Acoustic detection of excessive lung water using sub-band features

This paper proposes a fast and accurate feature extraction and classification algorithm to detect excess water in lungs using lung sounds. The proposed design uses a three-part Segmented Sub-band Feature Extractor to extract features. The first part extracts features by segmenting the frequencies found in these sounds into bins through sub-bands. The second part uses Principle Component Analysis and Support Vector Machine Recursive Feature Elimination for feature selection. In the third part, Support Machine Vector and k-Nearest Neighbor classification methods are used as classifiers and the accuracies are compared. Preliminary results obtained from the data collected show that the proposed method can achieve up to 99% accuracy. The proposed method was tested with real patient samples. Truncation of data up to ten bits was also tested with up to 95% accuracy. The algorithm was implemented on a Digital Signal Processor. The proposed method may be further developed to detect excessive water in lungs readily out of hospitals as a preliminary diagnostic device.