This retrospective cohort study included patients diagnosed with NSCLC and BM at Yunnan Cancer Hospital (Yunnan, China) between January 2017 and December 2019. Eligible patients met the following inclusion criteria: i) Pathologically confirmed NSCLC; ii) radiologically confirmed BM based on bone scintigraphy, computed tomography, or magnetic resonance imaging; iii) no history of other primary malignancies; and iv) availability of complete clinical and laboratory data at the time of BM diagnosis. Patients with small cell lung cancer, unclear primary tumor origin, severe chronic inflammatory or autoimmune diseases, active infections, liver cirrhosis, or incomplete data were excluded (Fig. 1).

According to the predefined inclusion criteria, a total of 233 patients who were newly diagnosed with NSCLC with BM at our institution were retrospectively enrolled. All cases were pathologically confirmed primary NSCLC. The cohort consisted of 141 males (60.5%) and 92 females (39.5%), with a median age of 59 years (range: 50–66 years). The median OS was 16 months (range: 6.0–50.0 months). A total of 160 patients (68.6%) were younger than 65 years, whereas 73 patients (31.3%) were aged 65 years or older.

The present study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of Yunnan Cancer Hospital (approval no. KYLX2023-038). All patients signed a written informed consent form, which included consent to participate in the present study, use of their medical data for research purposes and the publication of anonymized findings.

Medical records were accessed between January 2025 and March 2025 for the purpose of data collection and analysis. Clinical data were obtained from the hospital information system and included demographic characteristics, smoking history, Karnofsky Performance Status (KPS), pathological type, TNM stage according to the eighth edition of the American Joint Committee on Cancer (AJCC) staging system (20) treatment history before and after BM diagnosis and occurrence of skeletal-related events. Due to the retrospective nature of the study and changes in treatment strategies during the study period, detailed information regarding post-BM systemic therapies, including immunotherapy and targeted therapy regimens, was not consistently available for all patients.

Laboratory parameters collected at the time of BM diagnosis included complete blood counts, serum biochemistry, inflammatory markers, lipid profiles and tumor markers. Based on these data, 10 nutritional-inflammation indicators were calculated using established formulas: PLR, NLR, SII, PNI, CAR, C-reactive protein-to-lymphocyte ratio (CLR), ALI, body mass index (BMI), NLDA and HALP.

OS was defined as the interval from the diagnosis of BM to mortality from any cause or last follow-up. Survival status was obtained through medical records and telephone follow-up. Patients alive at the last follow-up were censored.

Continuous variables were assessed for normality using the Shapiro-Wilk test and expressed as mean ± standard deviation or median with interquartile range, as appropriate. Categorical variables were summarized as frequencies and percentages. Group comparisons were performed using the Student's t-test, Mann-Whitney U test, χ² test, or Fisher's exact test.

Univariate Cox proportional hazards regression was used to identify potential prognostic factors for OS. Variables with P<0.05 in univariate analysis were further evaluated using multivariate Cox regression with stepwise selection based on the Akaike information criterion. Optimal cut-off values for continuous variables were determined using maximally selected rank statistics, and Kaplan-Meier survival curves were generated with log-rank tests.

Lasso regression, Random Forest survival models, and extreme gradient boosting (XGBoost) were applied for variable selection. Based on selected variables, Cox regression models were constructed using bootstrap resampling to enhance robustness. Model performance was evaluated using the concordance index (C-index), time-dependent area under the curve (AUC), Brier score, calibration curves, and AIC. Statistical analyses were performed using R software (version 4.3.1) and a two-sided P<0.05 was considered to indicate a statistically significant difference.

Comparative analysis of baseline characteristics between survival groups demonstrated significant differences in age, presence of intrapulmonary metastasis, bisphosphonate treatment after BM, prior surgical history, ALP, LDH, CLR and several tumor markers (all P<0.05). The proportion of elderly patients, males, squamous cell carcinoma, and both intrathoracic and extrathoracic metastases was higher in the deceased group compared with the surviving group. Detailed baseline characteristics are summarized in Tables I and SI.

Univariate Cox proportional hazards analysis identified multiple variables markedly associated with OS, including age, histological subtype, bisphosphonate treatment, history of primary tumor surgery, erythrocyte sedimentation rate, white blood cell count, hemoglobin (Hb), CA199, CYFRA21-1, NSE, D-dimer, fibrinogen, LDH, ALP, total cholesterol (TC), NLR, SII, PNI, CAR, CLR and CRP (all P<0.05).

After adjustment for potential confounders, multivariate Cox regression analysis demonstrated that history of primary tumor surgery was independently associated with improved OS (HR=0.35; 95% CI: 0.23–0.55; P<0.001). Higher Hb levels were also protective (HR=0.99; 95% CI: 0.986–0.999; P=0.02), as was higher PNI (HR=0.98; 95% CI: 0.97–0.99; P=0.002). By contrast, elevated CYFRA21-1 (HR=1.01; 95% CI: 1.00–1.01; P<0.001), NSE (HR=1.03; 95% CI: 1.02–1.05; P<0.001) and TC (HR=1.05; 95% CI: 1.01–1.08; P=0.006) were independently associated with increased mortality risk. Although several inflammatory markers, including NLR, SII, ALI, CAR, CLR, HALP, NLDA, PLR and BMI, were significant in univariate analysis, they did not retain statistical significance in the multivariate model. Detailed results are shown in Table SII and Fig. 2.

To further evaluate the prognostic relevance of the identified factors, optimal cutoff values for continuous variables were determined using the surv_cutpoint function (Fig. S1). Kaplan-Meier survival analyses demonstrated that patients with a history of primary tumor surgery had markedly prolonged OS compared with those without surgery (P<0.001). Higher Hb and PNI levels were associated with improved OS (P=0.004 and P=0.018, respectively), whereas elevated CYFRA21-1 and NSE levels were markedly associated with poorer survival (both P<0.001). No significant survival difference was observed between high and low TC groups (P=0.15; Fig. 3).

The discriminative ability of individual nutritional-inflammatory biomarkers was subsequently assessed using Harrell's concordance index (C-index). All single-marker models yielded C-index values below 0.60, ranging from 0.48–0.58 (mean: 0.54), indicating limited prognostic discrimination and performance only slightly above random prediction (Fig. S2A). Subgroup analyses revealed heterogeneity in predictive performance across clinical strata. Relatively higher C-index values were observed in patients with SREs (C=0.64), in those without surgery (C=0.59), and in male patients (C=0.60); however, these values remained below the threshold generally considered indicative of moderate predictive accuracy (C>0.70; Fig. S2B).

Time-dependent receiver operating characteristic (ROC) analysis was further performed to evaluate dynamic predictive performance. The AUC values of individual inflammatory indicator models demonstrated temporal variability (Fig. S2C). The predictive performance of PLR, NLR and CLR gradually improved over time. The SII model maintained relatively stable and comparatively higher AUC values at 12 and 24 months. By contrast, models such as ALI and BMI showed fluctuating performance across different time points. The AUC of PNI declined at 12 months and subsequently recovered at 24 months, indicating temporal variation in its prognostic sensitivity. Despite these dynamic changes, most individual biomarkers consistently exhibited relatively low AUC values across time points, suggesting limited standalone prognostic utility. An integrated prognostic model incorporating clinical and inflammatory variables was constructed.

The present study systematically evaluated the prognostic value of nutritional-inflammatory biomarkers in patients with NSCLC with BM and developed integrated prognostic models incorporating clinical and laboratory variables. The findings can be summarized as follows: i) Several nutritional-inflammatory markers were markedly associated with OS, although most lacked independent predictive value when analyzed individually; ii) the prognostic performance of single inflammatory markers was limited and varied across time points and clinical subgroups; iii) integrated models incorporating nutritional-inflammatory indicators markedly improved predictive performance compared with models excluding these markers; and iv) a LASSO-based Cox model demonstrated the best overall discrimination and calibration and was used to construct a prognostic nomogram.

Traditional laboratory parameters remain widely accessible and cost-effective tools in oncology practice. In contrast to molecular profiling, these biomarkers reflect systemic host responses, including inflammation, nutritional status, and immune function. However, data regarding their prognostic relevance in patients with NSCLC with BM remain limited.

PNI, which integrates serum albumin and lymphocyte count, reflects the interaction between nutritional status and immune competence. In the present study, higher PNI was independently associated with improved survival, consistent with previous reports in advanced NSCLC (21,22). Hypoalbuminemia may indicate protein-energy malnutrition and systemic inflammation (14), whereas lymphopenia reflects impaired antitumor immune surveillance (23,24). These factors may collectively contribute to disease progression in metastatic settings. Hb, although not a classical inflammatory marker, also emerged as an independent prognostic factor (25). Cancer-related anemia is common in advanced lung cancer and may promote tumor progression through hypoxia-related pathways (26,27). The findings of the present study extend previous evidence by demonstrating its prognostic significance specifically in patients with BM.

Several inflammatory indices, including SII and NLR, were markedly associated with OS in univariate analysis but did not retain independent significance in multivariable models. This suggested that their prognostic contribution may be partially mediated through other clinical or laboratory factors. Importantly, markers such as CRP, CAR and CLR were also associated with survival, highlighting the relevance of systemic inflammation in this patient population.

Traditional tumor biomarkers, including CYFRA21-1 and NSE, were independently associated with survival, emphasizing the importance of tumor burden-related biological activity. Additionally, metabolic indicators such as total cholesterol demonstrated independent prognostic relevance, suggesting that tumor biology and host metabolic status jointly influence survival outcomes in patients with NSCLC with BM. These findings support the concept that prognosis in metastatic lung cancer is determined not only by tumor characteristics but also by systemic host responses.

Time-dependent ROC analysis demonstrated dynamic changes in predictive performance over follow-up. Certain markers, such as PLR and CLR, showed gradual improvement in discrimination at later time points, whereas others exhibited fluctuating performance. Subgroup analyses revealed heterogeneity in predictive ability across clinical strata, including patients with SREs and those without prior surgery.

Although the absolute discrimination of individual markers remained modest, these findings suggest that the prognostic impact of inflammatory and nutritional parameters may vary according to disease stage, treatment history and host condition. However, the relatively low C-index values indicate that single biomarkers alone are insufficient for accurate risk stratification.

To address the limited performance of individual markers, the present study constructed integrated prognostic models using LASSO, random forest, and XGBoost for feature selection. Models incorporating nutritional-inflammatory indicators consistently demonstrated superior discrimination, calibration, and overall performance compared with models excluding these variables.

Among all models, the LASSO-based Cox model achieved the highest C-index and lowest AIC, indicating optimal balance between predictive accuracy and model complexity. Based on this model, a prognostic nomogram was developed to facilitate individualized survival estimation in clinical practice.

These findings suggest that nutritional-inflammatory markers provide incremental prognostic information when integrated with tumor-related and clinical variables, rather than functioning as standalone predictors.

The present study had several strengths, including comprehensive evaluation of multiple inflammatory markers, application of machine learning-based feature selection methods and systematic comparison of models with and without nutritional-inflammatory indicators.

However, several limitations should be acknowledged. First, this was a single-center retrospective study with a relatively limited sample size, which may restrict the generalizability of the findings and introduce potential selection bias. In addition, the proposed nomogram has not yet been externally validated in independent cohorts. Therefore, the current model should be considered a preliminary prognostic tool and further multicenter prospective validation is required before broader clinical application. Second, progression-free survival was not included as a formal endpoint because retrospective progression assessment was considered insufficiently reliable owing to irregular imaging follow-up and incomplete documentation. Therefore, only OS was analyzed in the present study. Third, treatment heterogeneity may have influenced survival outcomes. During the study period (2017–2019), the clinical use of targeted therapy and immunotherapy evolved substantially, but detailed treatment records after BM diagnosis were incomplete and therefore could not be fully adjusted for in the current analysis. Finally, certain bone metabolism-related biochemical markers were not included in the dataset.

In addition, although multiple machine learning-based feature selection methods and bootstrap resampling were applied to reduce overfitting risk, the relatively limited sample size compared with the number of candidate variables may still introduce model instability. Therefore, the predictive performance of the current models should be interpreted cautiously until validated in independent external cohorts.

Future studies integrating multicenter data, longitudinal biomarker monitoring and molecular profiling may further refine prognostic stratification and improve clinical applicability.

In conclusion, although individual nutritional-inflammatory markers demonstrate limited standalone predictive performance, they provide meaningful incremental value when incorporated into integrated prognostic models. A LASSO-based nomogram combining clinical, tumor-related, and host inflammatory parameters offers a practical tool for individualized survival prediction in patients with NSCLC with BM. Further external validation is warranted to confirm its clinical utility.