (66.7)	33 (86.8)	27 (51.9)
>35	82 (39.4)	16 (18.4)	66 (54.5)			30 (33.3)	5 (13.2)	25 (48.1)
ProGRP (pg/mL)				<0.001*					<0.001*	0.72
≤67.42	146 (70.2)	75 (86.2)	71 (58.7)			65 (72.2)	36 (94.7)	29 (55.8)
>67.42	62 (29.8)	12 (13.8)	50 (41.3)			25 (27.8)	2 (5.3)	23 (44.2)
SCCA (ng/mL)				0.04*					0.40	0.03*
≤2.5	179 (86.1)	80 (92.0)	99 (81.8)			68 (75.6)	27 (71.1)	41 (78.8)
>2.5	29 (13.9)	7 (8.0)	22 (18.2)			22 (24.4)	11 (28.9)	11 (21.2)
NLR	3.7 [2.7–6.0]	3.4 [2.5–5.5]	4.0 [2.8–6.1]	0.15		3.5 [2.7–5.6]	3.4 [2.8–5.4]	3.9 [2.7–6.0]	0.70	0.87

Data are presented as n (%) or median [interquartile range]. *, P<0.05. CA125, carbohydrate antigen 125; CEA, carcinoembryonic antigen; COPD, chronic obstructive pulmonary disease; CT, computed tomography; LN, lymph node; NLR, neutrophil-to-lymphocyte ratio; NSE, neuron-specific enolase; ProGRP, pro-gastrin-releasing peptide; SCCA, squamous cell carcinoma antigen.

Variable selection for predictive model construction

The results of the univariate and multivariate logistic regression analyses of the training set are presented in Table 2. Univariate analysis revealed significant associations between multiple variables and primary lung cancer-related mediastinal or hilar LN metastasis (vs. benign LN enlargement), including sex, smoking history, CT-suspected malignancy, LN short-axis diameter, bronchoscopy-suspected mucosal invasion, CEA, NSE, CA125, ProGRP, and SCCA (all P<0.05). Subsequent multivariable logistic regression analysis identified seven independent predictors of primary lung cancer-related mediastinal/hilar LN metastasis (vs. benign LN enlargement): smoking history [odds ratio (OR) =5.998; 95% confidence interval (CI): 1.467–24.533; P=0.01], CT-suspected malignancy (OR =32.980; 95% CI: 6.091–178.570; P<0.001), bronchoscopy-suspected mucosal invasion (OR =3.481; 95% CI: 1.194–10.148; P=0.02), LN short-axis diameter (OR =1.108; 95% CI: 1.027–1.195; P=0.008), CEA (OR =10.600; 95% CI: 3.241–34.670; P<0.001), NSE (OR =2.841; 95% CI: 1.004–8.033; P=0.049), and CA125 (OR =5.724; 95% CI: 1.797–18.231; P=0.003). Sex, ProGRP, and SCCA were not identified as independent predictors in the multivariable model (all P>0.05).

Table 2. Univariate and multivariate logistic regression analysis in the training set.

Variables	Univariable analysis				Multivariable analysis
	OR	95% CI	P value		OR	95% CI	P value
Sex	3.176	1.694–5.957	<0.001*		0.729	0.175–3.041	0.67
Age	1.024	0.997–1.052	0.09
Smoking history	2.802	1.585–4.954	<0.001*		5.998	1.467–24.533	0.01*
Diabetes mellitus	0.682	0.307–1.516	0.35
Hypertension	1.609	0.830–3.119	0.16
Coronary heart disease	0.894	0.233–3.432	0.87
COPD	0.627	0.271–1.449	0.28
NLR	1.020	0.951–1.095	0.57
CT-suspected malignancy	31.339	10.625–92.439	<0.001*		32.980	6.091–178.570	<0.001*
Bronchoscopy-suspected mucosal invasion	7.053	3.630–13.704	<0.001*		3.481	1.194–10.148	0.02*
LN short-axis diameter	1.156	1.097–1.218	<0.001*		1.108	1.027–1.195	0.008*
CEA	12.554	5.906–26.688	<0.001*		10.600	3.241–34.670	<0.001*
NSE	5.429	2.958–9.965	<0.001*		2.841	1.004–8.033	0.049*
CA125	5.325	2.780–10.198	<0.001*		5.724	1.797–18.231	0.003*
ProGRP	4.401	2.167–8.941	<0.001*		2.901	0.842–9.992	0.09
SCCA	2.540	1.032–6.247	0.042*		0.743	0.138–4.015	0.73

*, P<0.05. CA125, carbohydrate antigen 125; CEA, carcinoembryonic antigen; CI, confidence interval; COPD, chronic obstructive pulmonary disease; CT, computed tomography; LN, lymph node; NLR, neutrophil-to-lymphocyte ratio; NSE, neuron-specific enolase; OR, odds ratio; ProGRP, pro-gastrin-releasing peptide; SCCA, squamous cell carcinoma antigen.

Development of the predictive nomogram

Based on the seven independent predictors identified by multivariable logistic regression analysis in the training set, a nomogram was developed to predict the probability that mediastinal/hilar lymphadenopathy detected by chest CT was attributable to lung cancer metastasis (Figure 2). This nomogram integrated clinical and laboratory variables, including smoking history, CT-suspected malignancy, bronchoscopy-suspected mucosal invasion, LN short-axis diameter, CEA, NSE, and CA125. For application, a point value was assigned to each variable according to its specific value or category for each patient. The total points were calculated by summing the individual point values, and the probability that the patient’s mediastinal/hilar lymphadenopathy was caused by lung cancer metastasis could be estimated by mapping the total points to the risk axis at the bottom of the nomogram.

Figure 2.

Nomogram for prediction of mediastinal/hilar lymphadenopathy related to lung cancer metastasis before EBUS-TBNA. CA125, carbohydrate antigen 125; CEA, carcinoembryonic antigen; CT, computed tomography; EBUS-TBNA, endobronchial ultrasound-guided transbronchial needle aspiration; LN, lymph node; NSE, neuron-specific enolase.

Performance validation of the predictive nomogram

The performance of our predictive nomogram for differentiating lung cancer metastasis-related mediastinal/hilar lymphadenopathy from benign lymphadenopathy was comprehensively validated in the training and validation sets, with key metrics summarized in Table 3. ROC curve analysis (Figure 3A) demonstrated excellent discriminative ability, with an area under the ROC curve (AUC) of 0.952 (95% CI: 0.925–0.978) in the training set and 0.932 (95% CI: 0.884–0.981) in the validation set, indicating robust capability to distinguish between the two conditions. At the optimal cutoff values (0.794 for the training set and 0.599 for the validation set), the nomogram maintained high sensitivity (training set: 0.810, 95% CI: 0.729–0.876; validation set: 0.827, 95% CI: 0.697–0.918) and specificity (training set: 0.954, 95% CI: 0.886–0.987; validation set: 0.947, 95% CI: 0.823–0.994). Additionally, favorable PPVs (0.961, 95% CI: 0.903–0.989 in the training set, 0.956; 95% CI: 0.849–0.995 in the validation set) and NPVs (0.783, 95% CI: 0.692–0.857 in the training set; 0.800, 95% CI: 0.654–0.904 in the validation set) further supported its reliability in clinical decision-making. Calibration assessment (Figure 3B,3C) revealed that the nomogram’s predicted probabilities were well aligned with actual observed outcomes, and the calibration regression intercepts and slopes were close to the ideal values of 0 and 1, respectively. Hosmer-Lemeshow test results (P=0.69 for the training set; P=0.26 for the validation set) confirmed no significant discrepancy between predicted and observed probabilities. DCA (Figure 3D) further verified the nomogram’s clinical utility, as it provided higher net benefit than the extreme strategies of “treating all” or “treating none” across a broad range of threshold probabilities in both sets. Collectively, these results confirmed that the nomogram exhibited excellent discrimination, good calibration, and substantial clinical utility for preoperative risk stratification before EBUS-TBNA.

Table 3. Performance metrics of the predictive nomogram.

Dataset	AUC (95% CI)	Optimal cut-off value	Sensitivity (95% CI)	Specificity (95% CI)	PPV (95% CI)	NPV (95% CI)	Intercept (95% CI)	Slope (95% CI)	H-L P value
Training set	0.952 (0.925–0.978)	0.794	0.810 (0.729–0.876)	0.954 (0.886–0.987)	0.961 (0.903–0.989)	0.783 (0.692–0.857)	0.004 (−0.067 to 0.076)	0.993 (0.891–1.095)	0.69
Validation set	0.932 (0.884–0.981)	0.599	0.827 (0.697–0.918)	0.947 (0.823–0.994)	0.956 (0.849–0.995)	0.800 (0.654–0.904)	0.063 (−0.053 to 0.178)	0.941 (0.770–1.111)	0.26

AUC, area under the curve; CI, confidence interval; H-L, Hosmer-Lemeshow; NPV, negative predictive value; PPV, positive predictive value.

Figure 3.

Performance validation of the predictive nomogram. (A) ROC curves in the training and validation sets. (B) Calibration curve of the nomogram in the training set. (C) Calibration curve of the nomogram in the validation set. (D) DCA of the nomogram in the training and validation sets. AUC, area under the curve; CI, confidence interval; DCA, decision curve analysis; ROC, receiver operating characteristic.

Development and validation of a simplified nomogram for resource-limited settings

A simplified nomogram was established using three easily accessible variables (CT-suspected malignancy, LN short-axis diameter, and smoking history) for broader clinical use in resource-limited scenarios (Figure S1). ROC curve analysis revealed favorable discriminative capacity, with an AUC of 0.874 (95% CI: 0.826–0.922) in the training set and 0.878 (95% CI: 0.811–0.946) in the validation set (Figure S2A). At the optimal cutoff values of 0.645 (training set) and 0.396 (validation set), the simplified nomogram achieved acceptable sensitivity of 0.769 (95% CI: 0.683–0.840) and 0.942 (95% CI: 0.841–0.988), as well as specificity of 0.828 (95% CI: 0.732–0.900) and 0.632 (95% CI: 0.460–0.782) in the training set and validation set, respectively. It also yielded favorable PPV (0.861, 95% CI: 0.781–0.920; 0.778, 95% CI: 0.655–0.873) and negative predictive value (NPV: 0.720, 95% CI: 0.621–0.805; 0.889, 95% CI: 0.708–0.976) (Table S1). Calibration assessment demonstrated excellent concordance between predicted and observed probabilities (Figure S2B,S2C), and Hosmer-Lemeshow tests confirmed acceptable model goodness-of-fit (training set: P=0.08; validation set: P=0.28). DCA further supported its clinical net benefit across a wide range of decision thresholds compared with extreme strategies (Figure S2D). The DeLong test performed in the validation set showed a significant AUC difference between the original and simplified nomograms (P=0.042), indicating that the original model possessed slightly superior predictive efficacy.

Discussion

This study developed a nomogram model based on four dimensions of patient characteristics (clinical characteristics, imaging features, serological markers, and endoscopic findings) to predict the probability that lymphadenopathy is attributable to lung cancer metastasis in patients with radiologically detected mediastinal/hilar lymphadenopathy prior to EBUS-TBNA. All the included variables were routinely and easily obtained in clinical practice. Internal validation demonstrated that the nomogram exhibited favorable predictive performance, with excellent discriminative ability, satisfactory calibration, and high sensitivity in identifying LN metastasis from lung cancer. Furthermore, the DCA confirmed that the model achieved a favorable net clinical benefit, supporting its practical clinical applicability.

In the clinical management of lung cancer, EBUS-TBNA serves as a key procedure for determining the nature of LNs (3). In current clinical practice, decisions regarding puncture mostly rely on single indicators (e.g., short-axis diameter of LNs) or clinicians’ empirical judgment, leading to high subjectivity and unsatisfactory accuracy of risk stratification. By integrating multi-source and fragmented preoperative data into an intuitive quantitative risk score, the nomogram provides an objective and quantitative decision-making tool. It enables clinicians to identify high-risk patients with lung cancer LN metastasis before EBUS-TBNA, optimize puncture strategies, reduce missed diagnoses of high-risk cases, and facilitate more accurate and targeted performance of EBUS-TBNA, thereby laying a reliable foundation for subsequent treatment decisions.

The optimal cutoff value of the model in the validation set was 0.599. DCA confirmed its favorable net clinical benefit across a threshold range of 0–0.8, with superior performance to the “treat all” strategy within 0–0.6, highlighting its value in the clinical “gray zone”. Based on the DCA results and the validation set cutoff, we propose a practical risk stratification: low-risk (<0.4, defer EBUS-TBNA with close follow-up), intermediate-risk (0.4–0.7, shared decision-making), and high-risk (>0.7, prioritized EBUS-TBNA). This framework is flexible for center-specific resources and patient preferences and serves as a decision-support tool rather than a substitute for clinical judgment. The clinical implications of each enrolled predictor are further analyzed below.

The predictors included in the model do not exist in isolation; they reflect the biological behavior and clinical status of the tumor from different dimensions, and jointly constitute a structured evaluation system. As a well-established clinical risk factor, smoking history may be associated with a more aggressive molecular phenotype of tobacco-related lung cancer (13,14), serving as a baseline element for risk assessment. For imaging features, this study intentionally incorporated qualitative CT assessment provided by radiologists (“CT-suspected malignancy”). This indicator was selected mainly because the conclusion of a CT report is intuitive, easily accessible, and readily available to clinicians without complex additional measurements, making it highly consistent with daily clinical practice. Combined with the quantitative measurement of the short-axis diameter of LNs (LN short-axis diameter, mm), CT-suspected malignancy serves as the most direct and sensitive basis for initial screening in clinical practice (15,16), whereas the short-axis diameter of LNs provides an objective and continuous anatomical indicator (17). Their complementarity enhances the reliability of the morphological evaluation. The model also incorporated a combination of serum tumor markers, including CEA, NSE, and CA125. These markers cover different histological subtypes of lung cancer, and their abnormal elevation provides supporting serological evidence for tumor activity and potential metastatic risk (5,18). Notably, CA125 is generally regarded as a biomarker of tumor burden or serosal involvement, rather than a specific indicator of LN metastasis. However, several recent studies have identified CA125 as an independent predictive factor (8,19), and its biological association with mediastinal/hilar LN metastasis remains incompletely understood. Bronchoscopy-suspected mucosal invasion is a direct endoscopic sign of local tumor aggressiveness and was included as a predictive factor. As a marker of central lesion involvement and local invasiveness, it supports the pathophysiological link between local infiltration and increased risk of regional LN metastasis (20,21). In clinical practice, this indicator can be used directly in patients with a prior preoperative bronchoscopy. For patients scheduled for simultaneous bronchoscopy and EBUS-TBNA, mucosal status can be preliminarily inferred from lesion location and features on CT and provisionally regarded as negative if CT findings are indeterminate. Of note, LN location (mediastinal vs. hilar) was not introduced as a candidate variable in the predictive modeling process. Our study aimed to conduct a unified risk assessment for all suspicious mediastinal and hilar LNs, without distinguishing based on anatomical site, and we therefore selected universal clinical and imaging indicators for model construction.

Compared with other recently published prediction models, our nomogram has a distinct research positioning and unique clinical value and can enrich the preoperative evaluation system for lung cancer-related lymphadenopathy from different perspectives. Specifically, the model by Han et al. (9) was developed for patients with confirmed lung cancer to predict LN metastasis risk using PET/CT metabolic information for lung cancer staging (AUC 0.883/0.877), providing an effective staging tool for centers with PET/CT availability. In contrast, this study focused on patients with unexplained mediastinal/hilar lymphadenopathy, aiming to preoperatively differentiate whether the lymphadenopathy was attributable to lung cancer metastasis. Our model relied only on routine and easily accessible clinical data, making it more practical for patients who cannot access or refuse high-cost advanced imaging (e.g., PET/CT) due to economic constraints, radiation concerns, or medical contraindications. These two models target different clinical scenarios and are complementary in clinical practice.

The model by Chen et al. (7) was constructed for patients with confirmed clinical stage IA peripheral NSCLC, focusing on precise staging of early-stage lung cancer by incorporating serum CA19-9 as one of the independent predictors for LN metastasis (AUC 0.78). The present study was not restricted to a lung cancer population, but instead enrolled patients with lymphadenopathy resulting from various etiologies, with the core goal of differentiating malignant metastasis from benign LN lesions. Our model covers a broader population and is more consistent with real-world scenarios of initial clinical screening. Furthermore, the application of a multi-dimensional serum tumor marker panel helps account for the heterogeneity of lung cancer and allows rapid differential diagnosis, thus complementing the staging model by Chen et al.

A study by Mei et al. (8), also based on lung cancer patients, integrated molecular and metabolic markers to predict LN metastasis (AUC 0.846/0.828), representing a cutting-edge direction for precise lung cancer staging. Our study focused on immediate preoperative differentiation and only used routine indicators available before EBUS-TBNA without relying on molecular testing, which is more suitable for rapid clinical triage. The work by Mei et al. also provides an optimization strategy for our model: adding molecular markers to the benign and malignant differentiation model established in this study may further improve differentiation accuracy in the future.

To further enhance the clinical accessibility of the predictive model in resource-limited settings, we developed a simplified nomogram incorporating only three readily available clinical indicators: smoking history, CT-suspected malignancy, and LN short-axis diameter. The simplified model maintained acceptable discriminative ability, calibration, and clinical net benefit, and exhibited high sensitivity of 0.942 (95% CI: 0.841–0.988) and relatively low specificity of 0.632 (95% CI: 0.460–0.782) in the validation set. This characteristic renders it ideal for preliminary screening, allowing the exclusion of low-risk patients and reduction of unnecessary referrals. The DeLong test confirmed that the original model had superior discriminative performance compared with the simplified model (P=0.042), reflecting a reasonable trade-off between predictive precision and clinical accessibility. The simplified nomogram is not an alternative to the original model, but a complementary triage tool for outpatient initial screening and resource-limited settings. In contrast, the original model is more suitable for clinical scenarios with complete preoperative workup to achieve refined risk stratification. This dual-model strategy can adapt to diverse clinical settings and resource conditions, further expanding the clinical applicability of the predictive approach developed in this study.

Despite their favorable predictive performance, both the original and simplified nomograms have several limitations. First, this single-center retrospective study only enrolled patients who underwent EBUS-TBNA based on routine clinical indications, representing a highly selected population with relatively higher clinical and imaging suspicion of malignancy. Such selection bias may restrict the direct generalizability of our model to the broader, unselected population of all patients with mediastinal/hilar lymphadenopathy detected on chest CT. Second, to ensure analytical quality, we only enrolled patients with complete routine clinical data and did not include individuals with missing tumor marker or clinical information, which may restrict sample representativeness. Third, the model performance was internally validated only (training and validation sets), requiring confirmation in an independent prospective external cohort for real-world applicability. Fourth, predictive variables were limited to routine clinical indicators, excluding high-level features [e.g., PET/CT standardized uptake value (SUV), circulating tumor DNA (ctDNA), and molecular markers] that may add predictive value. Additionally, subjective radiologist assessment (e.g., “CT-suspected malignancy”) may lead to interobserver variability. Fifth, the model is only applicable to patients with radiologically detected mediastinal/hilar lymphadenopathy for metastatic risk stratification, and is not intended for screening or diagnosing primary lung cancer, nor is it designed for assessing the metastatic risk of normal-appearing LNs on CT scan. Sixth, although a 6-month follow-up period was adopted to define benign LNs according to clinical consensus, we acknowledge that this duration may be insufficient to definitively exclude extremely indolent pulmonary malignancies. Given their inherent biological behavior with a low risk of mediastinal or hilar LN metastasis, their potential impact on the predictive model is minimal. Future studies will perform multi-center prospective validation and integrate advanced features to develop “basic” and “enhanced” models for different clinical settings, thereby improving accuracy and utility.

Conclusions

For patients with radiologically detected mediastinal/hilar lymphadenopathy, we developed and internally validated a nomogram to predict the risk of LN metastasis from lung cancer. This nomogram was established using clinical, imaging, serological, and endoscopic variables that are routinely available prior to EBUS-TBNA. By integrating multidimensional data, the model enables accurate risk stratification, and is particularly suitable for clinical scenarios in which patients decline high-cost, high-radiation advanced examinations, or in which such examinations are inaccessible due to individual or institutional factors. It provides objective, quantifiable decision support for clinicians prior to EBUS-TBNA, helps improve diagnostic yield and efficiency, and represents a practical tool for the precise and individualized management of lung cancer-related LN lesions. Additionally, a simplified nomogram with only three readily available clinical indicators was developed as a complementary screening tool for resource-limited settings, further extending the clinical applicability of our predictive strategy.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was approved by the Institutional Review Board of the Affiliated Hospital of Southwest Medical University (No. KY2026045). The requirement for informed consent was waived because of the retrospective nature of the study.

Data Sharing Statement

Available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-1-0447/dss

jtd-18-04-399-dss.pdf ^{(68.2KB, pdf)}

DOI: 10.21037/jtd-2026-1-0447