Feature extraction with time series models: application to lung sounds

A new method for the extraction of features from stationary stochastic processes has been applied to a medical detection problem. It is an illustration of new possibilities with automatic time series modeling. Firstly, the model type and the model order for time series prototype models are selected. The two prototypes represent the lung noises of a single healthy subject, before and after the application of methacholine. Using the model error ME as a measure for the difference between time series models, new observations can be divided into classes that belong to the prototype models for this person. The prototype models are obtained from a few expiration cycles under known conditions. This is sufficient to detect the presence of methacholine in new data of the same subject. It is not necessary to use the same model type and order for prototype and new data. Automatically and individually selected models for prototypes and data give a good detection of metacholine