Design and development of an intelligent algorithm for optimizing biomedical data acquisition on miniaturized devices

This work contributes to the development of robust and interpretable diagnostic models, and supports the future integration of real-time, low-cost breath analysis into clinical workflows. Helicobacter pylori is a spiral-shaped bacterium that colonizes the human stomach and causes chronic inflammation of the gastric mucosa. Its urease activity leads to ulcer formation, which is a known risk factor for gastric cancer. The ¹³C-Urea Breath Test (UBT) detects this infection by measuring ¹³CO₂ in exhaled breath after ingestion of labeled urea. The diagnostic decision traditionally relies on the Delta Over Baseline (DOB) metric, defined as the difference in ¹³C values before and after the ingestion. However, it will be demonstrated that the DOB measure is sensitive to instrument drift, ion current variability and acquisition modes which can all affect precision and diagnostic reliability particularly around the decision boundary. This study, perfomed in collaboration with NanoTech Analysis S.r.l. and the Department of Medical and Surgical Sciences, University of Bologna introduces a methodology for improving the reliability, accuracy, and interpretability of UBT using miniaturized isotope-ratio-mass-spectrometry. This methodology was developed and tested on data measured on two different prototypes: one used in preliminary testing, and a compact and automated system designed for clinical use. The work begins by comparing data partitioning strategies, training classifiers on both raw and corrected DOB values. Model performance is evaluated using fixed-patient and sliding-window splits, selecting the most reliable configuration for assessing generalization. Variability in the r45 isotopic ratio, defined as the ratio between the ion intensities of the signal measured at mass 45 and mass 44, is addressed through a correction method called TIC3. This approach combines statistical modeling techniques with machine learning-based adjustments to reduce both systematic and residual fluctuations in r45. The classification threshold is then optimized using a performance-based criterion, resulting in more stable and reliable DOB estimates. To account for acquisition-related variability, the method integrates the acceleration voltage (HV) applied to the detector. Patients are grouped by HV intervals, and correction factors are computed from regression models fitted on negative samples within each group. This HV-aware adjustment enhances consistency across different acquisition sessions. An alternative classification strategy is also developed, avoiding reliance on composite metrics such as DOB. The model extracts spectral features directly from the signal, including r45 values before and after ingestion, the absolute intensity of the CO₂ peak at mass 44 (p44), its variation between pre- and post-ingestion (Δp44), and intra-patient variability (std_p44). A data-driven model trained on these features yields recall and AUC (Area Under the ROC Curve) comparable to those obtained with DOB-based models. Dimensionality reduction techniques, including projection and visualization methods, support the analysis of feature space separability and visualization of normalization and correction effects. Although not yet ready for clinical application, these tools reveal meaningful patterns in instrument behavior and spectral structure, offering a solid foundation for further refinement of breath-based diagnostic models.