Deep Learning Outperforms Descriptor-Based Classification of Food Items Using Chromatography–Mass Spectrometry Data

Abstract

Rationale

Food classification and adulteration detection are challenged by the complexity of chemical matrices. Traditional descriptor-based methods fail to capture interactions in chromatography–mass spectrometry (CMS) data. This study compares a convolutional neural network (CNN) against a random forest (RF) baseline using engineered descriptors to classify 15 food items to demonstrate deep learning advantage for objective, scalable food authentication.

Methods

Three thousand CMS spectra is computationally simulated (200 per class) based on FooDB compound profiles. Six engineered descriptors (mean intensity, standard deviation, spectral entropy, etc.) are extracted for RF training. A SimpleCNN with two convolutional layers (32/64 filters, kernel size 5) followed by max-pooling, a fully connected layer (128 units, dropout 0.3), and softmax output is trained (Adam, lr = 0.001, 50 epochs). t-SNE and K-means (k = 5) visualize the chemical space, while adulteration are simulated by mixing corn syrup into honey spectra (10%–50% v/v).

Results

The CNN achieves 93.1% accuracy, significantly outperforming RF (87.3%). t-SNE reveales five coherent clusters (silhouette coefficient = 0.71; Davies–Bouldin index = 0.43). Adulteration detection reaches 94.2% sensitivity and 96.8% specificity. The chemical similarity network identified novel pairings (e.g., chocolate–coffee, cosine similarity = 0.804) consistent with shared Maillard reaction products.

Conclusion

End-to-end deep learning on CMS fingerprints substantially surpasses descriptor-based classification to provide a robust pipeline for food authentication, adulteration screening, and data-driven pairing discovery. Validation on experimentally acquired spectra remains a necessary next step.