Data were retrospectively collected from the records of consecutive patients who received a diagnosis of advanced NSCLC in two hospitals from 2013 to 2019. Clinical staging of the disease was conducted renewedly with reference to the eighth edition for tumor-node-metastasis (TNM) classification (19), at the time of data collection. The inclusion criteria in this study were: (1) a diagnosis of LUAD confirmed from pathological or cytological specimens, or both; (2) evidences of bone metastasis confirmed by imaging examinations, such as plain radiograph, CT, PET-CT, MRI, and bone scan, or a bone biopsy performed during surgery; (3) the number of bone metastases was ≤ 5; (4) a data of gene mutation status identified via next-generation sequencing; (5) did not receive immunotherapy in the first-line. Patients were excluded if they had second primary tumor; a site of metastasis other than bone; without gene sequence result; or incomplete medical records.

In this study, positively sensitive mutations (SM⁺) included: epidermal growth factor receptor (EGFR) exon 19 deletion, EGFR exon 21 Leu858Arg mutation, and anaplastic lymphoma kinase (ALK) mutation. EGFR uncommon mutations, such as exon 18 mutations, exon 20 insertion mutations and so on, no-targeted therapy mutations or without any mutation, were defined as sensitive mutations negative (SM^–).

Laboratory examinations including routine blood test data, hepatic, and renal function test data of patients were collected before initial treatment. The calculation formulas of neutrophils to lymphocyte ratio (NLR), platelet to lymphocyte ratio (PLR), and systemic inflammatory index (SII) were as follows: NLR = neutrophil number (10⁹/L)/lymphocyte count (10⁹/L); PLR = number of platelets (10⁹/L)/number of lymphocytes (10⁹/L); SII = number of platelets(10⁹/L) × number of neutrophils (10⁹/L)/number of lymphocytes (10⁹/L). Corrected calcemia was computed according to the formula: c-Ca = measured Ca + (40-albumin)/40. The best cutoff values for albumin, alkaline phosphatase (ALP), leukocyte, PLR, NLR, SII, and c-Ca were obtained according to overall survival (OS).

All patients with EGFR non-sensitive mutations, no-targeted therapy mutations, or without mutation underwent first-line chemotherapy after confirmation of the initial LUAD diagnosis. The treatment included platinum-based doublet chemotherapy such as pemetrexed or paclitaxel combined with cisplatin, carboplatin, or nedaplatin. Each chemotherapy session was separated by an interval of 3 to 4 weeks.

Patients with EGFR-sensitive mutations (exon 19 deletion, exon 21 Leu858Arg mutations) were administered first-line treatment with EGFR tyrosine kinase inhibitors (TKIs), such as gefitinib, erlotinib, and icotinib; or with chemotherapy mentioned above and then TKIs after disease progression. All patients with ALK mutation were administered first-line treatment with crizotinib, or with chemotherapy as aforesaid and then TKI after disease progression.

OS was the primary endpoint, defined as the time from the date of diagnosis until death or the last follow-up. The follow-up schedule began from the time of treatment to the final follow-up on November 22, 2021. The data on the date of death or at the final follow-up visit were acquired from hospital records or through direct correspondence with the family of patients. R 4.1.1 software and SPSS 24.0 software were used to perform the statistical analyses. The chi-squared test (or Fisher’s exact test as applicable) and independent-samples T test were used to compare the clinical characteristics between the internal and external groups. OS was estimated using Kaplan-Meier method and between-group difference in OS was assessed using log-rank test. The optimal cutoff values of hematology markers were determined using the package “survminer” based on OS. LASSO-Cox regression analysis was performed to select the optimal prognostic factors using packages “glmnet,” “survival,” and “MASS” and the backward-forward stepwise method. The “predict” function of package “survival” was used to calculate the risk-score of each patient.

Nomograms, including clinical variables alone or clinical variables plus Risk-H, were constructed by using the package “regplot”. The concordance index (C-index) and calibration curves were evaluated to assess the consistency between the predicted and observed probabilities using package “pec”. The time-dependent receiver operating characteristic (ROC) analysis and area under the curves (AUC) were evaluated to assess the discrimination using packages “survivalROC” and “riskRegression”. The decision curve analysis (DCA) was formulated to evaluate the clinical practicality of constructed models using package “ggDCA”. All P-values were two-sided, with P < 0.05 considered statistically significant.

Data were collected for 983 patients and 654 patients with advanced NSCLC, who had been treated in two hospitals from January 2013 to December 2019. The detailed patients selecting process is shown in Figure 1 . Eventually, 125 and 69 patients, respectively, were enrolled in the internal set and the external set. The demographic and clinic-pathological features of patients are displayed in Table 1 . The median follow-up time, median OS, and 1-, 3-, 5-year survival rates were 35.5 vs. 43.0 months, 28.3 vs. 32.7 months, 87% vs. 88.4%, 43.4% vs. 49.0%, and 17.4% vs. 13.0%, respectively, in the internal set and external set.

In the internal set, the optimal cutoff values of hematology markers were determined based on OS and the result showed that high albumin (p=0.020) was a favorable indicator of OS, whereas high ALP (p=0.005), high c-Ca (p=0.015), and high leukocyte (p=0.039) were accompanied by inferior OS. But the NLR (p=0.113), PLR (p=0.219), and SII (p=0.113) level had no significant statistical effects for survival ( Table 2 ).

NLR, PLR, and SII had been reported as important prognostic factors for OS, so, these three variables and four other variables with p value < 0.05 were all included in the LASSO-Cox regression model to select the optimal prognostic variables in the internal set. Finally, ALP, albumin, and leukocyte were significantly independent prognostic factors, and were included to formulate the risk scoring system ( Figures 2A, B , AIC value=666.31, p=8.176×e-06). A formula named Risk-H was constructed as follows: Risk-H=1 * HR-value (ALP) * HR-value (Albumin) * HR-value (Leukocyte) ( Table 3 ). According to the median value of risk score (2.783129), patients in the internal set were divided into low-risk and high-risk groups and the median OS were 46.0 and 28.3 months, respectively (p<0.001, Figure 3A ). Patients in the external set were calculated risk score according to the Risk-H formula and Table 3 and were divided into low- and high-risk groups based on the median score mentioned above. This significant prognostic difference was also observed (p=0.011, Figure 3C ). The prognostic accuracy of Risk-H was evaluated using time-dependent ROC analysis, yielding comparable AUC values between the internal and external sets with 2-, 3-, and 4-year AUC values of 0.698 vs. 0.672, 0.747 vs. 0.640, 0.729 vs. 0.690, respectively ( Figures 3B, D ), which confirmed the excellent prognostic power of Risk-H in another heterogeneous population.

In the internal set, Risk-H, and clinical variables, such as gender, age, smoking, karnofsky performance status scores, loss of weight, gene mutation status, T stage, N stage, number of bone metastases, antiresorptive drugs treatment, and weight bearing bone metastasis were included in the LASSO-Cox regression model to select optimal prognostic factors associated with OS and PFS using the backward-forward stepwise method ( Figures 4A, B ). After adjusting clinical characteristics, Risk-H (p < 0.001) remained as an independent negative prognostic indicator for OS. Beside Risk-H, another four clinical variables, including loss of weight, SM status, T stage, and N stage were added to establish the optimum model with the smallest AIC value (660.98, p=1.117×e-06) ( Table 4 ). However, the significant effect of Risk-H on progression free survival (PFS) was not observed in the univariate analysis in the internal set (p=0.051).

In the internal set, nomogram models were established to predict the survival probability of LUAD with ≤ 5 BOM using four clinical variables mentioned above with or without Risk-H (Model 1 vs. Model 2). Figures 5A, B demonstrated an example of using the two nomogram models to predict the survival probability of a given patient which revealed that the addition of Risk-H to nomogram harbored improved predictive accuracy for survival when compared with that of clinical factors alone. The consistency between the predicted and observed probabilities were also improved as demonstrated by time-dependent C-index ( Figure 6A ) and 3- and 4-year calibration curves ( Figures 6B, C ).

We generated ROC analysis to assess the discrimination abilities of two nomograms and the result displayed nomogram constructed by clinical variables plus Risk-H had higher AUC value in predicting 3- and 4-year survival probability in the internal set ( Figures 6D, E ). Furthermore, time-dependent AUC curve also proved the superiority of combined prediction model again ( Figure 6F ).

Because the calibration curves and ROC analysis are based on the sensitivity and specificity of the predictive model, they cannot recognize false positive and false negative cases. While DCA was widely adopted to assess the clinical utility and net clinical benefits when the predictive model guides clinical practice. So, we performed 3- and 4-year DCA to evaluate the clinical utility and net clinical benefits that two nomograms would bring to patients and the results revealed that the integrated nomogram was significantly superior to the clinical nomogram in the internal set ( Figures 6G, H ). To sum up, the above results elucidated that the integrated nomogram had better predictive ability for the survival probability of LUAD patients with 1-5 BOM.