ARTIFICIAL NEURAL NETWORKS IMPROVE THE ACCURACY OF CANCER SURVIVAL PREDICTION

BACKGROUND. The TNM staging system originated as a response to the nee d for an accurate, consistent, universal cancer outcome prediction sys tem. Since the TNM staging system was introduced in the 1950s, new pro gnostic factors have been identified and new methods for integrating p rognostic factors have been developed. This study compares the predict ion accuracy of the TNM staging system with that of artificial neural network statistical models. METHODS. For 5-year survival of patients w ith breast or colorectal carcinoma, the authors compared the TNM stagi ng system's predictive accuracy with that of artificial neural network s (ANN). The area under the receiver operating characteristic curve, a s applied to an independent validation data set, was the measure of ac curacy. RESULTS. For the American College of Surgeons' Patient Care Ev aluation (PCE) data set, using only the TNM variables (tumor size, num ber of positive regional lymph nodes, and distant metastasis), the art ificial neural network's predictions of the 5-year survival of patient s with breast carcinoma were significantly more accurate than those of the TNM staging system (TNM, 0.720; ANN, 0.770; P < 0.001). For the N ational Cancer Institute's Surveillance, Epidemiology, and End Results breast carcinoma data set, using only the TNM variables, the artifici al neural network's predictions of 10-year survival were significantly more accurate than those of the TNM staging system (TNM, 0.692; ANN, 0.730; P < 0.01). For the PCE colorectal data set, using only the TNM variables, the artificial neural network's predictions of the 5-year s urvival of patients with colorectal carcinoma were significantly more accurate than those of the TNM staging system (TNM, 0.737; ANN, 0.815; Pt 0.001). Adding commonly collected demographic and anatomic variabl es to the TNM variables further increased the accuracy of the artifici al neural network's predictions of breast carcinoma survival (0.784) a nd colorectal carcinoma survival (0.869). CONCLUSIONS. Artificial neur al networks are significantly more accurate than the TNM staging syste m when both use the TNM prognostic factors alone. New prognostic facto rs can be added to artificial neural networks to increase prognostic a ccuracy further. These results are robust across different data sets a nd cancer sites. (C) 1997 American Cancer Society.