Determination of microalbuminuria and increased urine albumin excretion by immunoturbidimetric assay and neural networks

A new technique is demonstrated for the determination of urine albumin concentration. A commercially available microalbuminuria test was combined with neural network analysis of reaction kinetic data. In total 102 patient urine samples were analyzed [27 diabetes patients, 21 with nephrosis or nephritis, 54 with hypertension]. Due to the prozone-effect in standard immunoturbidimetric assay technology we found 1 sample in the diabetes group, 6 in the nephrosis/nephritis group and 4 in the hypertension group, that yielded false negative results, i.e. misleading low instead of high urine protein concentrations. By means of albumin dilution series in a range of 0 to 40,000 mg/l a non-monotonous calibration curve (Heidelberger curve) was obtained. The measured kinetic data were split for training, testing and validation of a backpropagation neural net. It could be demostrated that such a net yields a correct correlation between measured signals and concentration, even for the difficult task of classification between very high and very low concentrations. Moreover also the false negatively assigned patient data were all classified correctly.