Signal processing methods for non-invasive respiration monitoring
This thesis investigates the feasibility of using a set of non-invasive biomedical signals to monitor respiration. The signals of interest being the electrocardiogram (EGG), photoplethysmography (PPG) and impedance plethysmography (IP) signals. The work has two main aims; the first being to estimate breathing rates from the signals, the second being to detect apnoeas from the signals. The fusion of information from different signals is used throughout in developing algorithms that give more accurate respiratory information than that obtained using one signal alone. Respiratory waveforms are derived from the signals, and the accuracy of detecting individual breaths from the waveforms is assessed and compared objectively. Results from evaluations on two separate databases show there is no waveform that gives sufficient accuracy to consider using it alone. A novel fusion method is developed which uses measurements from all three signals. This fusion method is based on weighting the estimates from each signal, according to the innovation from a Kalman filter model, applied to each respiratory waveform separately. The fused estimates give a higher overall correlation with respect to the reference breathing rate values than any of the breathing estimates derived from a single waveform. The detection of both central and obstructive sleep apnoea from the signals is investigated. It is shown that the accuracy of detecting central apnoeas from the IP signal using a timedomain method, often used in practice, can be improved by combining it with information from the frequency-domain. When discriminating between obstructive sleep apnoeic and non-apnoeic data it is seen that combining features from two signals results in a superior classification accuracy than is possible by using features from just one signal. The proposed classification system using just one of these signals, the EGG, is shown to give a performance accuracy comparable to that found in the literature. In conclusion this thesis shows that by fusing information from a number of non-invasive biomedical signals, estimations of breathing rates can be found with correlation 0.8. This is superior to estimation using only the impedance pneumography signal (correlation 0.64) which is currently used to monitor respiration. The fusion approach could potentially be applied to improve other non-invasive physiological monitoring systems.