Skip to main content

Machine Learning Models to Predict 5-Year Post-Surgery Survival of Elderly Gastric Cancer Patients: Multicenter Retrospective Cohort Study

Authors
Xing-Qi Zhang, MD1,2,#, Ze-Ning Huang, MD1,2,#, Ju Wu, MD1,2, Chang-Yue Zheng, MD1,2, Xiao-Dong Liu, MD3, Yan-Bing Zhou, MD, PhD3, Ying-Qi Huang, MD1,2, Jian-Xian Lin, MD, PhD1,2, Qi-Yue Chen, PhD1,2, Ping Li, MD, PhD1,2, Jian-Wei Xie1,2, Chao-Hui Zheng, MD, PhD1,2, Chang-Ming Huang, MD, FACS1,2

# Zhang XQ and Huang ZN contributed equally to this work and should be considered co-first authors.

Affiliations

  1. Department of Gastric Surgery, Fujian Medical University Union Hospital, Fuzhou, Fujian Province, China.
  2. Key Laboratory of Ministry of Education of Gastrointestinal Cancer, Fujian Medical University, 350108 Fuzhou, Fujian Province, China.
  3. Department of General Surgery, Affiliated Hospital of Qingdao University, Qingdao, Shandong Province, China.

Corresponding Author
Chang-Ming Huang (hcmlr2002@163.com)

Background
Oxidative stress plays a critical role in aging and cancer. However, few studies have explored the relationship between oxidative stress and prognosis in elderly patients with gastric cancer (GC). This study aimed to construct and validate a machine learning-based model to predict the five-year overall survival (OS) postoperatively in this demographic population.

Patients and methods
Conducted in four tertiary hospitals, this multicenter retrospective cohort study included 2,185 GC patients aged ≥ 65 years, who underwent radical gastrectomy from January 2012 to April 2018. We developed the GC Oxidative Stress Score (GIOSS) via Cox regression, analyzing the association between biochemical markers changing and prognosis. Predictive models for 5-year OS were constructed using Random Forest (RF), Decision Tree (DT), and Support Vector Machine (SVM), validated and compared using Area Under the Receiver Operating Characteristic Curve (AUC) and calibration plots.

Results
The study comprised 1,367 derivation and 492 external validation cohort patients. Multivariate analysis indicated low GIOSS as a predictor of poor prognosis. The RF model emerged as the most accurate, with AUCs of 0.999 (training set), 0.869 (internal validation set), and 0.796 (external validation set). DT and SVM models showed AUCs of 0.784, 0.796, 0.741, and 0.832, 0.839, 0.789, respectively. Decision curve analysis demonstrated RF's superior net benefit. In the 24-variable RF model, GIOSS was the fourth most significant predictor after pN, pT, and tumor location. Conclusion: Incorporating GIOSS, the RF model effectively predicts postoperative OS in elderly GC patients, offering a novel, robust tool for prognosis in this demographic. Our findings emphasize the importance of oxidative stress in cancer prognosis and provide a pathway for improved management and treatment of GC in elderly patients.

Keywords
Elderly; Gastric cancer; Oxidative stress; Machine learning; Overall survival