Abnormality detection in noisy biosignals.

Although great strides have been achieved in computer-aided diagnosis (CAD) research, a major remaining problem is the ability to perform well under the presence of significant noise. In this work, we propose a mechanism to find instances of potential interest in time series for further analysis. Adaptive Kalman filters are employed in parallel among different feature axes. Lung sounds recorded in noisy conditions are used as an example application, with spectro-temporal feature extraction to capture the complex variabilities in sound. We demonstrate that both disease indicators and distortion events can be detected, reducing long time series signals into a sparse set of relevant events.