Non-contact breathing abnormality detection using machine learning

Respiratory diseases are widely seen in the world and they are not seriously handled until they start affecting the patient's life very badly. Respiration motion contains information about the patient's health status which can be measured with non-contact measurement techniques. Non-contact continuous measurement of respiration rate and pattern is desirable for both the patients and the caregivers. Doppler radar can measure the chest wall displacement, accurately. It is also cheap and accessible. Once the chest wall motion is captured, machine learning algorithms can predict the type of the breathing pattern. Different types of breathing patterns contain distinctive features that the classification algorithms can focus on. In this study, a Doppler radar measurement setup was prepared. The accuracy of the system was tested with a linear actuator and it found to be accurate enough to measure the chest wall displacement. 5 breathing patterns including normal, hypoventilation, Kussmaul, Cheyne-Stokes and Biot's breathing were collected from 10 subjects. Since each subject reproduced 5 breathing patterns, a total of 50 measurements were taken. Results show that prediction accuracy is 96% for linear discriminant and subspace ensemble classifier, and other used algorithms also predict the patterns with more than 90% accuracy.