Classification and recognition of asthmatic breathing sounds

Spectral information of breathing sounds was used to classify the sounds into two or three classes of seriousness of asthma. With both supervised and unsupervised neural networks linear separation of the spectral vectors appeared to be possible, with significant better results when classification was restricted to two instead of three classes of asthma. Classification according to the nearest neighbour classification method yielded results comparable with those of the neural networks when restricted to two classes. Performance of the nearest neighbour method was better than the neural method in case of three classes. Different representations of the spectral information were studied; best results were obtained for power spectra on a near-decibel scale. Introduction For diseases causing airways obstruction, it is not yet possible to continuously monitor the rate of obstruction. Sound observation by medical practitioners is laborious and highly subjective. Automated analysis of the breathing sound spectra would provide a valuable diagnostic tool. Up to now, sound spectra were not considered in detail but described by rough parameters as mean or median frequency [1,2]. Spectra are often obtained under artificial conditions such as air-flow standardisation or induced airways obstruction [1,2]. So far, no unambiguous relation between the degree of airways obstruction and breath sound could be established [3]. This study describes the frequency analysis of real life sound recordings of an asthmatic patient, using more detailed spectral information. Methods From one female patient 18 yr. of age one night-time registration was available for analysis. By two medical practitioners 54 samples from this registration were classified by ear into three classes: "no wheeze audible", indicating (as to our assumption) no airways obstruction, "few wheeze audible", indicating a slight degree of airways obstruction, and "much wheeze audible", which indicates a severe degree of airways obstruction. The three classes are related to three successive periods in time, with durations of 4 1/2 hours, 1 hour, and 3/4 hour respectively, containing 20 samples "no wheeze", 18 "few wheeze" and 16 "much wheeze" respectively. Classification experiments were performed on this set of three classes. Since the class "few wheeze" is a rather feeble one due to its instationary nature, experiments considering only the two remaining classes (36 data points) were performed as well. Samples were analysed by determining Fourier spectra in the spectral range of 100-1300 Hz. This range was divided into 26 intervals. Each spectrum represents the frequency contents of one full breathing cycle. Power vectors were calculated and normalised as to total power. Though often used, the choice of power (instead of e.g. amplitude) is not an obvious choice. In order to establish the optimal representation of the spectra, transforms of the Fourier spectra are investigated. The supervised networks consisted of an input layer of 26 neurons, and none or one hidden layer with up to 7 neurons. To reduce the a priori probability of correct classification two output neurons were used in all supervised networks. The three sound classes were represented by binary target vectors. In order to obtain a continuous error descent, a sigmoid transfer function was chosen, in stead of a step function. For the training of the networks the Levenberg-Marquardt optimisation algorithm was chosen, as it appeared to yield an efficient training procedure. The supervised networks were trained on all-but-one spectral vectors and tested on the excluded vector. This procedure was carried out for each spectral vector, therewith testing all vectors. This study was supported by grant 94.17 of The Netherlands Asthma Foundation.