SUPERIORITY OF NEURAL NETWORKS OVER DISCRIMINANT FUNCTIONS FOR THALASSEMIA MINOR SCREENING OF RED-BLOOD-CELL MICROCYTOSIS

We compared the utility of screening red blood cell (RBC) microcytosis for thalassemia minor using backpropagation neural networks, linear a nd quadratic discriminant functions, and previously reported discrimin ant functions based on RBC indices. Screening classification of cases representing possible thalassemia minor (n = 60) and nonthalassemic mi crocytosis (n = 60) were studied. Among eight RBC indices evaluated, t he RBC count was the best univariate discriminant function. Multivaria te stepwise discriminant analysis selected the RBC count, the mean cor puscular volume, and the percentage of hypochromic cells as the most d iscriminatory subset of RBC indices. Optimized linear and quadratic di scriminant functions based on these indices performed better than seve n previously reported multivariate discriminant functions. However, op timized neural networks were superior to all other discriminant method s studied, averaging 94.1% discriminant efficiency, 94.2% sensitivity, and 94.2% specificity.