This study retrospectively enrolled 615 patients with LUAD, including 401 with BM and 214 without BM, matched by age and sex. Between June 2021 and June 2024, patients with LUAD who developed BM were consecutively identified from the medical records of the First Affiliated Hospital of Guangzhou Medical University (Guangzhou, China); a control group of patients with LUAD without BM was subsequently selected from the same institution during the same period, matched to the BM group by age and sex. LUAD diagnosis was established by a pathologist and respiratory physicians based on histopathological examination of lung biopsy specimens. BM diagnosis was established by a multidisciplinary team comprising neurologists, radiologists and respiratory physicians, based on findings from contrast-enhanced 3.0T MRI or detection of tumor cells in the cerebrospinal fluid. All patients in the non-BM group also underwent contrast-enhanced 3.0T MRI, with the absence of BM confirmed on the basis of negative imaging findings, thereby ensuring that BM status was determined by standardized imaging assessment across all enrolled patients. The primary inclusion criterion was a confirmed diagnosis of primary LC in the absence of other concurrent malignancies. The exclusion criteria were as follows: i) History of psychiatric complaints, disturbance of consciousness, dementia, central nervous system (CNS) infection or other conditions associated with impaired brain function; ii) severe hepatic or renal dysfunction; and iii) absence of key diagnostic examinations required for clinical diagnosis. All participants provided written informed consent for the collection, processing and analysis of blood samples and clinical data for research and publication purposes. This study was conducted in accordance with the principles of the Declaration of Helsinki and was approved by the Ethics Committee of the First Affiliated Hospital of Guangzhou Medical University (approval no. GZMC2021-06-1237; Fig. S1).

Blood samples from patients in the BM group were collected within 3 days of BM diagnosis; this sampling window was intended to ensure that collection preceded the initiation of BM-specific treatments, including corticosteroids, whole-brain radiotherapy and stereotactic radiosurgery. Blood samples from patients in the non-BM group were collected at the time of initial LUAD diagnosis, prior to the initiation of any antitumor therapy, representing a treatment-naive baseline state. All samples were collected in the morning under fasting conditions and stored in anticoagulant-containing polypropylene tubes. Plasma was separated by centrifugation at 500 × g for 10 min at 4°C. Serum samples were processed for LC biomarker detection within 24 h of collection; remaining samples were stored at −80°C. The LC biomarkers assessed included neuron-specific enolase (NSE), carcinoembryonic antigen (CEA), cancer antigen 125 (CA125), CA153 and cytokeratin 19 fragment (CYFRA21-1). CEA concentrations were measured by electrochemiluminescence immunoassay using a commercial kit (Architect CEA Reagent Kit; Abbott Diagnostics Division), whereas NSE, CA125, CA153 and CYFRA21-1 concentrations were determined by immunoradiometric assay using commercially available kits (Cytokeratin 19 Fragment IRMA Kit and NSE IRMA Kit; Beckman Coulter, Inc.). The procedure followed the manufacturer's instructions, as reported previously (14). The upper reference limits were as follows: NSE, 16.30 ng/ml; CEA, 5.00 ng/ml; CA125, 35.00 U/ml; CA153, 25.00 U/ml; and CYFRA21-1, 3.30 ng/ml. Each sample was measured in triplicate and the mean value was used for subsequent analyses. No samples were subjected to repeated freeze-thaw cycles.

Total DNA was extracted from formalin-fixed, paraffin-embedded tumor tissue using a QIAamp DNA Tissue Kit (Qiagen GmbH), following previously described methods (15). DNA purity was assessed by spectrophotometric measurement of the 260/280 nm absorbance ratio using a NanoDrop 2000 spectrophotometer (Thermo Fisher Scientific, Inc.). Targeted next-generation sequencing was performed by Burning Rock Biotech, a commercial clinical laboratory certified by the Clinical Laboratory Improvement Amendments and accredited by the College of American Pathologists (15). A commercial capture panel targeting 520 genes of the human genome (OncoScreen Plus; Burning Rock Biotech) was used for target enrichment according to the manufacturer's protocol. DNA fragments of 200–400 bp were purified using a Bioanalyzer 2100 (Agilent Technologies, Inc.), hybridized with capture probes, selected with magnetic beads and amplified. Library sequencing was performed on the NextSeq 500 platform (Illumina, Inc.), with a mean sequencing depth of ~1,000 ×. Further methodological details were reported previously (15).

A total of 9 candidate variables were initially entered into a multivariate analysis, encompassing TNM staging components (T stage, N stage and other tissue metastases as a surrogate for distant metastatic burden), EGFR exon 19 deletion (19del) mutation status, and five serum biomarkers (NSE, CEA, CA125, CA153 and CYFRA21-1). M stage was excluded to avoid collinearity with the outcome variable, as BM itself constitutes M1 disease by definition. Bidirectional stepwise selection based on the Akaike information criterion (AIC) was applied to identify the optimal variable subset; T stage was eliminated during this process and eight variables were retained in the final model (AIC=439.42). A nomogram was subsequently constructed using these variables to predict BM risk in patients with LUAD (16). Calibration curve and receiver operating characteristic (ROC) curve analyses were used to evaluate the calibration and discriminatory ability of the nomogram, respectively, and decision curve analysis (DCA) was performed to assess its clinical utility. Subsequently, 10-fold cross-validation was performed for the internal validation of the nomogram. The original dataset was randomly partitioned into 10 equal subsets; at each iteration, 9 subsets were pooled for model training and the remaining subset was withheld for validation, with this process repeated 10 times. The mean area under the ROC curve (AUC), sensitivity and specificity across the 10 validation iterations were calculated to assess the generalizability of the model. A web-based calculator was also developed (https://clinicalnomogram.shinyapps.io/BMsPredictionApp/) to facilitate the clinical application of the nomogram, providing individualized BM probability estimates based on patient-specific variable inputs.

Statistical analyses were performed using GraphPad Prism (version 8.0.2; Dotmatics), SPSS (version 19.0; SPSS, Inc.) and R (version 4.1.0; R Foundation for Statistical Computing). Key R packages used included rms, pROC, ResourceSelection and caret. Non-normally distributed data are presented as median (interquartile range), while normally distributed data are expressed as mean ± standard deviation. Categorical data were expressed as n (%). Group comparisons were performed using the Mann-Whitney U test, Kruskal-Wallis test or chi-square test, as appropriate for the data distribution. Logistic regression was used to identify factors independently associated with BM in patients with LUAD. ROC curves and the corresponding area under the curve (AUC) values were generated to evaluate and compare the predictive performance of individual and combined markers for BM. The net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were calculated to quantitatively compare the discriminative ability of the nomogram with that of the TNM staging model. The optimal cut-off value for each serum biomarker was determined by maximizing the Youden index (sensitivity + specificity-1) on the training set ROC curves: NSE, 18.18 ng/ml; CEA, 6.01 ng/ml; CA125, 19.59 U/ml; CA153, 14.82 U/ml; and CYFRA21-1, 4.17 ng/ml. A post-hoc power analysis was conducted for the primary comparison of BM incidence between 19del and non-19del groups within a chi-square test framework (Cohen's w=0.10; N=615; achieved power, 70.3%; α=0.05). A two-sided P<0.05 was considered to indicate statistical significance for all analyses.

In accordance with the inclusion and exclusion criteria, 401 patients with LUAD with BM and 214 patients without BM were recruited, matched by age [BM group: Median, 61 years (range, 24–88 years); non-BM group: Median, 60 years (range, 30–93 years)] and sex [BM group: 55.4% male; non-BM group: 49.5% male] and were well-balanced with respect to smoking history [BM group: 9.2% (37/401); non-BM group: 9.3% (20/214); P=0.961; Table SI]. Demographic and clinicopathological characteristics, including pathologic stage, primary tumor status, lymph node metastasis and number of BMs, were recorded for all enrolled patients. The proportion of patients with other tissue metastases was significantly higher in the BM group than in the non-BM group (54.1 vs. 22.0%, P<0.001), whereas the overall distribution of EGFR mutations did not differ significantly between groups (P>0.999; Table SI). These findings suggest that BM development in LUAD may be associated with enhanced tumor invasiveness; however, the lack of a significant difference in overall EGFR mutation distribution indicates that distinct mutation subtypes may exert differential effects on BM risk, warranting further subtype-level analysis. Serum concentrations of NSE, CEA, CA125, CA153 and CYFRA21-1 were subsequently measured in all patients. As shown in Fig. S2, serum levels of NSE, CEA, CA125, CA153 and CYFRA21-1 were significantly elevated in patients with BM compared with those in the non-BM group, by 1.44-, 4.93-, 3.33-, 2.13- and 2.30-fold, respectively (all P<0.0001), suggesting potential predictive value for BM in clinical settings. To further characterize these relationships, patients were stratified into four subgroups according to BM and EGFR mutation status. As shown in Fig. 1A-E, serum levels of all five biomarkers were significantly elevated in both BM subgroups compared with both non-BM subgroups (all P<0.0001), while EGFR mutation status exerted no significant effect within either the non-BM or BM stratum (all P>0.9999). The sole exception was CA153 (Fig. 1D), for which a modest but statistically significant difference was observed between EGFR-mutated and non-mutated patients within the non-BM group (P=0.038), but not within the BM group. Collectively, these results indicate that elevation of all five serum biomarkers was primarily driven by BM status rather than EGFR mutation status. Nevertheless, the complex interplay among BM, EGFR mutations and serum biomarkers suggests that subtype-level analysis of EGFR mutations may help elucidate underlying mechanisms.

All 615 patients with LUAD were classified into three subgroups according to EGFR genotype: 19del, 21L858R and other genotypes. Baseline clinicopathological characteristics, including pathologic stage, primary tumor status, lymph node metastasis and other tissue metastases, were comparable among the three genotype groups (Table SII). Compared with other genotypes, the 19del mutation was associated with a significantly higher BM rate (105/142, 73.9% vs. 296/473, 62.6%, P=0.013), while the 21L858R mutation showed a trend toward a lower BM rate that did not reach statistical significance (P=0.083; Table SII).

Patients were further stratified into four subgroups according to BM and 19del mutation status. As shown in Fig. 2A-E, serum levels of all five biomarkers were significantly elevated in both BM subgroups compared with both non-BM subgroups (all P<0.0001), whereas no significant difference was observed between 19del and non-19del subjects within either the non-BM or BM stratum (all P>0.999).

Univariate logistic regression analysis was performed with BM status as the dependent variable and nine candidate variables as independent variables, including primary tumor stage, lymph node metastasis, other tissue metastasis, EGFR 19del status and five serum biomarkers (NSE, CEA, CA125, CA153 and CYFRA21-1). All nine variables were significantly associated with BM in the univariate analysis (Table SIII). Among these, lymph node metastasis showed a stepwise increase in BM risk with advancing N stage (vs. N0: N1, OR=4.24; N2, OR=10.01; N3, OR=12.31; all P<0.001), and EGFR 19del was independently associated with elevated BM risk (OR=2.10, 95% CI: 1.25–3.51, P=0.005). All five serum biomarkers were also significantly associated with BM (all P<0.001; Table SIII). In the multivariate analysis, five variables were independently associated with BM: N stage (N1, N2 and N3 vs. N0), EGFR 19del, NSE, CEA and CYFRA21-1 (Table SIII). These results suggest that the EGFR 19del mutation is independently associated with an elevated risk of BM in LUAD, and that the concurrent elevation of serum NSE, CEA, CA125, CA153 and CYFRA21-1 reflects increased tumor burden at the time of BM development.

Building on the above findings, the predictive performance of serum biomarkers for BM was further evaluated within EGFR mutation-defined subgroups. In the overall cohort, ROC curve analysis yielded AUC values of 0.709, 0.756, 0.739, 0.742 and 0.751 for NSE, CEA, CA125, CA153 and CYFRA21-1, respectively, as individual biomarkers (Fig. S3). Following stratification by EGFR mutation status, AUC values in the EGFR-mutated subgroup increased substantially for NSE, CEA and CA153, reaching 0.746, 0.781 and 0.800, respectively, while those for CA125 and CYFRA21-1 showed modest changes, reaching 0.753 and 0.763. Further stratification by the 19del subtype yielded additional AUC gains for CA125 and CA153 in the 19del subgroup, reaching 0.803 and 0.815, respectively, while AUC values for NSE, CEA and CYFRA21-1 showed modest changes, reaching 0.775, 0.750 and 0.772, respectively (Fig. 3). These results indicate that stratifying patients by EGFR mutation status improves the BM predictive performance of serum biomarkers compared with biomarkers alone, and that further stratification by the 19del subtype yields additional gains for selected biomarkers. These findings suggest that distinct EGFR mutation subtypes exert differential effects on biomarker-based BM prediction in LUAD.

A combined approach incorporating EGFR mutation status and serum concentrations of NSE, CEA, CA125, CA153 and CYFRA21-1 was evaluated for BM prediction. ROC curve analysis of the combined approach in the overall cohort yielded an AUC of 0.776. Following stratification by EGFR mutation status, the AUC increased to 0.820 in the EGFR-mutated subgroup and further to 0.871 in the 19del subgroup (Fig. 4). These findings demonstrate that the combined approach outperforms individual biomarkers for BM prediction in LUAD, with the performance further enhanced in EGFR mutation-enriched subgroups.

Multivariate logistic regression analysis identified eight variables independently associated with BM in patients with LUAD, including N stage, other tissue metastasis, EGFR 19del, NSE, CEA, CA125, CA153 and CYFRA21-1. The resulting risk prediction formula was as follows: Logit (P)=−2.4028 + 1.0930(N1) + 1.7697(N2) + 1.5743(N3) + 0.3853 (other tissue metastasis) + 0.8107(EGFR 19del) + 0.7753 (NSE >18.18 ng/ml) + 0.7440 (CEA > 6.01 ng/ml) + 0.4539 (CA125 >19.59 U/ml) + 0.3879 (CA153 >14.82 U/ml) + 0.7526 (CYFRA21-1 >4.17 ng/ml), with an optimal probability cutoff of 0.591. A nomogram incorporating these eight variables was constructed for individualized BM risk prediction (Fig. 5A). In the 10-fold cross-validation, the nomogram achieved a mean AUC of 0.814, sensitivity of 0.564 and specificity of 0.906 across the 10 iterations (Table SIV). The nomogram achieved an AUC of 0.835 on the training set ROC curve (Fig. 5B). Calibration curves demonstrated close agreement between predicted and observed BM frequencies across the cohort, confirming good model calibration (Fig. 5C). DCA demonstrated that the nomogram provided greater net clinical benefit than the TNM staging model across a wide range of threshold probabilities (Fig. 5D). Furthermore, compared with the TNM-only model, the nomogram yielded an NRI of 0.210 (P<0.001) and an IDI of 0.147 (P<0.001), indicating that incorporation of EGFR 19del mutation status and the five serum biomarkers significantly improved the predictive performance over TNM staging alone.