Artificial neural network predictions of urinary calculus compositions analyzed with infrared spectroscopy.

Infrared (IR) spectroscopy is used to analyze urinary calculus (renal stone) constituents. However, interpretation of IR spectra for quantifying urinary calculus constituents in mixtures is difficult, requiring expert knowledge by trained technicians. In our laboratory IR spectra of unknown calculi are compared with references spectra in a computerized library search of 235 reference spectra from various mixtures of constituents in different proportions, followed by visual interpretation of band intensities for more precise semiquantitative determination of the composition. To minimize the need for this last step, we tested artificial neural network models for detecting the most frequently occurring compositions of urinary calculi. Using constrained mixture designs, we prepared various samples containing ammonium hydrogen urate, brushite, carbonate apatite, cystine, struvite, uric acid, weddellite, and whewellite for use as a training set. We assayed known artificial mixtures as well as selected patients' samples from which the semiquantitative compositions were determined by computerized library search followed by visual interpretation. Neural network analysis was more accurate than the library search and required less expert knowledge because careful visual inspection of the band intensities could be omitted. We conclude that neural networks are promising tools for routine quantification of urinary calculus compositions and for other related types of analyses in the clinical laboratory.