TinyML-oriented workflow for VOC classification: proof-of-concept in breath acetone analysis

Описание: Breath acetone has emerged as a reliable, non-invasive biomarker for metabolic health monitoring and early diabetes screening. However, deploying machine learning models for breath analysis on embedded platforms remains challenging due to memory and computational constraints. In this study, we leveraged publicly available data from metal-oxide semiconductor gas sensors (UCI Gas Sensor Array Drift) exposed to acetone, ethanol, and ammonia at 100 and 200~ppm. Signals were transformed into 224 X 224 RGB scalograms using the continuous wavelet transform, enabling image-based classification. To assess embedded deployment, the model was quantized, exported to TensorFlow Lite, and validated using Edge Impulse, emulating TinyML inference on microcontrollers. Performance remained high after quantization (Accuracy=93.3%, Precision/Recall/F1=97%), demonstrating robustness for real-time embedded classification. This work introduces a compact CWT–scalogram and lightweight CNN pipeline for low-power classification of volatile organic compounds, including breath acetone. The results validate the feasibility of TinyML-based breath sensing, supporting future integration with nanostructured acetone sensors and experimental breath datasets to enable portable, energy-efficient health monitoring solutions. This work was developed by Xenia Azareth Ayón-Gómez, as part of her posgraduate studies.