Establishment of a prognostic risk prediction modelfor non-small cell lung cancer patients with brainmetastases: a retrospective study

Study population and follow-up

This retrospective study included 300 patients with NSCLC who were diagnosed with BM between January 2006 and May 2020 at Yunnan Cancer Hospital, Third Affiliated Hospital of Kunming Medical University. The inclusion criteria for this study were as follows: (1) NSCLC confirmed via pathological examination; (2) magnetic resonance imaging-confirmed BM; (3) availability of patient demographic characteristics, clinicopathological features, serological indicators, and treatment information; (4) absence of other concurrent cancers. The survival duration of the patients was determined by reviewing the medical records and conducting telephonic inquiries. The OS was defined as the interval from the initial diagnosis until death from any cause or the date of the last follow-up visit (Pilz, Manegold & Schmid-Bindert, 2012).

Data collection

Baseline clinical data were obtained from medical records at the time of initial diagnosis of BM. The collected data encompassed various aspects, including general conditions (age, sex, body mass index (BMI), smoking history, and KPS), tumour markers (carcinoembryonic antigen (CEA), neuron-specific enolase (NSE), cytokeratins (cytoplasmic protein fragment of cytokeratin 19 (CYFRA21)), and squamous cell carcinoma antigen (SCCA)), serological indicators (albumin (ALB), lactate dehydrogenase, and alkaline phosphatase (ALP)), serum inflammatory indicators (neutrophils, platelets, lymphocytes, monocytes, platelet/lymphocyte ratio (PLR), neutrophil/lymphocyte ratio (NLR), systemic immune-inflammation index (SII) = platelet × neutrophil/lymphocyte, advance lung cancer inflammation index (ALI) = BMI × ALB/NLR), prognostic nutritional index (PNI) = ALB + 5 × lymphocyte), advance distant metastases (number of BM, lung metastasis, intrathoracic metastasis (malignant pleural effusion, pericardial effusion, or pleural metastasis), liver metastasis, bone metastasis, adrenal metastasis, and metastases to other sites), signs and symptoms of BM (intracranial hypertension, focal signs and symptoms, epilepsy, and decreased cognitive function), type of pathology, pathological stage (tumour stage, lymph node stage (N_stage), or metastasis stage (M_stage)/TNM stage), epidermal growth factor receptor (EGFR) gene mutation status, treatment status (surgery for primary lung cancer foci, radiotherapy of primary lung cancer, radiotherapy for BM lesions (whole-brain radiation therapy or stereotactic radiation therapy), surgical treatment of metastatic brain lesions, chemotherapy, or EGFR-tyrosine kinase inhibitor (TKI) treatment), and classification information of RPA, GPA, Lung-molGPA, and BSBM models. The above predictors were complete and comprised objective data. All predictors were assessed independently of each other, without any knowledge of the clinical outcome. All continuity predictors maintained their continuity and were not categorised. The categorised predictors were all predetermined before model construction. The continuous and categorical variables are presented in Table S1. The sample size of this study met the criterion of having events per variable (EPV) of >10 (Peduzzi et al., 1996; Austin, Allignol & Fine, 2017; Moons et al., 2019).

Model construction and evaluation

The final predictors were selected via a 10-fold cross-validation of LASSO-Cox regression, whereby the λ-value associated with the minimum standard error was chosen. Subsequently, the final predictors were incorporated into a multivariate Cox regression analysis, and the risk score for each patient was calculated using the “predict ()” function. Finally, a prognostic model was constructed. ${\displaystyle Riskscore=h_0\left(t\right)\times exp\left({\beta }_1{X}_1+{\beta }_2{X}_2+\cdots +{\beta }_{\mathrm{n}}{X}_{\mathrm{n}}\right)}$ The discriminatory ability of the model was assessed by evaluating the area under the receiver operating characteristic (ROC) curve (AUC). Furthermore, the calibration curve was plotted and the Brier score was calculated to measure the calibration of the model. Internal validation was conducted using the bootstrap method (resampling 1,000 times). The discriminative ability and clinical utility of the novel prognostic models were compared with RPA, GPA, Lung-molGPA, BSBM, and TNM staging using time-dependent C-index and decision curve analysis (DCA). A larger AUC value indicated better predictability of the model (Carrington et al., 2023). DCA demonstrated the relationship between benefits and risks across models by examining various cut points (thresholds) in different models (Van Calster et al., 2018). Furthermore, the integrated discrimination improvement (IDI) and net reclassification improvement (NRI) were employed to assess the reclassification performance and discrimination of our novel prediction models compared with RPA, GPA, Lung-molGPA, BSBM, and TNM staging. A nomogram was developed based on the selected predictors to facilitate individual survival prediction in NSCLC patients with BM. Subsequently, patients were classified into low-risk, intermediate-risk, and high-risk groups based on the new prediction model RiskScore. The differences in OS among these three subgroups were analysed using the Kaplan–Meier method. All statistical analyses were performed using R software (version 4.2.1; R Core Team, 2022), and statistical significance was set at P ≤ 0.05.

Patient characteristics

A total of 300 NSCLC patients with BM were included in this study, all of whom had complete baseline clinical and laboratory data. The clinicopathological characteristics and laboratory results of these patients are summarised in Table S1. The mean age of the patients was 55.4 years (range 31–83 years). Among the patients, 185 were males and 115 were females. The median follow-up duration of the patients was 13.9 months, with a minimum and maximum follow-up duration of 0.1 months and 173.83 months, respectively. The last follow-up was performed on 16 June 2021. The OS rates for these patients at 1, 3, and 5 years were 75%, 49%, and 40.3%, respectively.

Construction of the prognostic models

First, LASSO-Cox regression analysis was used to identify the optimal predictors and constructed the model. Cross-validation was used to select the λ-value associated with lambda.min (λ = 0.054), which yielded the highest model fit (Fig. 1). This λ-value corresponds to the most significant prognostic factor for OS. The final model comprised 15 predictors, namely age, KPS, NSE, PLR, lymphocyte count, ALP, smoking history, intrathoracic metastasis, metastases to other sites, N_stage, M_stage, surgery for primary lung cancer foci, chemotherapy EGFR mutation, and TKI treatment. The EPV for each variable was 12.4. The regression coefficients of these predictors were used to construct the prognostic model. The risk score of the prognostic model was calculated as follows: risk score = h0(t) exp [(age × 0.0105543) + (kps × −0.0082627) + (NSE × 0.0017274) + (PLR × 0.0007071) + (lymphocyte count ×−0.0174690)+ (ALP ×−0.0001577) + (smoke × 0.1433414) + (intrathoracic_metastasis ×0.4382260) + (metastases_to_other_sites × 0.0330467) + (N_stage × 0.1062274)+(M_stage × 0.0277756) + (surgery × −0.2624173) + (chemotherapy × −0.0027364) + (TKI_therapy × −0.3508606) + (EGFR mutation × −0.0952148)]. The continuous variables in the formula were based on the original numerical levels, while the categorical variables were represented by codes, as presented in Table S1.

We conducted Cox proportional risk regression analysis using the 15 predictors identified through LASSO regression (Fig. 2). The time-dependent ROC curves suggested AUC values of 0.746 (0.678−0.814), 0.819 (0.761−0.877), and 0.865 (0.774−0.957) for our prediction model at 1, 3, and 5 years, respectively. The Brier score was calculated to measure the overall performance of the model, yielding a value of 0.103 (0.072−0.134). The Brier score combines discrimination and calibration measures, with lower values indicating better model performance (0 for perfect overall performance and 0.25 for worthless model) (Schneider et al., 2022; Assel, Sjoberg & Vickers, 2017). Additionally, internal validation using the bootstrap method (resampling of 1,000) demonstrated an AUC of 0.811 (0.638−0.950) and a Brier score of 0.123 (0.066−0.188; based on 5 years), confirming the model’s good discriminative ability and calibration (Fig. 3).

Evaluation of the performance between our novel prognostic model, RPA, GPA, Lung-molGPA, BSBM, and TNM staging

The time-dependent C-index was used to evaluate the accuracy of our model’s predictions. Our model demonstrated superior discriminatory ability compared with RPA, GPA, Lung-molGPA, BSBM, and TNM staging, and these results were consistent with our bootstrap validation (Fig. 4).

Furthermore, DCA was employed to evaluate which focus the clinical applicability of our predictive model. Notably, the applicable threshold range varied among the six DCA curves. The widest threshold range was observed for our model, approximately 0.4−0.8, among the six curves. Moreover, within most of the threshold ranges, our model yielded the highest net benefits (Fig. 5). These findings indicate that our model is the optimal choice.

Lastly, the IDI and NRI metrics were incorporated to assess the performance of our model. The IDI analysis allowed us to measure the extent to which our new model improved its predictive power compared with the existing model, and the NRI analysis was used to determine the extent to which the new model improved the proportion of correct reclassifications compared with the existing model. The IDI analysis revealed that our model demonstrated positive improvements compared with GPA 0.152 (0.063−0.287), RPA 0.209 (0.113−0.347), Lung-molGPA 0.106 (0.030−0.240), BSBM 0.120 (0.044−0.247), and TNM staging 0.218 (0.122−0.354) with IDI >0. Meanwhile, the NRI analysis revealed that our model demonstrated positive improvements compared with GPA 0.537 (0.172−0.676), RPA 0.474 (0.270−0.722), Lung-molGPA 0.525 (0.124−0.641), BSBM 0.457 (0.149−0.644), and TNM staging 0.536 (0.269−0.711) with NRI >0 (Table 1). These results indicate that our new model has effectively improved the outcome events compared with the previous model.

Constructing a nomogram for predicting OS

A nomogram was constructed to visualise our model. This provided a convenient, personalised tool to predict the 1-, 3-, and 5-year OS in NSCLC patients with BM. Each predictor is associated with a score, and the scores of all predictors were summed together to obtain a total score, from which the OS at 1, 3, and 5 years could be obtained (Fig. 6).

Risk stratification based on our model

The risk score of each patient was calculated, following which they were classified into three groups, namely the low-risk ( n = 75), intermediate-risk (n = 150), and high-risk (n = 75) groups based on the tertiles of their risk score to assess whether our model could accurately assess patient risk. The OS was significantly lower (P < 0.001) in the high-risk group (risk score >2.125) compared with the low-risk (risk score ≤0.789) and intermediate-risk (0.789 < risk score ≤2.125) groups. Additionally, patients in the intermediate-risk group had a lower OS than those in the low-risk group (P < 0.001).

The Lung-molGPA and BSBM models, along with our model, help distinguish the OS of patients. However, when using risk groupings derived from the GPA and RPA models, the differentiation of OS among patients is not fully effective. Specifically, according to the GPA model, there was no statistically significant difference in the OS between patients in the “GPA 1.5−2.5” group and the “GPA 3” group (P = 0.275). Similarly, based on the RPA model, there was no statistically significant difference between the OS of patients in the “class II” group and the “class III” group (P = 0.122).

While the dichotomous risk grouping based on TNM staging can differentiate patients’ OS, it lacks the level of detail provided by our model, which comprises three subgroups. The results demonstrate the superior performance of our prognostic model in accurately distinguishing the prognosis of NSCLC patients with BM (Figs. 7, 8).