Exhaled breath contains a complex mixture of volatile organic compounds (VOCs) that carry information about underlying physiology and disease. The study presents a physics-guided computational sensing framework designed to map raw breathomic feature spaces into representations that are more diagnostically separable. Rather than relying solely on downstream machine-learning transformations, the approach explicitly models sensing physics (spectral selectivity and nonlinear sensor responses) to engineer data representations that emphasize disease-discriminative structure.
The evaluation used a dataset of 121 breath samples associated with three respiratory conditions: asthma, bronchiectasis, and chronic obstructive pulmonary disease (COPD). The goal was to determine whether physics-informed sensor transformations could overcome the substantial overlap in raw breathomic signals and yield representations suitable for reliable multi-class disease classification.
The computational model simulated several photonic sensing paradigms relevant to breathomics. These included mid-infrared spectral selectivity, nonlinear plasmonic responses, and hybrid combinations of these sensing modalities. Each modality was represented in a manner that reflected the physical principles governing spectral discrimination, resonance-based nonlinearities, and potential hybridization effects when modalities are combined.
To mimic practical deployment conditions, the framework introduced real-time perturbations into the simulated sensing process. Perturbations included measurement noise, temporal drift, fabrication dissimilarities across sensor instances, and humidity interference. These factors were explicitly modeled to assess the stability and robustness of the transformed representations under conditions that commonly challenge breath-based sensing systems.
Comparison between the original breathomic representation and the sensor-transformed representations showed a marked difference in classification performance. In the native feature space, average classification accuracy was approximately 0.5. After applying the physics-guided sensor transformations, repeated cross-validation yielded accuracies exceeding approximately 0.96. The authors report that simple linear classifiers performed optimally on the transformed data, which indicates that the physics-guided sensing induced a near-linear separability among disease-specific signatures.
The study examined how different sensing physics affected the geometric structure of the breathomic data. Mid-infrared sensing was associated with strong preservation of the intrinsic geometry of the breathome, quantified by a Spearman correlation of about 0.89 between original and transformed pairwise relationships. In contrast, nonlinear plasmonic responses introduced controlled geometric distortions that enhanced local differentiation between disease classes. Hybrid architectures interpolated continuously between these behaviors, revealing a tunable trade-off between conserving the original data geometry and maximizing class separability through deliberate distortion.
Robustness testing explored the impact of modeled perturbations on classification and representation stability. The transformed representations demonstrated invariance to several perturbations, including additive noise, temporal drift, and fabrication variability across sensor instances. Among the perturbations considered, humidity emerged as the dominant confounding factor that most affected performance, identifying environmental moisture as a key consideration for real-world breathomic sensing systems.
Beyond classification metrics, the authors evaluated whether the sensor-induced transformations preserved biologically meaningful signals. Analysis of disease-discriminative VOCs and sensitivity of interaction matrices suggested that the transformed representations maintained biologically significant signatures while remaining robust across different sensor–VOC interaction scenarios. This indicates that the physics-guided transformations did not simply create separable mathematical structures at the expense of losing relevant biochemical information.
The results demonstrate the potential of integrating sensing physics into the data-representation pipeline for biomedical diagnostics. By treating sensors as active contributors to data geometry rather than passive data collectors, the approach provides a geometry-transforming pathway to improve disease classification from breathomics. The findings suggest a strategy for co-designing sensor hardware and machine-intelligence models to achieve robust, interpretable, and high-performing breath-based diagnostics for respiratory diseases such as asthma, bronchiectasis, and COPD.
Note: Details such as exact model architectures, hyperparameters, per-class performance metrics, and full experimental protocols were not reported in the abstract and would require consultation of the full text for complete reproducibility and technical specifics.