BREAST CANCER PATIENT SURVIVAL PREDICTION--A NEURAL NETWORK APPROACH

点击翻译按钮获取中文摘要

In FY91 we reported preliminary experiences in developing a back-error propagation Artificial Neural Network (ANN) to predict survival with a small (170 cases) breast cancer database using as explanatory covariates tumor size, number of positive lymph nodes, histologic grade, and estrogen and progesterone receptor status. We concluded that individual patient survival curves could be computed and that bootstrapping methods could be used to compute survival confidence intervals with larger databases. Progress in FY92 included a comparison of the Cox regression model with an ANN approach to survival analysis, leading to the conclusion that an ANN approach would not be constrained by the proportional hazards assumption of the Cox model, thus suggesting factors for which predictive associations are not yet known. A collaboration was begun with Dr. Donald Henson (NCI), present chairman of the American Joint Committee on Cancer (AJCC). Dr. Henson and the AJCC are very interested in the application of appropriate computing methodologies to prognostic factors for evaluation and use in patient outcome prediction and management. A program was written for computing group actuarial survival functions based on the assumption of equal interval hazard rates for censored and non-censored events. Studies were conducted with data extracted from the NCI Surveillance, Epidemiology, End Results (SEER) program. A 6,000-case melanoma database was used to demonstrate basic concepts in ANN survival prediction. The back-error propagation ANN developed last year was refined and applied to a 44,000-case breast cancer database to demonstrate ANN survival prediction based on multi-explanatory factors. Presentations of methods and results were made at the AJCC annual Meeting in January 1992 in San Diego and before the AJCC Task Force on Multiple Prognostic Factors in Chicago in June 1992.