Artificial intelligence for quality control of oscillometry measures.

Area of Science:

• Respiratory Medicine
• Medical Technology
• Machine Learning in Healthcare

Background:

• Forced oscillation technique (FOT) offers non-invasive lung function testing during quiet breathing.
• Manual exclusion of breathing artefacts (swallowing, glottis closure, coughing) is operator-dependent and time-consuming.
• Automating artefact detection in FOT is needed for wider clinical application.

Purpose of the Study:

• To evaluate supervised machine learning methods for automatic exclusion of breathing artefacts in FOT data.
• To assess the reliability and performance of machine learning models in FOT analysis.

Main Methods:

• Collected 932 FOT measurements from 155 patients according to European Respiratory Society standards.
• Extracted 71 features per breath, including anthropometric, pressure, breathing pattern, and oscillometry data.
• Employed feature selection methods (univariate, multivariate, wrapper) combined with classification models (e.g., AdaBoost tree).

Main Results:

• A wrapper feature selection method with an AdaBoost tree model achieved high performance on the test set (Balanced Accuracy=85%, Sensitivity=79%, Specificity=91%, AUC-ROC=0.93).
• Manual and automatic breath selection resulted in minimal differences in FOT parameters (<0.25 cmH2O*s/L for 95% of cases).

Conclusions:

• Supervised machine learning enables reliable artefact detection in FOT diagnostic tests.
• Automating artefact detection is crucial for home monitoring, telemedicine, and point-of-care FOT applications.
• This automation opens new possibilities for respiratory and community medicine.