Dynamic feature–based breath ammonia analysis using a SAW sensor integrated with shallow neural network classification and nonlinear regression

Breath ammonia (NH 3 ) serves as a promising non-invasive biomarker for gastrointestinal and metabolic disorders. However, accurately interpreting breath-derived NH₃ signals remains challenging due to humidity interference and nonlinear sensor responses. This study presents a surface acoustic wave (SAW)-based breath analysis platform designed for reliable NH 3 detection through dynamic signal processing and machine learning. The platform features a dual-track SAW resonator integrated with a hybrid nanocomposite film made of AuNPs, Cu₂O, rGO, and PPy, which converts NH 3 adsorption into time-resolved frequency responses at room temperature. Rather than relying solely on steady-state frequency shifts, we extract dynamic frequency-domain features—such as frequency shift, rise time, recovery time, and response slope—to characterize adsorption and desorption kinetics and mitigate the impact of humidity-induced signal variation under breath-relevant conditions. For classification of normal and abnormal breath conditions, we utilize a shallow neural network (SNN) achieving 95.0% accuracy with 40 normal and 40 abnormal samples. Additionally, a Levenberg–Marquardt (LM) nonlinear regression model estimates NH₃ concentration quantitatively with R²=0.959. Our proposed framework is tested under both dry air conditions and humid conditions (35% relative humidity), demonstrating robust performance with classification accuracy of 98.9% (dry) and 95.0% (humid), and regression R² values of 0.986 (dry) and 0.969 (humid). This work offers a comprehensive hardware-software co-design approach for non-invasive breath NH 3 detection and classification in physiologically relevant conditions.