A Prognostic Nomogram via Competing Risk Model with Patterns of Extracranial Metastasis in Elderly NSCLC Patients with Synchronous Brain- metastasis: A Large Population-based Study

Mingwei Zhang The First A liated Hospital of Fujian Medical University Zijing Zhu Fujian Normal University Wenying Jiang The Third A liated Hospital of Soochow University Shaoli Peng The First A liated Hospital of Fujian Medical University Weitong Zhou The First A liated Hospital of Fujian Medical University Jinfu Zhuang The First A liated Hospital of Fujian Medical University Jinsheng Hong The First A liated Hospital of Fujian Medical University Xiaoping Chen (  xpchen@fjnu.edu.cn ) Fujian Normal University


Introduction
Globally, lung cancer (LC) is the number one cause of cancer-associated death [1]. Most LC-related deaths are due to metastasis, with brain being a common site. Non-small cell lung cancer (NSCLC) is the most prevalent subtype and comprises about 85% of cases [2]. The incidence of brain metastasis (BM) in NSCLC is growing due to imaging technique advancements, improved survival from novel regimens and population aging [3]. The conventional therapies for NSCLC with BM include surgery, radiation and chemotherapy with a median overall survival (mOS) ranging from 4.0 to 31.0 months [4]. The introduction of tyrosine kinase inhibitors and the immune checkpoint inhibitors has improved the 5-year survival rate from 30-60% in selected patients with metastatic NSCLC [5]. NSCLC with de novo BM has been recognized as an entity with heterogeneous prognosis[6] and its risk strati cation is a critical component of standard management.
Several prognostic models have been applied to predict BM patients' outcomes, such as the Disease-Speci c Graded Prognostic Assessment (DS-GPA). Besides age, DS-GPA considers four clinical factors: Karnofsky Performance Status (KPS) [7], extracranial metastasis (ECM) and BM lesions. Genomic and proteomic biomarkers have been reported to stratify patients, which necessitates invasive biopsy and costly molecular sequencing [6]. The invasive nature of biopsy and the high cost of genomic tests limit their large-scale application. Meanwhile, utilizing readily available clinical data, recombination of the M1 stage for stage IV breast cancer can forecast prognosis and effects of primary surgery better [8].
This study on breast cancer has added additional details to DS-GPA, but no such research to modify DS-GPA has been conducted for NSCLC yet. While more ECM lesions have proven to have a worse prognosis in NSCLC patients with BM [9], the subdivision of ECM, other than the total number, might more su ciently address the medical challenges in NSCLC with BM. Noteworthy, half NSCLC patients are over 70-years-old [10]. Current epidemiological trends make LC in the geriatric population increasingly recognized as a public health concern [11]. Yet, their prognostic estimates mainly stem from younger, healthier patients enrolled in clinical trials [12], because patients over 70 years old have historically been excluded from major landmark trials [13]. According to the WHO and published estimation, cancer is not the number one cause of death in people older than 70-years [14]. Compared with the Cox proportional hazards model, the competing risk model improves the strati cation of senior patients with cancer by the risk of dying from cancer relative to overall mortality [15].
A hospital-based cohort study of elderly NSCLC patients with de novo BM would be extremely challenging, regarding to obtain samples su cient to detect ECM patterns' impact on survival. Due to limitations of registries, the incidence of BM remains indistinct. Yet, the Surveillance Epidemiology and End Results (SEER) program revealed data relevant to with or without BM at diagnosis of malignancies in 2016 [16]. Given that it is a large population-based database, selection bias is largely reduced. Exploiting NSCLC cohorts in SEER datasets, we analyzed the prognostic value of the ECM patterns in septuagenarians and octogenarians with de novo BM and plotted a nomogram for clinical usage.

Patients
The study data set was retrieved from the SEER cancer registries, which are freely accessible via SEER*Stat 8.3.8 (https://seer.cancer.gov). The SEER registry comprises 18 cancer registries across America, taking up about 30% of the US population. The SEER data are periodically updated with clinical information, such as tumor basics, therapies, secondary tumors, demographics and cause of death. The present study was conducted in compliance with the Helsinki Declaration. Given that the SEER registry contains only de-identi ed information, ethical review and informed consent requirements were waived. Given that elderly metastatic lung cancer patients rarely underwent surgeries, those who received surgery (< 5%) were excluded from this study to improve the homogeneity of the study population. Other exclusion criteria were: 1) aged younger than 70 years; 2) with prior or subsequent cancers; 3) underwent surgical treatments; 4) less than one month of follow-up; 5) patients with unavailable information regarding survival duration, status, cause of death or other important items.
We included clinical characteristics, such as gender, race, age, American Joint Committee on Cancer (AJCC) T and N stage, histologic type, chemoradiotherapy, and ECM involvement. The included patients were then randomly divided into a training and a hold-out test set in a 50/50 ratio. The former was used to train the weights of the model, while the latter was to validate the predictive ability of the trained model.

Statistics
Based on the accessibility of the SEER databank, ECM involvement in de novo brain-metastatic NSCLC patients were separated into eight categories: none, bone alone, liver alone, lung alone, both bone and liver, bone and lung, liver and lung, or all of the three. The main outcome was disease-speci c survival (DSS), and a competing risk is an event whose occurrence precludes the occurrence of the primary event of interest.
The univariable and multivariable analyses of the competing risk model were conducted via Fine and Gray's regression, which led to the recombination of extracranial involvements for further investigation.
Multivariable analysis of the competing risk model was recalculated in the training, test, and whole sets, followed by subgroup analysis and interaction tests. Multivariate analysis of overall survival (OS) was carried out via Cox proportional hazards model. Sensitivity analyses were done to support the primary analysis. Cumulative incidence curves for DSS and OS were estimated via Gray's and log-rank tests, respectively.
A nomogram of independent factors predicting DSS was depicted based on multivariate analysis of the competing risk model. A stepwise variable selection method was applied. For calibration, estimated probabilities were contrasted with the actual ones. Time-dependent ROC curves were used to judge discrimination. Decision curves analysis (DCA) was applied to visualize the clinical bene ts of alternative models by calculating net bene ts under different threshold probabilities. Curves of therapy plan (highest cost) and no plan (no bene t) were used as references.
Chi-square and Fisher exact test were used for categorical data. Statistical signi cance level was set at a P value (two-tailed) < 0.05, except for those that were speci cally stated. Chi-square, Fisher exact test, Cox proportional hazards model, Kaplan-Meier method, Gray's test, and Fine-Gray regression model were performed using R (version 3.4.3; The R foundation for statistics, Vienna, Austria).

characteristics
In total, 4,974 cases were identi ed as de novo brain-metastatic NSCLC aged over 70 years, with both training and hold-out test sets allocated 2,487 cases each. The most frequently involved organs were bone, contra-lateral lung and liver. Clinicopathologic characteristics were well-balanced between the two data sets. Speci c basic characteristics are outlined in Table 1 and supplemental Table 1.

Correlation between patterns of distant metastasis and DSS
As depicted in

Subcategories of M1 cohort
Based on independent prognostic factors from multivariable analysis of competing risk model, the eight groups of extracranial involvements were recombined into MA, MB and MC subcategories (Supplemental Table 2). The group with all three lesions topped the list, while the group without ECM ranked at the bottom. Given that the SHR of the other three groups with liver metastasis was higher than that of the bone solo group, they ranked higher. The two groups without independent prognosis ranked after the bone solo group. Only two of the seven P values were < 0.05, making the cutoff among MA, MB, and MC.

Multivariable analysis of the competing risk model
Validation of the M1 subcategory system is described and depicted in supplemental

Cumulative incidence curves of DSS and OS
Cumulative incidence curves plotted the mortality augments across M1 subcategories in training, test and whole data sets. Again, the validation of the M1 subdivision was con rmed (Fig. 1).

Subgroup analysis and interaction tests
To study the interaction between treatment mortality and M1 subcategories, MB and MC were combined and compared to MA (Fig. 2)

Kaplan-Meier curves and the Cox proportional hazard model
Kaplan-Meier curves were depicted to investigate survival differences between the covariates (Fig. 3). Compared to MA, MB and MC were associated with increased risk for OS (Supplemental Table 4

Construction and validation of a competing risk nomogram in the training set
A competing risk nomogram was built via the Fine and Gray model to predict the 6-month, 12-month and 18-month cumulative death probabilities (Fig. 4A). Calculation details are listed in Supplemental Table 5.
As depicted in Fig. 4B, the calibration curves for DSS were close to the standard curves. The ROC curve used to assess the nomogram of DSS is shown in Fig. 5. AUC of DSS was 0.757, 0.748 and 0.738 for 6month, 12-month and 18-month prediction (Fig. 4C), respectively. DCA analysis proved the clinical value of this model for 6-and 12-month prediction. (Fig. 4D-F).

Validation of the nomogram in the hold-out test set
The calibration curves for DSS in the hold-out test set were plot in Fig. 5A. Again, calibration curves were close to standard curves. The ROC curve used to assess the nomogram of DSS is shown in Fig. 5. AUC of DSS was 0.778, 0.769 and 0.756 for the 6-month, 12-month, and 18-month prediction (Fig. 5B), respectively. The DCA analysis indicated the good value of this model for 6-and 12-month predictions.

Discussion
In the present study, the Fine and Gray proportional sub-distribution hazard model revealed that the anatomic extent of ECM is an independent prognostic factor for DSS in NSCLC with de novo BM. After recombination of ECM into three subtypes (MA, MB and MC), multivariable analysis based on the competing risk model was re-performed, with ECM subdivision, AJCC T and N, chemoradiotherapy, and histology found signi cantly associated with DSS. Furthermore, these independent predictive factors were employed to create 6-month, 12-month, and 18-month nomograms for DSS. While in the training cohort, the AUC for the 6-month DSS prediction reached 0.757, it reached 0.778 in the internal validation cohort. The calibration curves did not deviate from the reference line. Decision curve analyses (DCA) revealed the net bene t of the nomogram for clinical utility.
The TNM 8th classi cation of LC addressed the existence of synchronous oligo-metastatic patients with a single extrathoracic metastatic lesion (M1b), which is almost the same in prognosis to those presenting with the intra-thoracic metastatic disease only (M1a) [17]. In terms of prognosis, these two groups stood out from the other metastatic population (M1c), which involved multiple metastatic organs. Identi cation of this new category can be related to several factors, including improvement in diagnostic and imaging accuracy, the continuous increase in systemic oncological treatment e cacy, and the favorable tolerability of local ablative treatments [18]. Given the outcome variability within M1c stages, identifying independent, but complementary prognostic factors is of intense interest [19]. As mentioned in the introduction, various models have been proposed to prognosticate patients with brain metastasis [20][21][22].
The DS-GPA incorporates the presence/ absence of ECM, but a wide range of individualized survival probability was found within each DS-GPA class. In addition to its presence alone, more ECM lesions were associated with a poorer prognosis in NSCLC patients with BM [9,23].
According to our analyses, ECM patterns in NSCLC with BM is quali ed to stratify prognosis. Patients were divided into MA (neither liver, bone, nor lung involvement), MB (without liver involvement), and MC (with liver involvement). The MA subcategory had relatively favorable prognosis, while the MC subcategory had signi cant survival disadvantage. Our strati cation is consistent with previous research [9,24]. It has been proposed that distant metastases in LC are nonrandom and there may be speci c patterns of distant metastases [25]. According to the seed and soil hypothesis [26], there are complex ways for disseminated cells to interact with the host organ microenvironment, hence different interactions may result in distinctive patterns of metastatic events. According to our analyses, once the liver becomes involved, patient survival would be impaired. One reported theory is that, as the liver is an immunosuppressive organ when metastases occur, liver metastasis impedes the liver's immune surveillance of other ongoing metastases [27]. Similarly, NSCLC patients with metastasis to their brain and liver were more likely to progress when treated with PD-1 inhibitors. with cancer-speci c death to overall mortality [15,29]. Although the competing risk model was less than perfect, it was validated in our data.
The nomogram allowed for individualized estimation of DSS and highlighted the potential to tailor predictions to individual patients rather than group-wise estimations from other models like the DS-GPA.

Conclusion
In conclusion, the extracranial subdivision of M1 categories harbors the type and number of metastatic sites and provides extra information for prognosis in septuagenarians and octogenarians with De Novo  Competing risk nomogram predicting 6-month, 12-month, and 18-month cumulative probabilities of DSS in elderly NSCLC patients with synchronous BM (A), calibration curves and ROC curves with AUC for 6month, 12-month, and 18-month prediction (B and C), DCA curves for 6-month (D), 12-month (E), and 18month (F) prediction, in the training set.

Figure 3
Validation of the nomogram in the hold-out test set: calibration curves and ROC curves with AUC for 6month, 12-month, and 18-month prediction (A and B), DCA curves for 6-month (C), 12-month (D), and 18month (E) prediction, in the hold-out test set.