Discrimination of omega-3 fatty acid oil forms by combining NMR spectroscopy with artificial intelligence

Abstract

This study presents an approach for discriminating omega-3 fatty acid forms using proton nuclear magnetic resonance (¹H-NMR) spectroscopy combined with machine learning and deep learning techniques. A total of 90 samples, comprising triglyceride, re-esterified triglyceride, and ethyl ester forms, were analyzed. Principal component analysis–linear discriminant analysis, support vector machine (SVM), artificial neural network (ANN), and one-dimensional convolutional neural network (1D CNN) models were applied using binned spectral data. In contrast, a two-dimensional convolutional neural network (2D CNN) was constructed using spectral images. To prevent overfitting and optimize model hyperparameters, early stopping, cross-validation, and Bayesian optimization were used across the different machine learning and deep learning models. The 1D and 2D CNN models both achieved 100% accuracy on the training and test sets, while the SVM and ANN models yielded slightly lower but still excellent performance, with a test accuracy of 94.4%. Model interpretability was enhanced through SHapley Additive exPlanations and Gradient-weighted Class Activation Mapping, which identified critical spectral regions associated with classification decisions. These results demonstrate that the integration of artificial intelligence techniques with ¹H-NMR spectroscopy enables accurate, interpretable discrimination of omega-3 fatty acid forms, offering a promising strategy for supplement authentication and quality control.