Abstract:
The aim of this study is estimate the respiratory airflow and phases from the respiratory sounds recorded at the chest wall. In order to estimate the absolute airflow curve, time varying autoregressive (TVAR) model coefficients are used. TVAR coeffi cients are calculated with three approaches: windowing based autoregressive modeling, TVAR modeling with basis functions, TVAR modeling with Kalman filter. Then evolu tion in magnitudes of spectral band is used as an estimation of airflow curve. A Wiener filter approach is presented for fusion of different features to estimate the airflow curve. Average of correlation coefficients up to 0.75 for absolute airflow and 0.72 for airflow are achieved. In the second part of this thesis, respiratory phases are estimated using a neural network and the estimated absolute airflow curve. TVAR coefficients, Shan non entropy estimate, percentile frequencies, variance, spectral magnitude and kurtosis are used as inputs. Distributions of these features for different phases and Kullback Leibler divergence of these distributions are presented. For phase estimation from es timated airflow, heuristic methods are used for local minima extraction and selection of the transition points. 97 and 83 milliseconds (3% and 2.6% of average full cycle) of average deviation from true transition point are achieved with neural networks for inspiration to expiration and expiration to inspiration transitions respectively. 120 and 131 milliseconds (3.8% and 4.1% of average full cycle) of average deviation from true transition point are achieved with heuristic methods for inspiration to expiration and expiration to inspiration transitions, respectively.
