FTIR Spectroscopy Combined With Chemometric and Artificial Neural Network Modeling for Multicomponent Carbohydrate Quantification in Dairy-Derived Powders.

Abstract

Carbohydrate assessment is important in dairy production for process control, product characterization, and labeling. This study presents an FTIR-based chemometric framework integrating FTIR spectroscopy, partial least squares (PLS2) regression, and artificial neural network (ANN) modeling for the simultaneous quantification of glucose, galactose, and lactose in powdered dairy matrices. A structured spectral dataset was generated from pure carbohydrate standards, multicomponent mixtures, and dairy-derived samples, including skim milk powder and α-serum. Spectral preprocessing included Savitzky-Golay filtering, second-derivative transformation, standard normal variate normalization, and selected windows in the carbohydrate fingerprint region. PLS2 served as a linear chemometric benchmark, while a feedforward ANN modeled multicomponent spectral-concentration relationships. In the frozen group-based held-out set, the ANN achieved R2 values of 0.64, 0.89, and 0.77 for glucose, galactose, and lactose, respectively, with corresponding RMSE values of 16.82, 8.81, and 13.16 percentage points. These estimates came from 12 independent experimental groups and should therefore be interpreted cautiously when generalizing. Spectral variable-importance analysis showed that prediction incorporated characteristic carbohydrate-associated signals, together with distributed spectral information across the selected fingerprint regions. Evaluation of independent samples and reference materials showed matrix-dependent performance. Applicability-domain analysis distinguished extrapolative predictions from samples represented within the model-development spectral space, while showing that domain membership alone did not guarantee quantitative accuracy. Overall, the framework demonstrates the potential and current limitations of FTIR-based multicomponent prediction and is intended for rapid preliminary screening of carbohydrate composition, with confirmatory analysis using appropriate reference methods when required. PRACTICAL APPLICATIONS: This study proposes an FTIR-based chemometric approach for rapid screening of carbohydrate composition in dairy-derived powders. However, the current calibration does not provide sufficient accuracy for practical quantitative glucose measurement. Thus, the results should be considered as introductory screening data rather than precise quantitative values. The approach may complement conventional analytical methods by providing an initial assessment of carbohydrate profiles, while samples requiring confirmatory quantification can be evaluated using established reference methods. This application may support more efficient preliminary assessment of dairy-derived powders while reserving reference analytical methods for confirmatory analysis.

Abstract as supplied by the indexing source. © the authors / publisher.