Application of Survival Quilts for prognosis prediction of gastrectomy patients based on the Surveillance, Epidemiology, and End Results database and China National Cancer Center Gastric Cancer database

Abstract

Objective

Accurate prognosis prediction is critical for individualized-therapy making of gastric cancer patients. We aimed to develop and test 6-month, 1-, 2-, 3-, 5-, and 10-year overall survival (OS) and cancer-specific survival (CSS) prediction models for gastric cancer patients following gastrectomy.

Methods

We derived and tested Survival Quilts, a machine learning-based model, to develop 6-month, 1-, 2-, 3-, 5-, and 10-year OS and CSS prediction models. Gastrectomy patients in the development set (n = 20,583) and the internal validation set (n = 5,106) were recruited from the Surveillance, Epidemiology, and End Results (SEER) database, while those in the external validation set (n = 6,352) were recruited from the China National Cancer Center Gastric Cancer (NCCGC) database. Furthermore, we selected gastrectomy patients without neoadjuvant therapy as a subgroup to train and test the prognostic models in order to keep the accuracy of tumor-node-metastasis (TNM) stage. Prognostic performances of these OS and CSS models were assessed using the Concordance Index (C-index) and area under the curve (AUC) values.

Results

The machine learning model had a consistently high accuracy in predicting 6-month, 1-, 2-, 3-, 5-, and 10-year OS in the SEER development set (C-index = 0.861, 0.832, 0.789, 0.766, 0.740, and 0.709; AUC = 0.784, 0.828, 0.840, 0.849, 0.869, and 0.902, respectively), SEER validation set (C-index = 0.782, 0.739, 0.712, 0.698, 0.681, and 0.660; AUC = 0.751, 0.772, 0.767, 0.762, 0.766, and 0.787, respectively), and NCCGC set (C-index = 0.691, 0.756, 0.751, 0.737, 0.722, and 0.701; AUC = 0.769, 0.788, 0.790, 0.790, 0.787, and 0.788, respectively). The model was able to predict 6-month, 1-, 2-, 3-, 5-, and 10-year CSS in the SEER development set (C-index = 0.879, 0.858, 0.820, 0.802, 0.784, and 0.774; AUC = 0.756, 0.827, 0.852, 0.863, 0.874, and 0.884, respectively) and SEER validation set (C-index = 0.790, 0.763, 0.741, 0.729, 0.718, and 0.708; AUC = 0.706, 0.758, 0.767, 0.766, 0.766, and 0.764, respectively). In multivariate analysis, the high-risk group with risk score output by 5-year OS model was proved to be a strong survival predictor both in the SEER development set (hazard ratio [HR] = 14.59, 95% confidence interval [CI]: 1.872–2.774, P < 0.001), SEER validation set (HR = 2.28, 95% CI: 13.089–16.293, P < 0.001), and NCCGC set (HR = 1.98, 95% CI: 1.617–2.437, P < 0.001). We further explored the prognostic value of risk score resulted 5-year CSS model of gastrectomy patients, and found that high-risk group remained as an independent CSS factor in the SEER development set (HR = 12.81, 95% CI: 11.568–14.194, P < 0.001) and SEER validation set (HR = 1.61, 95% CI: 1.338–1.935, P < 0.001).

Conclusion

Survival Quilts could allow accurate prediction of 6-month, 1-, 2-, 3-, 5-, and 10-year OS and CSS in gastric cancer patients following gastrectomy.