Development of Korean Gastric Cancer Risk Model using Logistic Regression and Neural Network

Abstract

Introduction: High incidence and mortality rates of gastric cancer demands for reduction of incidence and mortality rates by public intervention strategies in high risk groups and general population, or by preventive efforts by individuals. This study aims to comprehensively evaluate the risk and protective factors of gastric cancer and develop a valid model to predict gastric cancer risk. Methods: Candidate predictors were selected combining expert opinion and literature search. Using a case-control study with 4,603 Korean subjects, a logistic regression model for linear regression and a neural network model for nonlinear regression were developed. By comparing the discriminatory ability by AUC, one final model was chosen for further development of absolute risk projection model. The developed model was validated using an independent data, a Japanese population-based cohort study. Results: All 9 factors categorized to be at evidence level of sufficient, probable and possible were used for constructing the final model. Logistic regression model and neural network model did not show a distinct discriminatory ability, AUC of 0.731 vs. 0.732 in men, 0.760 vs. 0.745 in women. For robustness and applicability, logistic regression model was chosen for development of absolute risk model. The model was calibrated using a ratio of expected to observed number of gastric cancer cases, 1.05 (95% CI 0.98 – 1.12) for men and 0.93 (95% CI 0.83 – 1.04) in women. Conclusions: The mathematical model developed in the present study will help predict the occurrence of gastric cancer for an individual considering combined risk factors which will help at a personalized level by enabling early detection and preventive efforts. Moreover, the model can also be used as a source for developing a national guideline for prevention of gastric cancer and a reference to develop future preventive trials.