AIMS: To advise a system of neuronal networks which can classify the densitometric patterns of serum electrophoresis. METHODS: Digitised data containing 83 normal and 132 pathological serum protein electrophoresis patterns were presented to four neuronal networks containing 1900 neurons. Network 1 evaluates the integrated values of the albumin, alpha 1, alpha 2, beta and gamma fractions together with total protein (Biuret method). Networks 2, 3, and 4 analyse the shape of the albumin, beta and gamma fractions. To increase the sensitivity for the detection of monoclonal gammopathies a Fourier transformation was applied to the beta and gamma fractions. RESULTS: After a learning period of 20 minutes (back-propagation learning algorithm) the system was tested with a set of electrophoresis patterns comprising 446 routinely collected samples. It differentiated between physiological and pathological curves with a sensitivity of 97.5% and a specificity of 98.8%, with 86% correct diagnoses. All monoclonal gammopathies were recognised by the Fourier detector. CONCLUSIONS: Neuronal networks could be useful for certain medical uses. Unlike rule based systems, neuronal networks do not have to be programmed but have the capacity to "learn" quickly.
AIMS: To advise a system of neuronal networks which can classify the densitometric patterns of serum electrophoresis. METHODS: Digitised data containing 83 normal and 132 pathological serum protein electrophoresis patterns were presented to four neuronal networks containing 1900 neurons. Network 1 evaluates the integrated values of the albumin, alpha 1, alpha 2, beta and gamma fractions together with total protein (Biuret method). Networks 2, 3, and 4 analyse the shape of the albumin, beta and gamma fractions. To increase the sensitivity for the detection of monoclonal gammopathies a Fourier transformation was applied to the beta and gamma fractions. RESULTS: After a learning period of 20 minutes (back-propagation learning algorithm) the system was tested with a set of electrophoresis patterns comprising 446 routinely collected samples. It differentiated between physiological and pathological curves with a sensitivity of 97.5% and a specificity of 98.8%, with 86% correct diagnoses. All monoclonal gammopathies were recognised by the Fourier detector. CONCLUSIONS: Neuronal networks could be useful for certain medical uses. Unlike rule based systems, neuronal networks do not have to be programmed but have the capacity to "learn" quickly.