Alignment-free convolutional neural network classification of glycemic states in type 1 diabetes from GC–IMS breath spectra

Abstract

Abstract Reliable detection of glycemic dysregulation, particularly hypoglycemia, is essential for people with diabetes, yet current approaches rely on capillary blood sampling or require subcutaneous sensor insertion. Breath analysis offers a non-invasive alternative by capturing volatile organic compounds associated with metabolic changes. This study investigated whether a convolutional neural network can classify glycemic states from gas chromatography–ion mobility spectrometry data of exhaled breath. Breath samples from ten individuals with type 1 diabetes were collected under standardized fasting, postprandial, and insulin-induced hypoglycemic conditions. In a proof-of-concept cohort of ten participants, the model trained on minimally preprocessed spectra achieved a mean balanced accuracy of 86.6% (±4.0%) in leave-2-patients-out cross-validation, with specificity exceeding 92% across all classes. Importantly, the network maintained stable performance despite retention time variability caused by instrumental drift, without requiring spectral alignment. In contrast, traditional machine learning models required extensive preprocessing and showed reduced robustness under shifted conditions. Saliency map visualization highlighted physiologically relevant breath markers, including acetone, isoprene, and 2-propanol. To our knowledge, this is the first application of a convolutional neural network to ion mobility spectrometry breath data. The findings establish a methodological foundation for applying neural network-based analysis to complex breathomics datasets and motivate further validation toward scalable digital health tools that complement existing glucose monitoring technologies.