OPTIMIZATION OF AN ARTIFICIAL NEURAL NETWORK USED FOR THE PROGNOSTIC OF CANCER PATIENTS

Abstract

An artificial neural network model was developed for providing an estimation of the survival time for cancer patients. Data from 31 dogs and cats treated as oncologic patients was used for training and validating the network which used 28 elementary predictors selected from clinical and para-clinical, macroscopic pathology, and histology information. The network was optimized to avert the overfitting, which blocks the learning process. Among the techniques tested and illustrated are: random noise in synapses, automatic cloning of well performing neurons, extended learning with/without jitter, jogging of connection weights, freezing of weights and biases, pruning of less performing nodes, cross-validation and bootstrapping methods for the selection of training and validation data sets. Once overfitting is avoided, the model provides not only reliable predictions, but also an identification of the most effective predictors.