A TIME-SERIES PRE-PROCESSING METHODOLOGY FOR BIOSIGNAL CLASSIFICATION USING STATISTICAL FEATURE EXTRACTION

摘要

Biosignal classification is an important diagnosis tool inbiomedical application that helps medical experts toautomatically classify whether a sample of biosignalunder test/monitor belongs to the normal type orotherwise. Most biosignals are stochastic and nonstationaryin nature, that means their values are timedependentand their statistics vary over different points oftime. However, most classification algorithms in datamining are designed to work with data that possessmultiple attributes in order to capture the non-linearrelationships between the values of the attributes to thepredicted target class. Therefore it has been a crucialresearch topic for transforming univariate time-series tomultivariate dataset in order to fit into classificationalgorithms. For this, we propose a pre-processingmethodology, called Statistical Feature Extraction (SFX).Using the SFX we can faithfully remodel statisticalcharacteristics of the time-series via a sequence ofpiecewise transform functions. The new methodology istested through simulation experiments over threerepresentative types of biosignals, namely EEG, ECG andEMG. The experiments yield encouraging resultssupporting the fact that SFX indeed produces betterperformance in biosignal classification than traditionalanalyses techniques like Wavelets and LPC-CC.