Exploring the survival prognosis of lung adenocarcinoma based on the cancer genome atlas database using artificial neural network.

The aim of this study was to investigate the clinical factors affecting the survival prognosis of lung adenocarcinoma, and to establish a predictive model of survival prognosis of lung adenocarcinoma by artificial neural network.Download the cancer genome atlas (TCGA) database for lung adenocarcinoma research data, perform cox regression analysis and descriptive statistics on the obtained clinical data, draw the survival curve by Kaplan-Meier method, select the independent variables that are statistically significant for constructing the artificial neural networks (ANN) model, and establish artificial neural network model.The number of valid cases included in the study was 524, including 280 men and 244 women, with an age range of 33 to 88 years, mean age 66.87 years, and median progression-free survival (PFS) was 37.7 months. The median overall survival time (OS) was 41.1 months. Cox multivariate analysis showed that smoking history, tumor stage, and surgical margin resection status were independently associated with PFS, and tumor stage and surgical margin resection status were independently associated with OS. The accuracy of the established ANN model itself was predicted to be 65.8%. The accuracy of correctly predicting the prognosis of the predicted samples was 75.0%, and the area under the receiver operating characteristic curve was 0.712.The clinical prognostic factors of lung adenocarcinoma include: smoking history, tumor stage, and surgical margin resection status. The established ANN model can be used to predict the prognosis of lung adenocarcinoma.

Introduction

Lung cancer is the most common cause of cancer-related mortality worldwide, causing >1 million deaths each year.[ Lung adenocarcinoma (LUAD) is the most common histological type of lung cancer, and its effect on recurrence and metastasis is still unsatisfactory.[ Therefore, it is important to improve the rate of early diagnosis and individualized treatment options, which forces clinicians and researchers to continuously study the prognosis of their survival in order to guide the clinical.

The cancer genome atlas (TCGA) database is a joint project of the National Cancer Institute and the National Human Genome Research Institute, allowing researchers to search and download tumor-related data for analyze. The Genomic Data Commons Data Portal (GDC) is the TCGA database data-driven platform. As of November 2018, the GDC platform has released 43 projects, including 33,096 samples.

Artificial neural networks (ANN) is an artificial intelligence model with a highly nonlinear super-large-scale continuous-time dynamic system, a network formed by a large number of processing units (neurons) interconnected. In the field of oncology, ANN is increasingly used in cancers such as gastric cancer, prostate cancer, colon cancer, and breast cancer for diagnosis, differential diagnosis, prognosis, etc.,[ and has been continuously recognized by relevant medical professionals.

In this study, the data of lung adenocarcinoma are deeply explored through the TCGA database, and the clinical data are statistically analyzed to explore the prognostic factors, and an artificial neural network prediction model is established.

Materials and methods

Source of information

Data were collected from the TCGA database (https://portal.gdc.cancer.gov/), and accessed its external link GDC platform (http://www.cbioportal.org/) to obtain data on lung adenocarcinoma (TCGA, provisional) published on the GDC platform. A total of 586 lung adenocarcinoma samples were included in this study. In the total of all samples, 62 cases of clinical information deficiency were excluded and 524 valid data were included.

In this study, a total of 524 patients with pathological diagnosis of lung adenocarcinoma were included, including 280 men and 244 women. Among these patients, the age fluctuated between 33 and 88 years, with an average age of 66.87 years. Three hundred forty six cases of smoking history can be traced, and each tumor stage (IA-IV) can be seen. The lungs of each lung can be seen in the tumor site. After 88.80% of patients had surgery, the pathological grade of the margin was R0, and 43.89% of the cases had different types of gene mutations, including Kras and ALK mutations (Table 1).

Table 1

Clinical data statistics for case samples.

The TCGA database is a joint project in which the receipt collection phase of the database comes from the National Cancer Institute and the National Human Genome Research Institute, and it is ethical. All the data involved in our research are from the public platform TCGA database, and does not involve ethical issues and patient consent. It is not necessary.

Statistical methods

The clinical data obtained by SPSS22.0 software (Chicago, Illinois) were processed and statistically analyzed, including log-rank test and cox regression analysis and descriptive statistics. It was considered that P < .05 had significant statistical difference. Through single factor and multi-factor analysis, the independent variables which are statistically significant for constructing the ANN model are selected, and the artificial neural network model is established. The prediction performance of the ANN model is evaluated by the area under the receiver operating characteristic (ROC) curve (AUC). When AUC >0.5, the more tending to 1, the better the prediction performance.

Results

Univariate and multivariate analyses were performed using COX regression to determine the relationship between the selected clinical variables and progression-free survival (PFS) or overall survival (OS), with bold emphasis on statistically significant differences (Tables 2 and 3).

Table 2

Univariate analysis of overall survival and progression-free survival of case samples.

Table 3

Multivariate analysis of overall survival and progression-free survival of case samples.

In the univariate analysis, we can conclude that tumor stage is associated with PFS and OS in patients with lung adenocarcinoma (P < .001).

In the multivariate analysis, P = .002 < .05, rejecting the null hypothesis, and considering that the partial regression coefficient is not 0, it is worth further analysis. At the significant level of 0.05, smoking history, tumor stage, and surgical margin level have statistics. Learning differences, tips: smoking history, tumor stage, and surgical margin resection status were independently associated with PFS (P = .026; P = .009; .001). Tumor staging and surgical margin resection status were independently associated with OS (P < .001; P = .003).

Through the cox regression univariate and multivariate analysis, the variables with statistical significance for the survival prognosis of lung adenocarcinoma were screened, and the log-rank test was performed by Kaplan–Meier method (Figs. 1 and 2). Smoking history is related to PFS, and different tumor stages have significant differences in survival prognosis, the higher the stage, the shorter the PFS and OS. Different surgical margins also have significant differences in PFS and OS. Therefore, it can be concluded that the smoking history is related to PFS, tumor stage, and different surgical margin resection status are related to PFS and OS. Through the above analysis, the independent variables which are statistically significant for constructing the ANN model are selected and an artificial neural network model is established. The topology includes input layer, hidden layer, and output layer. The data analysis selects “neural network” → “multilayer perceptron.” The input node has 13 neural nodes and 8 hidden layer neural nodes. Transfer function, 2 output neural nodes, corresponding to the survival state: survival = 1, death = 0, transmitted by softmax function, the established ANN model is as follows (Fig. 3).

Figure 1

Relationship between tumor stage and PFS or OS. OS = overall survival, PFS = progression-free survival.

Figure 2

Relationship between surgical margin level and PFS or OS. OS = overall survival, PFS = progression-free survival.

Figure 3

Artificial neural network pattern.

The ANN model was used to predict the training set samples, and the accuracy of correctly predicting the prognosis was 65.8%. The ANN model was used to predict the prediction set samples. The accuracy of correct prognosis was 75.0%, and the area under the ROC curve was 0.712 (Fig. 4).

Figure 4

Artificial neural network ROC curve, prediction-measured map. ROC = receiver operating characteristic.

To evaluate the predictive power of the ANN model, we used SPSS software to reconstruct the traditional binary logistic regression model and compare the 2 models. In this model, “smoking history, tumor staging, and surgical margin resection status” were used as independent variables, and “patient's vital status” was used as a dependent variable to construct a binary logistic regression model. The model was used to predict the 524 cases included. And using the joint probability to draw the ROC curve, the results show that the accuracy of correctly predicting the prognosis is 60.6%, and the area under the ROC curve is 0.662 (Fig. 5).

Figure 5

Binary logistic regression model ROC curve. ROC = receiver operating characteristic.

Discussion

LUAD is the most common subtype of non-small cell lung cancer (NSCLC), accounting for about 40% of all histological types, and is prone to recurrence. For recurrence and metastatic LUAD, the current treatment effect is poor.[ Due to the existence of tumor heterogeneity, the choice of individualized programs and the evaluation of prognosis have important clinical significance. Through the study of the clinical features of LUAD, the assessment of survival prognosis can provide a basis for clinical individualized treatment.

In this study, by cox regression univariate and multivariate analysis, we found that smoking history, tumor stage, and surgical margin resection status were independently associated with PFS. Tumor staging and surgical margin resection status are independently associated with OS, which is consistent with previous reports.[ The differences in other factors (including sex, age, smoking history, location of tumors, and whether there were genetic mutations) were not significant.

In previous studies, there were different reports on whether sex and age were the prognostic factors of lung cancer. Jubelirer et al[ retrospectively analyzed 2207 patients with lung adenocarcinoma, and concluded that sex does not affect the prognosis of lung cancer, and the significant prognosis of death within 5 years. The indicators are staging, grading, age, and histopathology. In the study by Pitz et al,[ the survival rate and prognosis of women with lung adenocarcinoma were significantly better than those of men. In this study, statistical analysis showed that age was not an independent factor affecting prognosis, which was consistent with previous studies.[

The PS score was considered to be an independent prognostic factor for LUAD, which was not covered in this study. In the literature, Kawaguchi et al[ in a study of survival prognosis factors in 26,957 patients with NSCLC, by single factor and multivariate analysis found that the median OS of PS = 0 and PS = 1, respectively for 51.5 and 15.4 months, the difference was significant and the PS score was an independent prognostic factor for NSCLC.

The degree of tumor differentiation and the expression of Ki67 are also considered to be important factors influencing survival prognosis. Zhang et al[ found in a retrospective study of 616 patients with lung adenocarcinoma, the prognosis of patients with poorly differentiated adenocarcinoma is not low. The prognosis of patients with differentiated adenocarcinoma is poor; the prognosis of patients with high expression of Ki67 is significantly worse than that of patients with low or no expression of Ki67.

In addition to the above factors, in recent years, the relationship between miRNA and survival and prognosis of lung adenocarcinoma was found. Lin et al[ used the R language to analyze the differential miRNAs in LUAD and paracancerous tissues by mining the TCGA database. The results showed that miR-101-3p, miR-148a-3p, miR-192-5p, miR-193b-3p, miR-505-3p, miR-584-5p, and miR-99a-5p are associated with prognosis in LUAD patients.

Wang et al[ discovered that the methylation level of methylation site cg12013757 of KRI1 gene has an effect on the prognosis of lung adenocarcinoma by mining the whole genome methylation data of TCGA database. The mRNA expression of corresponding gene is also related to lung. It is related to the prognosis of adenocarcinoma and can be further studied as a biomarker for the prognosis of lung cancer.

The above clinical factors are simple, easy to obtain, and have many clinical applications. In this study, we establish the clinical prognostic factors and construct the binary logistic regression model and the ANN model, respectively. The accuracy of correctly predicting prognosis using a binary logistic regression model was 60.6%, and the area under the ROC curve was 0.662. Compared with the traditional binary logistic regression model, the accuracy of the ANN model's own prediction is 65.8%. The accuracy of correct prognosis for predictive samples is 75.0%, and the area under ROC curve is 0.712. The prediction accuracy of the 2 models is roughly the same. The ANN model is slightly more accurate, and it can provide a new idea for the inference of clinical disease prognosis.

In summary, this study used the existing TCGA database to initially analyze the factors affecting the survival prognosis of LUAD, and established the ANN prediction model, which has higher prediction accuracy for the disease. The TCGA database has a large sample size and a wide variety of cancers, which requires us to further digging deeper to provide more guidance for the clinic.

Author contributions

Conceptualization: Xianrong Xu.

Data curation: Na Jiang.

Formal analysis: Na Jiang.

Funding acquisition: Xianrong Xu.

Methodology: Na Jiang.

Project administration: Xianrong Xu.

Resources: Na Jiang.

Software: Na Jiang.

Supervision: Na Jiang, Xianrong Xu.

Validation: Xianrong Xu.

Visualization: Xianrong Xu.

Writing – original draft: Na Jiang.

Writing – review & editing: Xianrong Xu.