Use of an artificial neural network in estimating prevalence and assessing underdiagnosis of asthma

Abstract

An artificial neural network was trained in the recognition of asthmatics in a general practice population, employing crossvalidation on a subset of 350 patients of known asthmatic status. The trained network was then run on the data from 3139 patients whose asthmatic status was unknown. Using the values from the test set as estimates of sensitivity and specificity, the number predicted positive was adjusted to allow for false positives and false negatives to give an estimate of asthma prevalence and the minimum underdiagnosis rate that this suggested for the population. Using different data sets and network structures, prevalence rates of approximately 16–21% were measured providing evidence, even after allowing for maximum variablity in the estimates, consistent with under-diagnosis of at least a small percentage (0.7–4.0%). To provide a more precise estimate of the rate of this under-diagnosis and associated prevalence, a larger training and testing set of more accurately labelled cases is planned.