Prediction of Caffeine in Tablets Containing Acetylsalicylic Acid, Dipyrone, and Paracetamol by Near-Infrared Spectroscopy, Raman Scattering, and Partial Least Squares Regression

Two chemometric models drawing on diffuse reflectance near-infrared spectroscopy and Raman scattering are proposed to predict caffeine content in tablets based on acetylsalicylic acid, dipyrone, and paracetamol contents. However, data mining from these analyses to create models generally requires a prior comparison between spectral data and the results from reference values obtained by analytical methodology. Therefore, the construction of a robust calibration model entails that both analytical methods are simultaneously employed on several samples, which represents a limiting factor for the widespread use of spectroscopy. In this case, grounded tablets of different brands, containing only the active principles acetylsalicylic acid, dipyrone, or paracetamol and their excipients, were doped with controlled amounts of pure caffeine ranging from 0 to 10%(w/w) and used as calibration samples. Thus, caffeine quantification with a reference method was not necessary. The prediction samples had at least one of the aforementioned active ingredients and caffeine in its original formulation. Hence, the %(w/w) values of caffeine used as reference for the prediction steps were calculated from the values described on the drug description leaflet and the tablet final mass. Partial least squares regression was used as a multivariate method to construct the models. The near-infrared and Raman prediction models for caffeine, using four latent variables, presented the respective values of 0.79 and 0.78 of root-mean-square errors of cross validation, 0.74 and 1.00 of root-mean-square errors of prediction, and 0.97 and 0.97 of correlation coefficients.

Automated summary

Two chemometric models were proposed to predict caffeine content in tablets using diffuse reflectance near-infrared spectroscopy and Raman scattering. The construction of robust calibration models requires simultaneous use of both analytical methods, which limits the widespread use of spectroscopy. The prediction models for caffeine presented high correlation coefficients and low root-mean-square errors.