Usefulness of Neural Network as a Novel Method of Predicting Outcome for Gastric Cancer Patients Compared with Logistic Regression

Iwao Kumazawa, Takamasa Hiraoka, Yoshihiro Kawaguchi, Katsuyuki Kunieda, Takao Umemoto and Shigetoyo Saji

Second Department of Surgery, Gifu University School of Medicine

Introduction: To establish tailor-made therapy for gastric cancer patients, we evaluated the usefulness of a neural network (NN), a computer-based mathematical model superior in pattern recognition. Materials and Methods: Predictions of 1- and 3-year survival were compared between the NN and logistic regression (LR) retrospectively using 672 gastrectomized patients with stomach cancer. As prognostic factors, we selected peritoneal metastasis (P), hepatic metastasis (H), invasion depth, lymph node metastasis (n), curability, lymph node dissection (D), age, histlogy, INF, ly and v, and categorized them into 21 dichotomous (0, 1) variables to suit each model. We then evaluated accuracy using a 2×2 matrix and Az (area under the receiver operating characteristic curve). Both models were tested using the "leave-1-out" method. Results: The accuracy of 1-year survival predicted by the NN was significantly better than that of LR (training data: NN 90.0%, LR 86.8%, test data: NN 88.1%, LR 85.3%; p<0.01). The accuracy of 3-year survival predicted by the NN was relatively better than that of LR (training data: NN 85.3%, LR 83.9%, test data: NN 83.0%, LR 82.7%). The Az of the NN was statistically similar to that of LR. Conclusion: The neural network showed superior or similar predictions for gastric cancer patients compared logistic regression. The neural network may thus be used as an index for deciding the risk of individual postoperative patients.

Key words
Prediction of postoperative survival, gastric cancer, neural network, logistic regression, receiver operating characteristic analysis

Reprint requests
Iwao Kumazawa Second Department of Surgery, Gifu University School of Medicine 40 Tsukasamachi, Gifu, 500-8705 JAPAN

Accepted
January 31, 2001

To read the PDF file you will need Abobe Reader installed on your computer.