Development and validation of nomograms for predicting overall survival and cancer-specific survival in elderly patients with locally advanced gastric cancer: a population-based study | BMC Gastroenterology

Gastric cancer (GC) is a complex gastrointestinal malignancy that has the fifth highest incidence of any cancer type worldwide [1]. To date, radical resection is the cornerstone in the treatment of resectable gastric cancer. With the continuous progress of biochemical technology, chemotherapy including targeted drugs has been an emerging trend for practising precision medicine and improving the treatment effects of gastric cancer, but the overall survival rate is still not satisfactory [2]. By 2022, gastric cancer had become the fourth leading cause of cancer-related mortality worldwide [1, 3]. Meanwhile, as the worldwide population ages, the incidence of gastric cancer in elderly patients is increasing [4]. According to statistics, more than 60% of gastric cancer patients are aged 65 years [5]. The treatment of elderly gastric cancer patients (ELGC) consumes a large amount of social and medical resources and increases the heavy burden on families and society. However, few clinical studies have focused exclusively on ELGC, and limited evidence has been mainly derived from subgroup analyses.

ELGC patients have more comorbidities, decreased physiological reserves, and poor tumour immune responses, which eventually lead to immune escape and tumour metastasis [6]. Meanwhile, due to the high degree of malignancy and insidious onset, the majority of patients are in the mid-late stage of the disease when diagnosed. Recent data from the China Gastrointestinal Cancer Surgery Union showed that the proportion of people with locally advanced gastric cancer (LAGC) was as high as 70.8% [7]. In recent years, LAGC has evolved from a single surgical resection to multidisciplinary therapy centering on the role of surgery [8]. However, for elderly patients with LAGC, the clinicopathologic characteristics of these patients and the factors influencing prognosis have not been fully elucidated. Sufficient evidence-based medical evidence is lacking for the surgical treatment of elderly LAGC patients.

To provide optimal therapeutic strategies for this population, the assessment of factors affecting life expectancy has become of tremendous importance. To date, the American Joint Committee on Cancer (AJCC) TNM staging system has been widely used for the assessment of risk stratification and prognosis in oncology [9]. Among them, this staging system for gastric cancer has relied on a limited number of pathological variables (including tumour depth, lymph node metastasis, and distant metastasis), and assumed homogeneity within the same stage groups. In general, the health status of elderly patients with LAGC is highly complex and heterogeneous [10]. Long-term survival is affected by multiple factors, such as sex, tumour stage and pathological state, so relevant studies must combine demographic and epidemiological data [11]. In addition, compared with a single predictor, the establishment of a multivariate prediction model is more likely to increase the sensitivities and specificities of predicting prognosis at the macro level and improve the reliability of the conclusion.

Given the limitations of the AJCC TNM staging system, clinical prediction models (CPMs) have become popular among oncologists and patients as risk assessment tools [12]. On the one hand, CPMs are increasingly able to estimate individual risk based on patient and disease characteristics. On the other hand, CPMs could combine multiple predictors, including molecular, histological and clinical features, to improve the accuracy of prognostic estimates [13]. CPMs include disease occurrence models, diagnostic models and prognostic models [14]. Nomograms, as a common tool in CPMs, have been constructed successfully and proven to be effective in a variety of tumour diseases. For prognostic nomograms, researchers have often assigned corresponding values to different variables, and the total score was transformed into the occurrence probability of the outcome event. After the population was divided according to individual scores, different clinical interventions were implemented. Currently, nomograms have been used to identify high-risk patients, monitor and direct personalized therapeutics and improve the design of clinical trials. Technical guidelines for nomogram development have been published by the AJCC Precision Medicine Core to improve the validity and quality of research on accurate predictive models [12].

Recently, nomograms for predicting lymph node metastasis or prognosis have been widely used in the field of gastric cancer. However, a nomogram for predicting the survival of elderly patients with LAGC has yet to be developed and validated. This study aimed to evaluate multiple factors influencing the survival of gastric cancer patients with LAGC based on a retrospective population-based study. Novel nomograms for predicting the overall survival (OS) and cancer-specific survival (CSS) were developed and validated. On this basis, external validation of the prediction model was carried out to demonstrate its applicability in Asian populations.

Patients and methods

Data sources

This study combined data from two sources. The data source of this retrospective training and internal validation cohort was from the Surveillance, Epidemiology, and End Results Program (SEER). At present, the SEER database consists of cancer registries from 21 geographic areas, covering approximately one-third of the American population [15]. As the largest publicly authoritative data system, the SEER database includes more than 100 sociodemographic and clinical characteristics. Moreover, the SEER data are available to the public for research purposes, and no ethics committee approval or consent procedures are needed.

In addition, data from the Affiliated Hospital of Qingdao University, Qilu Hospital of Shandong University and Shandong Provincial Hospital were used to externally validate the model. According to the prespecified protocol, all medical records were retrieved, and data were extracted by two reviewers to improve the validity. The interrater reliability between the evaluators was found to be excellent (Cohen κ index 0.9). Furthermore, we adhered to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines for cohort studies to ensure the quality of the research [16]. All procedures were approved by the Ethics Committee of the three medical centres. Oral informed consent was obtained from all patients.

Study population

In this study, the clinical features of 123,964 patients with stomach cancer were downloaded between 2000 and 2018 from the SEER database using SEER*Stat software (v8.3.6).

The eligibility criteria were as follows: (1) At or over the age of 65; (2) All patients had been pathologically confirmed to have gastric adenocarcinoma by preoperative gastroscopy biopsy or postoperative pathology; (3) All patients underwent radical (R0) surgical treatment consisting of gastrectomy and lymphadenectomy; (4) Histologically proven locally advanced gastric cancer patients (T1-2N + M0 or T3-4NanyM0); and (5) All the patients had complete follow-up data. The exclusion criteria were as follows: (1) Patients under the age of 65 or early gastric cancer (EGC); (2) Patients with multiple tumours, or distant metastasis; (3) Patients with confirmed pathology diagnosis of nonadenocarcinoma, GIST or a neuroendocrine tumour; (4) Patients who did not undergo gastrectomy or underwent partial gastrectomy; (5) All-cause mortality within 30 days of surgery; and (6) Patients with incomplete clinical data (medical records or follow-up data).

The data screening process is shown in the flow diagram (Fig. 1). A total of 4991 eligible patients were included in this study. The elderly patients with LAGC were randomly divided into a training cohort (n = 3494) and an internal validation cohort (n = 1497) with an allocation of 7:3 ratio by R software. In the external validation set, 841 elderly patients with LAGC at the three medical centres were retrospectively collected and reviewed between January 2015 and December 2018.

Flow chart of patient selection

Observation indicators and endpoints

The main observation indices of this study included the demographics of the patients (sex, age, race, marital status at diagnosis), the clinicopathological features of the cancer (tumour location, size, tumour differentiation, histology, gastrectomy type, depth of invasion, lymph node metastasis, distant metastasis, tumour stage, chemotherapy record) and survival data (survival time and death reason). According to the specific circumstances and goals of the study, as well as the nature of the data and the relationship between the variables, this study converted the continuous variables into categorical variables in the regression analysis. It should be noted that categorizing continuous variables can be helpful in cases where there is a nonlinear relationship between the predictor and outcome variables, and it may be difficult to find a suitable model to fit. While splines can be used as an alternative, they can be computationally intensive. X-tile is a bioinformatics tool for risk factor assessment assessment and outcome-based cut-point optimization. As an alternative, the optimal cut-off values of age, tumour size and lymph node ratio (LNR) were determined using the X-tile program (X-tile software version 3.6.1, Yale University) [17], and the continuous variables were converted into classification variables.

Age was categorized into three groups: 65 ~ 70 years old, 71 ~ 80 years old, and ≥ 81 years old. Race was divided into four groups: white, black, Asian or Pacific Islander and Indian or unknown. The two marriage categories were married and unmarried (including single, widowed, divorced and informal union). Tumour size was divided into three groups (< 3.5, 3.5 ≤ tumour size < 9.5, and ≥ 9.5). The location of the tumour was divided into the cardia/fundus, body and antrum/pylorus. Tumour differentiation was defined according to the cellular differentiation degree, which may be classified as I-II and III-IV [18]. Pathology type was classified as adenocarcinoma and signet-ring cell carcinoma. The type of surgery included proximal gastrectomy, distal gastrectomy and total gastrectomy. Cancer stage was categorized according to the Staging Manual of the AJCC [9]. The positive rate of lymph node metastasis was classified into 2 groups with a cut-off of 33%. In addition, cause-specific survival (CSS) and overall survival (OS) were used as the main study endpoints. In this study, CSS was defined as the time from gastric cancer diagnosis until gastric cancer-related death or end of follow-up. OS was defined as the time to death from any cause or the end of follow-up.

Development and validation of the nomogram

Univariate and multivariate analyses of the prognostic values were performed using the Cox proportional hazards regression model, which was fundamental to the survival prediction model. Factors with P < 0.10 in the univariate analyses were entered into the multivariate regression model. The covariates included in the nomogram models were selected based on the independent risk factors affecting survival. Thereafter, nomograms predicting 1-, 3- and 5-year OS as well as 1-, 3- and 5-year CSS were constructed using the “rms” package (6.2–0) of R software 3.5.0.

The goal of a forecasting model is to predict the outcome as quickly and accurately as possible. The predictive power of the nomogram was assessed by both discrimination and calibration [19]. Discrimination referred to the ability to separate patients with different outcomes and used the Harrell’s concordance index (C-index) as the measurement tool [20, 21]. Moreover, the C-index and 95% confidence interval (CI) were calculated on the basis of bootstrap resampling with 1000 replicates. A C-index of 1 indicated perfect discrimination, and a C-index of 0.5 indicated that the model was not better than random chance. The calibration of the models could be assessed using a calibration chart, which was used to evaluate the difference between the predicted probability and the actual result, and the 45-degree line denoted the optimal prediction [20]. To avoid overfitting, fivefold cross-validation was adopted for the nomogram model. Finally, the clinical usefulness of the nomogram was the last component in evaluating the value of the nomogram. Decision curve analysis (DCA) was utilized to investigate whether the nomogram-assisted decisions effectively improved the outcome for individual patients [22].

Statistical analysis

The randomization sequences were generated using the RANDBETWEEN function in Microsoft Excel. The difference distribution of the categorical variables between the subgroups was assessed using Pearson’s χ² test and Fisher’s exact test. OS and CSS curves were plotted using the Kaplan–Meier method, and the differences between the survival curves were evaluated with the log-rank test. Receiver operating characteristic (ROC) curves were plotted by SPSS, and the area under the ROC curve (AUC) was determined to evaluate the accuracy of the model. All statistical graphics and analyses were performed using SPSS software (IBM SPSS Statistics 26.0) or R software (Version 3.5.0). P values < 0.05 were regarded to indicate statistically significant differences.