Stage, EGFR Status and CEA Predict Brain Metastasis Risk in Lung Cancer

Patients with lung adenocarcinoma whose tumors have already reached advanced stages, spread to lymph nodes, or carry an EGFR mutation face a substantially higher hazard of developing brain metastases months or years later, according to a retrospective cohort study published in Frontiers in Genetics. A research team at The Second People’s Hospital of Liaocheng in Shandong Province, China, followed 66 patients who initially showed no sign of brain metastasis on a baseline contrast-enhanced brain MRI and found that a simple combination of four routinely measured factors—T classification, N classification, EGFR mutation status, and serum carcinoembryonic antigen (CEA)—could estimate an individual’s probability of remaining free of brain metastasis at 12 and 24 months with moderate accuracy.

The study screened 94 consecutive patients with pathologically confirmed lung adenocarcinoma diagnosed between January 2018 and December 2024. Twenty-eight were excluded, most commonly because brain metastasis was already visible on the baseline MRI (17 patients), because baseline clinicopathological or molecular data were incomplete (6), because of a concurrent or recent active malignancy (3), or because no post-baseline follow-up information existed (2). The remaining 66 patients entered the analysis at the date of a negative baseline brain MRI, which the researchers defined as the index date. Any metastasis detected after that point, even within the first three months, counted as an incident event rather than being excluded as synchronous disease.

During a median follow-up of 27.8 months—ranging from 1.8 to 72.0 months—26 of the 66 patients developed incident brain metastasis. The median time from the index date to brain metastasis was 15.7 months. Nine patients died before any documented brain metastasis, and 31 were alive without documented brain metastasis at their last follow-up, meaning 40 patients were censored without an event in the primary analysis. The estimated brain-metastasis-free survival for the overall cohort was 82.1 percent at 12 months and 66.4 percent at 24 months.

The baseline characteristics of patients who later developed brain metastasis differed noticeably from those who did not. Among patients with incident brain metastasis, 61.5 percent had T3–4 disease compared with 30.0 percent of those who remained free of brain metastasis; 80.8 percent had N1–3 nodal involvement compared with 52.5 percent; 73.1 percent carried an EGFR mutation compared with 42.5 percent; and 65.4 percent had serum CEA above the institutional upper reference limit of 5 ng/mL compared with 37.5 percent. The mean age of the cohort was 58.7 years, and 56.1 percent were male.

Because follow-up duration varied widely and some patients were censored, the researchers analyzed brain metastasis as a time-to-event outcome using Cox proportional hazards regression rather than treating it as a simple binary endpoint. In univariable analyses, T3–4 disease was associated with a hazard ratio of 2.63 (95 percent CI, 1.20–5.76), N1–3 involvement with 2.71 (95 percent CI, 1.12–6.54), EGFR mutation with 2.58 (95 percent CI, 1.12–5.94), and CEA above 5 ng/mL with a less precise 2.12 (95 percent CI, 0.98–4.60). Kaplan-Meier analysis showed shorter brain-metastasis-free survival among patients with EGFR-mutant tumors, with a median of 22.4 months versus not reached in the wild-type group (log-rank P = 0.009).

To build the prediction model without overfitting a small dataset, the team prespecified candidate predictors on clinical grounds, checked for multicollinearity, and used LASSO-penalized Cox regression for variable reduction rather than stepwise screening. The penalized procedure retained T classification, N classification, EGFR mutation status, and serum CEA. In the resulting parsimonious Cox model, T3–4 disease carried a hazard ratio of 2.31 (95 percent CI, 1.02–5.22), N1–3 involvement 2.42 (95 percent CI, 1.00–5.87), and EGFR mutation 2.49 (95 percent CI, 1.06–5.85), while the CEA estimate remained less precise at 2.04 (95 percent CI, 0.93–4.47). Overall clinical stage was deliberately excluded from the multivariable model because it is constructed from T and N categories and would introduce structural redundancy.

Model performance was assessed with internal validation only. The apparent Harrell C-index was 0.781, and bootstrap internal validation using 1,000 resamples—repeating the entire model-development process within each resample—yielded an optimism-corrected C-index of 0.742. Time-dependent ROC analysis showed an AUC of 0.773 (95 percent CI, 0.665–0.866) at 12 months and 0.758 (95 percent CI, 0.650–0.850) at 24 months. Bootstrap-corrected calibration curves showed reasonable agreement between predicted and observed brain-metastasis-free probabilities at both horizons, with an optimism-corrected calibration slope of 0.87 and a 24-month Brier score of 0.164. The resulting nomogram estimates individualized 12- and 24-month brain-metastasis-free probabilities.

Because death before brain metastasis prevents the event from ever being observed, the researchers also performed a competing-risk sensitivity analysis using the Fine Gray framework, treating the nine deaths before brain metastasis as competing events. The cumulative incidence of brain metastasis was 16.7 percent at 12 months and 30.8 percent at 24 months. The competing-risk analysis produced broadly consistent results: T3–4 disease carried a subdistribution hazard ratio of 2.18 (95 percent CI, 1.01–4.73), N1–3 involvement 2.27 (95 percent CI, 1.01–5.11), and EGFR mutation 2.36 (95 percent CI, 1.06–5.27), with CEA directionally similar but less precise at 1.88 (95 percent CI, 0.88–4.03). An additional exploratory analysis modeled EGFR-TKI exposure as a time-dependent covariate to reduce immortal-time bias, since treatment frequently began after cohort entry. In that analysis the associations remained directionally consistent but attenuated, and time-dependent EGFR-TKI exposure itself was associated with a lower hazard of brain metastasis (HR, 0.76; 95 percent CI, 0.33–1.75), though the confidence interval was wide and included the null.

The EGFR finding deserves particular context. Among the 36 patients with EGFR-mutant tumors, 19 developed brain metastasis, compared with 7 of 30 patients with EGFR-wild-type tumors. Within the mutant group, 17 had exon 19 deletions and 16 had L858R mutations, with brain metastasis occurring in 9 of each subgroup—too few events for formal comparison between subtypes. The authors note that the relationship between EGFR mutation and brain metastasis should not be read as a purely intrinsic biological effect, because EGFR-TKI therapy alters both survival duration and the pattern of central nervous system progression, and treatment allocation in this cohort was not randomized. During follow-up, 51.5 percent of patients received an EGFR-TKI, including 21.2 percent treated with osimertinib, alongside chemotherapy in 65.2 percent, immune checkpoint inhibitor therapy in 25.8 percent, and extracranial radiotherapy in 31.8 percent.

The study’s limitations are considerable and the authors emphasize them repeatedly. It was a single-center retrospective study with only 26 brain metastasis events, so coefficient estimates may be unstable and residual overfitting remains possible despite penalization and bootstrap correction. Surveillance schedules were not uniform, which could have introduced detection bias, although the median surveillance interval was similar between groups (4.8 versus 5.3 months, P = 0.41). Treatment regimens evolved substantially between 2018 and 2024, and molecular testing platforms changed during the study period, so alterations beyond EGFR were not systematically incorporated. No external validation cohort was available, and the study did not evaluate whether risk-adapted surveillance based on the model improves survival, neurological outcomes, or quality of life. The authors state that the model should be regarded as exploratory, should not be used to determine treatment selection or prophylactic cranial irradiation, and requires external validation in larger multicenter populations before any clinical implementation. Its principal value, they conclude, lies in hypothesis generation and preliminary risk stratification using variables that are already routinely available in clinical practice.

Subject of Research: Time-dependent risk factors and prediction modeling for brain metastasis in lung adenocarcinoma

Article Title: Risk factors and time-to-event prediction of brain metastasis in patients with lung adenocarcinoma: a retrospective cohort study

Article References: Wang, J., Shen, Y., Xu, Y., Liu, J., & Guo, Y. (2026). Risk factors and time-to-event prediction of brain metastasis in patients with lung adenocarcinoma: a retrospective cohort study. Frontiers in Genetics, 17, Article 1931724. https://doi.org/10.3389/fgene.2026.1931724

Image Credits: AI Generated

DOI: 10.3389/fgene.2026.1931724

Keywords: brain metastasis, lung adenocarcinoma, EGFR mutation, CEA, Cox regression, nomogram, competing risk, retrospective cohort, LASSO, prediction model, TNM staging, internal validation