We are excited to announce our latest work on deep learning-based analysis of breath volatile organic compounds for the classification of glycemic states in type 1 diabetes. Using a convolutional neural network, we achieved a mean balanced accuracy of 86.6% in distinguishing fasting, postprandial, and hypoglycemic states directly from minimally preprocessed GC–IMS spectra. Importantly, the approach remained robust to retention-time shifts caused by instrumental drift without requiring spectral alignment, demonstrating the potential of deep learning for robust analysis of complex breathomics data. This work is a collaborative effort involving Cléo Nicolier, Daniel Kerber, Alceu Bissoto, Seif Ben Bader, Lisa M. Koch, Juri Künzler, Stefanie Hossmann, Martina Rothenbühler, Markus Laimer, and Lilian Witthauer.
The full article is available in Scientific Reports:
