Automatic rule generation techniques supplement conventional statistics in identifying prognostic factors for head and neck cancer.

Aim of the study is to investigate the possibility to supplement the predictive ability of conventional statistics by taking into account the prognostic variables selected by automatic rule generation methods. The analysis involves patients affected by head and neck squamous cell carcinoma, treated by conventional radiotherapy, partly accelerated radiotherapy or combined chemo-radiotherapy. Univariate and multivariate statistic analysis are performed via SPSS, whereas the rule generation techniques considered are decision trees and logical neural networks. For each rule generation technique the prognostic variables are obtained by solving a proper classification problem (overall survival or loco-regional control) and by extracting a set of understandable rules underlying the problem at hand. Bayesian and neural networks are also used as a reference to evaluate the quality of the achieved accuracy. Conventional statistics selects as main predictive variables the dimension of the tumor, the involvement of lymphonodes, and the cancer site. Rule generation methods, besides extending similar results also to other candidate prognostic factors with reasonable accuracy, even when the number of available data is small, are also able to suggest non trivial rules linking the predictive factors.