NMR and Bayesian regularized neural network regression for impurity determination of 4-aminophenol

A method for the determination of 4-aminophenol as an impurity in paracetamol (N-(4-hydroxyphenyl)-acetamide) by proton nuclear magnetic resonance ((1)H-NMR) spectroscopy has been developed. The (13)C-satellite from the protons in the ortho position from the hydroxyl group in paracetamol was used as an internal standard, although these peaks interfered with the peaks from the protons in 4-aminophenol. Because of interference in the spectra and non-linearity over a wide calibration range, a Bayesian regularized neural network model was used for calibration. Various kinds of data preprocessing were examined: zero filling, multiplication by a negative exponential function (line broadening), followed by Fourier transformation of the free induction decay (FID). The NMR spectral data were automatically phased and shift-adjusted by means of a genetic algorithm. Multiplicative scatter correction and data compression by wavelets and sequential zeroing of weights variable selection were performed to obtain an optimal calibration model. Neither zero filling of the FID nor line broadening improved the calibration models with regard to error of prediction, so these processes were excluded in the final model. The generated Bayesian regularized network model was evaluated with an independent test set. Four different models with different test sets were constructed to explore the quality of the calibration. The mean error of the optimal calibration model was 25.3 x 10(-6) weight of 4-aminophenol per weight paracetamol. The method is characterized by being relative fast, simple and sufficient sensitive for typical pharmaceutical impurity determinations.