A literature review analysed correlations between breath VOCs and selected diseases. Spectral simulations identified absorption lines suitable for laser-based detection while minimizing cross-sensitivities from gases such as water and carbon dioxide. System architecture and specifications were defined. To reduce integration risks, established technologies were selected for the clinical prototype, while innovative concepts continue to be developed in an exploratory prototype targeting a more compact, powerful breath analyser.
Clinical settings and study protocols were established. The study includes baseline and validation phases using GC-MS as reference method. Patient recruitment and breath sample analysis are ongoing. The resulting data support AI model development and validation of disease-associated biomarkers identified during the literature review.
WP3 focuses on developing essential components for the final prototype. High-heat load DFB-QCL lasers at targeted wavelengths have been designed, with a five-wavelength fiber-packaged DFB array in progress. The DFB array has been simulated and designed, and fabrication is underway, with identified risks addressed. The multipass cell that is the core optical component of the VOCORDER spectrometer was manufactured according to the proposed design and thoroughly characterized. The following characteristics were successfully assessed:
• optical path length of 57 m,
• sufficient reflectivity of the optical-grade gold coating for 300 reflections,
• interference fringe noise below 5 × 10-4,
• vacuum integrity down to 0.01 atm.
A beam combining system was simulated, designed, implemented and tested to feed the multipass cell. A first complete version of the electronic subsystems, including current sources, temperature controllers, central control unit, detector, data acquisition and user interface, was developed together with the corresponding software. For the self-mixing scheme, a theoretical model guided the experimental setup and delivered promising results.
WP4 focuses on developing AI models for disease detection using breath and EMR data. EMR integration was achieved using HL7/FHIR standards, ensuring anonymised and regulation-compliant data transfer. Statistical analysis of six targeted VOCs (acetone, isoprene, ammonia, ethanol, ethylene, and methanol) across 183 subjects confirmed highly significant inter-group differences (ANOVA p<0.001). Dimensionality reduction techniques demonstrated clear separation between disease groups. A machine-learning framework with nested 5-fold cross-validation achieved Milestone MS4, with all five model architectures exceeding 95% classification accuracy for Kidney Insufficiency and Breast Cancer. SHAP-based explainability identified disease-specific VOC signatures consistent with known pathophysiological mechanisms.
To ensure predictions reflect disease biology rather than demographic or lifestyle effects, confounding validation was applied to both analytical tracks. On targeted VOC data, adversarial deconfounding and counterfactual invariance testing confirmed prediction stability across age, sex, smoking status and medication use. On the untargeted SESI-MS dataset (2,647 features), a three-method pipeline combining residualization, propensity score matching and inverse probability weighting identified robust features for Kidney Insufficiency and Breast Cancer, validated by permutation testing.
WP5 delivers the integrated VOCORDER system and its laboratory validation. An integration plan was established covering optoelectronic components, electronics, software and cloud-based AI analytics. The beam-combiner unit was provided by CAI to ARGOS for integration with the multipass cell and detector. Delays in control electronics are mitigated using commercially available electronics for module verification before full system integration.
Spectroscopic measurements on gases such as CO2 and CH4 will validate functionality and characterize system performance. The integrated prototype will subsequently be validated at EMPA using dedicated gas-mixture setups for reproducible evaluation of target VOC performance. First measurements of acetone using the fabricated multipass cell and a mid-infrared QC laser demonstrated the viability of the overall measurement concept.
