Development and validation of prognostic models for small cell lung cancer patients with liver metastasis: a SEER population-based study

The retrospective cohort study was based on the SEER program which covers approximately 30% of the total US population [16]. The records of SCLC patients with liver metastasis between 2010 and 2015 were extracted from the database ‘SEER 18 Regs Custom Data (including additional treatment fields), November 2018 sub (1975–2016 varying) database using SEER*stat 8.3.5 software. The International Classification of Diseases for Oncology third edition (ICD-O-3) was used to identify SCLC by site codes [8002, 8041, 8043, 8144, 8145] [17]. The inclusion criteria were as follows: (I) SCLC was the only primary cancer; (II) the staging of lymph nodes followed the 7th edition of the American Joint Committee on 2 Journal of International Medical Research Cancer; (III) aged ≥ 18 years old; (IV) patients with liver metastasis [SEER Combined Mets at DX-liver (2010 +)]. Since the clinical data in this study were collected from a publicly available database, there were no local or state ethical issues. In addition, because this retrospective study was based on public data from the SEER database, informed consent was not required. All methods were performed in accordance with the relevant guidelines and regulations.

The database reviewed retrospectively consisted of clinical characteristics of patients, pathological characteristics of tumors, and survival time (months). Continuous variables were transformed into categorical variables based on recognized cut-off values (for age). Clinical characteristics of patients included gender (female vs male), age (≤ 65 years, > 65 years), race (black, white, and other), and marital status (married and other). Pathological characteristics of tumors include primary site [multiple sites, upper lobe, middle lobe, lower lobe, main bronchus, not otherwise specified (NOS)], tumor size (≤ 24 mm, 25–37 mm, 38–59 mm, ≥ 60 mm), T stage in 7th edition AJCC system [T1, T2, T3, T4, and not specific known T stage (Tx)], N stage in 7th edition AJCC system [N1, N2, N3, N4, and not specific known N stage (Nx)], metastatic sites (bone, brain, lung, multiple other metastases, no metastases or unknown), surgery of primary site or not/unknown, surgery of metastatic site or not/unknown, radiotherapy or not/unknown, chemotherapy or not/unknown. The outcomes in this study were 1- and 2-year CSS and 1- and 2-year OS. The classification of tumor size was based on the inter-quartiles ranges. The outcomes in this study were 1- and 2-year CSS and 1- and 2-year OS. CSS was defined as the time between diagnosis and death owing to specific cancer during follow-up and OS was defined as death regardless of any causes. The total observation period was 2 years; follow-up would terminate if the patient died. The median follow-up time was 3 months.

For the development of the nomograms, all of the 8,587 cases were randomly divided into the training set and test set (ratio: 3:1) by using the random-number generation method. The prediction models were established on the basis of the predictive factors identified by Cox proportional hazards model. The performances of the prediction models were evaluated by measuring the concordance index (C-index),calibration plots, and decision curve analysis (DCA).

Measurement data by normal distribution are described in mean ± standard deviation (Mean ± SD). The independent sample t-test or analysis of one-way variance (ANOVA) was used for comparisons between groups. Non-normal data were described as a median and interquartile range [M (Q₁, Q₃)], and the Mann–Whitney U test or Kruskal–Wallis test was applied for comparisons between groups. Enumeration data were described as the number of cases and the constituent ratio [N (%)]. The Chi-square test was used for comparison between groups.

Hazard ratios (HRs) and their 95% confidence intervals (CIs) for each potential predictive variable were analyzed using Cox proportional risk models. The variables with multiple categories were transformed into dummy variables before further analyses. False-discovery rate (FDR) adjusted P values were calculated to correct for multiple testing. Based on the predictive factors, prediction models were conducted to predict the 1- and 2-year CSS and OS. The performances of the final nomogram were assessed by C-index, and calibration measures as the measuring tool. The C-index is a concordance measure analogous to ROC, which values range from 0.5 (no discrimination) to 1.0 (perfect discrimination). The higher the value between 0.5 and 1, the stronger the resolution of the nomogram. Calibration measures to what degree the predicted probabilities are close to actual outcomes. Calibration plots of the nomogram for 1- and 2-year CSS and OS were performed in the training set and the testing set. The prediction models were also verified by 5-fold cross validation. The 5-fold cross-validation was done to make the results more realistic and to avoid chance. The data set was divided into five groups by a 5-fold cross-validation operation. For each training session, one set was used as the validation set and the remaining four sets were used as the training set. After addressing the ability of the nomogram, we used DCA to test the reliability of the model, which was a method for evaluating alternative diagnostic and prognostic strategies that have advantages over other commonly used measures and techniques. If the threshold probability of net benefits is unpractical, then the applicability of a well-performing model may be limited, meaning that the benefits of the new prediction model will be less than the benefits of existing tools, and may even be detrimental. The prediction models with and without chemotherapy for predicting CSS and OS in SCLC patients with liver metastasis were conducted and were compared with the prediction models. Data analysis used R software version 4.1.2 (R Foundation). All tests were two-tailed and P < 0.05 was considered statistically significant,

In total, 8,587 eligible patients, who were diagnosed as SCLC patients with liver metastasis from 2010 to 2015 were identified in the SEER database. The flow chart of patient’s selection is shown in Fig. 1. Among them, 154 patients experienced CSS, and 154 patients experienced OS. The median follow-up time was 3 months. For all patients, there were 3,770 (43.90%) patients with age ≤ 65 years and 4,817 (56.10%) patients with age > 65 years. As for the race, the majority of patients (88.84%) were white. With respect to the primary site, most of them were located in the upper lobe. The majority of patients were classified as T4 (33.42%) and N2 (55.65%). The baseline characteristics of the patients are presented in Table 1.

Identifications of factors associated with CSS and OS are shown in Table 2. The results demonstrated that age > 65 years (HR: 1.303, 95% CI: 1.238–1.372, P < 0.001), being male (HR: 1.113, 95% CI: 1.057–1.171, P < 0.001), being married (HR: 0.907, 95% CI: 0.863–0.955, P < 0.001), Nx stage (HR: 1.264, 95% CI: 1.115–1.433, P < 0.001, FDR-adjusted P < 0.001), lung metastases (HR: 1.189, 95% CI: 1.075–1.316, P = 0.001, FDR-adjusted P = 0.004), multiple metastases (HR: 1.115, 95% CI: 1.031–1.206, P = 0.007 FDR-adjusted P = 0.021), surgery of metastatic site (HR: 0.927, 95% CI: 0.776–1.109, P = 0.407), chemotherapy (HR: 0.232, 95% CI: 0.219–0.246, P < 0.001), and radiotherapy (HR: 0.627, 95% CI: 0.593–0.664, P < 0.001) were associated with CSS in SCLC patients with liver metastasis. Age > 65 years, being male, being married, Nx stage, lung metastases, multiple metastases, no metastases/unknown, surgery of metastatic site, chemotherapy, and radiotherapy were associated with OS in SCLC patients with liver metastasis.

We developed two nomograms respectively for CSS and OS. Each of the variables was given a point according to the HR. Then, by adding the total score of each variable and locating the score on the total points scale, the 1- and 2-year CSS and OS could be obtained. The nomograms containing independent predictive factors for predicting 1- and 2-year CSS and OS prediction of SCLC patients with liver metastasis are shown in Figs. 2 and 3.

In the training set, the C-index for CSS predicted by the prediction model was 0.724 (95% CI: 0.716–0.731), whereas the C-index for OS was 0.732 (95% CI: 0.724–0.739) (Table 3). In the testing set, the C-index for CSS and OS were 0.734 (95% CI: 0.722–0.747), and 0.739 (95% CI: 0.727–0.751), respectively. The C-index of the prediction model for predicting 1-year CSS, and 2-year CSS was 0.727 (95% CI: 0.719–0.734), and 0.724 (95% CI: 0.717–0.731), respectively, in the training set. The C-index of the prediction model for predicting 1-year OS, and 2-year OS was 0.734 (95% CI: 0.727–0.742), and 0.732 (95% CI: 0.725–0.739), respectively, in the training set. C-indexes of the prediction models for CSS and OS in SCLC patients with liver metastasis are presented in Table 3. The 5-fold cross-validation results are shown in Table 4, which showed a similar performance as the prediction models conducted. Calibration and DCA curves show that the nomogram has good predictive accuracy and value. The results of calibration are shown in Fig. 4. The DCA curve for predicting 1- and 2-year CSS and OS prediction of SCLC patients with liver metastasis is presented in Fig. 5.

The nomogram indicated that chemotherapy carries significantly more importance compared to other variables, thereby, comparisons of prediction models with and without chemotherapy for predicting CSS and OS in SCLC patients with liver metastasis were performed. The C-index of the prediction models with only chemotherapy was 0.687 (95% CI: 0.681–0.693) for predicting CSS in the training set, and the C-index of the prediction models without chemotherapy was 0.613 (95% CI: 0.604–0.621) (Tables 5 and 6).