Pattern recognition in intra-breath oscillometry measurements

Intra-breath oscillometry (IBOsc) offers a high-resolution, non-invasive assessment of respiratory mechanics by tracking within-breath variations in respiratory impedance. Unlike traditional oscillometric approaches, which provide averaged impedance values over the entire respiratory cycle, IBOsc captures dynamic, nonlinear changes by analyzing single-frequency excitation signals. This study introduces a novel machine learning-based framework for automated pattern recognition in IBOsc data. The proposed pipeline incorporates artifact-tolerant preprocessing, impedance loop generation, feature engineering, and classification. Using carefully curated datasets from healthy individuals and patients with chronic obstructive pulmonary disease (COPD), interstitial lung disease (ILD), and obesity hypoventilation syndrome (OHS), the proposed framework successfully identifies clinically relevant patterns such as tidal expiratory flow limitation (tEFL). In the binary classification task distinguishing tEFL in healthy versus COPD patients, the best-performing model achieved an F1-score of 0.98 and an overall accuracy of 98.5% on a held-out test set of 204 samples. In the more complex three-class scenario involving healthy, COPD, and ILD patients, the model sustained strong performance, reaching a macro-averaged accuracy of 86.3% across 156 test samples, with a class-wise accuracy of 96.2% for tEFL detection.Beyond binary classification, the method proved effective in identifying both the presence and the resolution of tEFL patterns, which is a key clinical indicator for tracking therapeutic outcomes. The methodology demonstrated robustness across varying conditions and measurement setups, highlighting the potential of automated IBOsc analysis for enhancing clinical diagnostics and phenotyping of respiratory diseases.