EVALUATION OF LABORATORY DATA BY CONVENTIONAL STATISTICS AND BY 3 TYPES OF NEURAL NETWORKS

Sixty-three patients with lung (34 small-cell, 18 squamous, 11 adeno-) carcinomas and 43 patients with benign lung diseases were characteriz ed with seven tumor markers: neuron-specific enolase (NSE); cancer ant igens CA 19-9, CA 125, CA 15-3, and CA 50; carcinoembryonic antigen (C EA); and tissue polypeptide antigen. Diagnosis had been established by histological examination after surgery and used for classification. A fter vector transformation, the tumor marker data were fed into neural networks (NNs) based on three different types of learning algorithms: backpropagation (BP), competitive learning (CL), and Hopfield (H). Fo r comparison, the data were evaluated with multivariate stepwise discr iminant analysis (MVSDA). BP-NNs are equal to (NSE, CA 19-9, CEA) or b etter than (100% correct classification when using all seven markers) MVSDA in assigning the correct diagnosis to the patients. Cross-valida tion data yielded shrinkage effects ranging from 0% to 12.5%. Quality- control (QC) data were evaluated by traditional QC algorithms and comp ared with the results obtained by a BP-NN. The results show that the B P-NNs could only partly fulfill the tasks of three QC algorithms regar ding the violation of static borders but gave good results with respec t to dynamic changes.