Reti neurali artificiali per lo studio di fenomeni complessi: limiti e vantaggi delle applicazioni in biostatistica

The application of artificial neural networks for statistical modelling has been diffused in several fields. This class of models is oriented to inductive inference: the estimate of an unknown functional dependence relationship on the basis of a limited number of experimental observations. The largest developments have been achieved for multivariate non-linear regression, by the adoption of feed forward artificial neural networks for the flexible modelling of the effects of the independent variables. On the hand, methodological studies have enlighten good statistical properties of these models; on the other hand several eurystic applications, mainly in the biomedical field, have stimulated some criticism. This paper introduces basic statistical aspects which characterize feed forward artificial neural networks with specific reference to biostatistical problems. The extension of Generalized Linear Models as neural networks for processing censored survival data will be considered. Limits of the eurystic use and advantages of the integration with traditional statistical models will be finally discussed.