Lung Sound Recognition Using Model-Theory Based Feature Selection and Fusion

In this paper we describe the application of a new automatic signal recognition methodology to the recognition of lung sounds. Two main features of this methodology are: the use of symbolic knowledge about the signal sources (in addition to sensory inputs), and fusion of multisensor information. We show steps involved in the design of an automatic multisensor recognition algorithm that extracts symbolic features and utilizes symbolic knowledge for recognition. We also compare the performance of the resulting algorithm with both a single-sensor system and with a system that selects features using an entropy-based criterion. To evaluate the methodology we used both normal lung sounds (bronchial and vesicular) as well as adventitious sounds (rhonchi, wheezes and crackles). Our experiments show that the recognition accuracy can be improved through the use of symbolic knowledge about the signals, and that our methodology is feasible for this type of application.