Physics-guided transformation of breathomic feature spaces into disease-specific representations for respiratory disease classification

Exhaled breath is a complex mixture of volatile organic compounds (VOCs) that reflect underlying physiological states, yet significant overlaps between conditions limit the detection of disease-specific signatures. Here, we introduce a physics-guided computational sensing framework that transforms breathomic feature spaces into representations that are diagnostically separable. Using a dataset of 121 samples related to asthma, bronchiectasis, and chronic obstructive pulmonary disease, we develop a physics-guided computational model of photonic sensing based on mid-infrared spectral selectivity, plasmonic nonlinear response, and their hybrid combinations under real-time perturbations, including noise, drift, fabrication dissimilarities, and humidity interference. While classification in the original breathomic representation achieved average accuracy ∼0.5, sensor-transformed representations achieved accuracies exceeding ∼0.96 in repeated cross-validation. Linear classifiers worked optimally, indicating that photonic sensing induces near-linear separability of disease-specific signatures. Mid-infrared sensing preserves the intrinsic geometric structure of the breathome (Spearman ≈ 0.89), whereas nonlinear plasmonic responses introduce controlled geometric distortions, thereby enhancing local differentiation. Besides, hybrid architectures exhibit continuous tuning across these regimes, revealing a fundamental trade-off between structure preservation and classification, governed by sensing physics. Robustness analysis exhibited invariance to noise, drift, and fabrication variability, with humidity identified as the dominant confounding factor. Analysis of disease-discriminative VOCs and interaction-matrix sensitivity revealed that the transformed representations preserve biologically significant signatures while remaining robust across different sensor-VOC interactions. These findings demonstrate the potential of physics-guided computational sensing as a geometry-transforming framework for biomedical data representation and suggest a new strategy for integrating sensing physics with machine intelligence to improve breathomics-based disease diagnostics.