Application of multivariate linear discriminant analysis to lung sounds in some pulmonary diseases

In the past 15 yrs, a number of investigators have applied spectral analysis to respiratory sounds recorded from the chest wall or the trachea in order to objectively characterize them and to relate them with different pulmonary diseases. In the present study, we have applied multivariate linear discriminant analysis to the spectral features of respiratory sounds. Lung sounds and the airflow velocity were recorded from 15 normal adults and 37 patients falling into three different disease categories: chronic obstructive lung disease, bronchial asthma and bronchiectasis. All patients had prominent adventitious lung sounds (i.e. either wheezes or crackles). Amplitude spectra of five selected inspiratory and expiratory sound segments of each subject were calculated using the Fast Fourier Transform algorithm. Multi-variate linear discriminant analysis was then applied to the normalized and averaged spectral area values calculated for 10 unequal and arbitrarily selected frequency bands for each patient in the frequency range between 80 Hz and 1 kHz. Inspiratory and expiratory sounds were treated separately. Discriminant functions were computed from the spectral area values and plotted on graphs to classify the subjects into one of the disease categories or as normal (training set). While some separation was achieved among the different disease groups, a clearer separation was evident between normals and patients as a whole on the basis both of inspiratory and expiratory sounds. Inspiratory and expiratory sound frequency bands having the largest weights in classification were determined. Admittedly, the specific results of this study are preliminary or even tentative in view of the inadequacies of sound recording and signal conditioning techniques that were available to us at the time of recording. However, we believe that the investigation serves to illustrate the potential of multivariate discriminant analysis in the diagnostic classification of patients on the basis of their lung sound patterns. We suggest that this technique be considered by investigators involved in lung sound research, because it also allows other patient variables to be combined with the selected parameters of lung sounds.