A computational study transformed 121 breath samples from asthma, bronchiectasis and COPD through simulated mid-infrared, plasmonic and hybrid sensing. Accuracy rose from about 0.5 in the original features to above 0.96 in repeated cross-validation and remained robust to several perturbations. This is not yet a physical diagnostic device or prospective clinical test.
Key findings
- Original features averaged about 0.5 accuracy, while transformed representations exceeded roughly 0.96; linear classifiers performed best. Mid-IR sensing preserved geometry at Spearman about 0.89, plasmonic nonlinearity increased local separation, and humidity remained the dominant confounder despite robustness to noise, drift and fabrication variation.
Why this matters globally
Co-designing sensing physics with AI may simplify downstream models by making disease signals more separable in hardware, if physical fabrication and external validation succeed.
Thai researcher contribution
Vishal Chaudhary, affiliated with Mahidol University's Centre for Theoretical Physics and Natural Philosophy and Chitkara University, leads this Thai-Indian computational-sensing work.
Limitations to consider
A 121-sample three-disease dataset is small and vulnerable to overfitting or leakage if preprocessing is outside folds. Labels, controls and external cohorts require scrutiny. Simulated perturbations do not replace a fabricated prototype, and humidity may substantially degrade performance.