CENSORED LIFETIME DATA IN ADAPTIVE NEURAL NETWORKS

Background and aim of the study: In clinical research, survival, relia bility and failure analyses, the use of censored lifetime data often b ecomes a necessity. In this paper we present a novel methodology devel oper to allow for the use of censored data to train neural networks to predict the time of specific adverse events. Methods and results: Spe cifically, for patients with implanted bioprostheses, we were able to design and train a neural system to successfully predict the time from valve implant to valve dysfunction, Further, rue were able to demonst rate the clear improvement in performance and predictive accuracy of t he system when trained using this method. The assertion that censored data carry additional and extremely valuable information, especially i n cases of rare events, is substantiated by this correlation analysis. Conclusions: This new methodology, in combination with results obtain ed from previous models which were able to identify the patients most likely to experience such events, now completes the picture by pinpoin ting the 'who', as well as the 'when'.