Multiresolution similarity search in time series data: an application to EEG signals

Time series constitute a prevalent data type that arise in several diverse disciplines (e.g., biomedical data, sensor data, images, video data), and therefore analyzing time series is a significant task with a plethora of important applications. In this paper, we study the general problem of similarity search in time series databases and we propose a novel multiresolution indexing (i.e., representation) and retrieval method for time series similarity search. Our approach is motivated by the idea that if we examine a time series at different resolution levels, we could possibly acquire further insights about the data. The proposed algorithm adopts a combined, two-step pruning (filtering) strategy to further reduce data dimensionality by discarding irrelevant time series (i.e., false alarms). At a first level, the time series are represented by line segments and filtered by the triangular inequality property. Then, a Vector Quantization like scheme is applied to encode data and thus to reduce dimensionality. We test and demonstrate the performance of the proposed method, analyzing EEG time series data for retrieval of one of the constituent brain waveforms in EEG recordings, the K-complex, but the method can as well be applied for retrieval of other patterns of interest in time series analysis. The automatic detection and categorization of the EEG patterns will allow the advanced correlation analysis of large amounts of data and will lead to advanced decision making capabilities assisting diagnosis by medical professionals.