Benchtop volatilomics and advanced convolutional neural network workflows for accurate and explainable food...

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Spectroscopy

Background:

• Gas chromatography-ion mobility spectrometry (GC-IMS) is a sensitive technique for volatile organic compound (VOC) analysis.
• GC-IMS data presents challenges in peak alignment, feature extraction, and classification due to its second-order nature.
• Existing methods struggle with the complexity of GC-IMS data, limiting its practical application.

Purpose of the Study:

• To develop a robust and explainable deep learning (DL) framework for GC-IMS data analysis.
• To improve the classification accuracy of GC-IMS data using convolutional neural networks (CNNs).
• To address data limitations and enhance the interpretability of GC-IMS analysis.

Main Methods:

• A deep learning workflow utilizing convolutional neural networks (CNNs) was developed, treating GC-IMS chromatograms as images.
• Two CNN architectures (simplified and complex) were evaluated on olive oil and citrus datasets.
• A novel GC-IMS-specific data augmentation technique was introduced to simulate instrumental drift.
• Saliency maps and principal component analysis (PCA) were employed for data interpretability.

Main Results:

• The simplified CNN model achieved high classification accuracies of 96.4% for olive oil and 98.3% for citrus datasets.
• The developed data augmentation method effectively addressed challenges posed by limited sample sizes.
• Interpretability methods revealed stable, collective VOC signatures per class, moving beyond single-sample explanations.

Conclusions:

• The proposed DL framework offers a powerful and explainable approach for GC-IMS data analysis.
• Combining DL prediction with chemometric interpretability enhances the robustness of VOC analysis.
• This workflow facilitates practical laboratory deployment for complex GC-IMS datasets.