ProstAsure Index — A Serum-Based Neural Network-Derived Composite Index for Early Detection of Prostate Cancer

Abstract

ProstAsure is a neural network-derived algorithm which analyses the profile of multiple serum tumour markers and produces a single-valued diagnostic index (ProstAsure Index, or PI) for early detection of prostate cancer (CaP) in men with a relatively low level of serum prostate-specific antigen (PSA). PI has been validated through multiple retrospective clinical studies with a fairly large number of blind independent test patients and become the first of such tests to be commercially available through reference laboratories as a clinical information processing service. In this chapter, we first give a brief introduction to the clinical background of prostate cancer and then describe the derivation of the PI algorithm. Results from several clinical studies comparing PI with the PSA assay alone or the free to total (f/t) PSA ratio are presented. We will then focus the discussion on our experience in dealing with issues that are unique in the development of a clinical diagnostic system for the purpose of commercial deployment. The first issue is the construction of a training dataset. Computational learning theories assume independently and identically distributed (i.i.d.) sampling. In clinical reality, due to limitations of the current diagnostic techniques, the most commonly identifiable cases for a particular disease may not necessarily be the most informative ones to help define the decision boundaries of a classification system. When the total number of patients available for training are limited by patient source and cost of data collection, it makes sense to incorporate known medical knowledge to construct an ‘information-enriched’ training set to improve learning efficiency.