Design of an artificial neural network research framework to enhance the development of clinical prediction models

Abstract

This thesis presents an Artificial Neural Network Research Framework (ANN RFW) for predicting medical outcomes. The ANN RFW along with other new and pre-existing applications, and the steps linking them are presented as part of an Outcome Prediction Model Definition Process (OPMDP). Proof-of-concept experiments are performed on three outcomes from two Canadian Neonatal Network (CNN) databases. Successful results were obtained from the ANN RFW and a number of the subsequent applications. Results obtained using an ANN plus case based reasoner (CBR) were not yet favourable. In one of the intermediary steps, a modified method for extracting relative importance of ANN inputs was developed. The resulting relative weight results indicated that the importance of input variables of continuous outcomes may vary over the course of outcome's duration. Observing relative weights for three outcomes indicated that each outcome must have its own prediction model.