Overcoming Cross-Sensitivity for the Accurate Identification of Acetone, Isopropanol, and Clinical Mixtures Using a MEMS Dual-Sensor Array and Multi-task Deep Learning

This is the authors' abstract. We don't add key points for this paper.

Accurate detection of exhaled acetone (ACE) and isopropanol (IPA) mixtures is critical for the non-invasive screening of diabetic ketoacidosis and the continuous monitoring of lipid metabolism in type 1 diabetes mellitus. However, the broad clinical application of this approach remains severely constrained by the inherent cross-sensitivity of metal oxide semiconductor (MOS) sensors. To address this, a microelectromechanical systems (MEMS) dual-sensor array integrating PdO-modified SnO2/ZnO and MoS2/ZIF-67 nanocomposites was developed for simultaneous gas identification and concentration regression. By pioneering the extraction of distinct transient thermokinetic responses of the materials, combined with a multi-task learning architecture featuring a 1D Convolutional Neural Network and a Bidirectional Long Short-Term Memory (CNN-BiLSTM) network, the system efficiently analyzed dynamic sensing data using a sliding time window (K). At an optimal 15-s window, the model demonstrated robust real-time predictive capabilities, achieving 95.4% classification accuracy and a regression mean absolute error (MAE) of 0.625 ppm across clinical IPA/ACE ratios. This deep learning-enabled framework circumvents traditional steady-state limitations, providing a scalable, low-power solution for precise clinical breathomics and metabolic disease screening.