Individual Risk Prediction of Gastric Cancer Using Fully Connected Neural Network with Weighted Neighborhood Feature

Gastric cancer is one of the common malignant tumor diseases, with its incidence increasing year after year. The primary issue of preventing gastric cancer is to assess its risk level. Most of the current approaches merely contain histological and demographic information for cancer risk assessment. However, the factors of patient-based information, such as family history, diet, high-risk habits and exposure to environmental carcinogens are often not considered in cancer prediction. To address the issue, we present a fully connected neural network method with weighted neighborhood feature (WNF-FCNN) to improve the accuracy of gastric cancer risk prediction based on patient-based information. We first adopt two feature selection schemes to choose weighted neighborhood feature from the 121 raw medical attributes about gastric cancer. Then, a four-layer FCNN is constructed to model the risk level of gastric cancer with Adam optimization algorithm being used to speed up the convergence during training stages. Empirical evaluations on a real hospital cancer medical dataset containing 50,611 records demonstrate the competitiveness of our approach. The results show that our approach is viable to assess individual gastric cancer risk using patient-based information, with the accuracy rate of 96.2%. Our experimental results also show that the neural network-based approach outperforms other competing approaches in terms of prediction accuracy.