Application of Ica for Automatic Noise and Interference Cancellation in Multisensory Biomedical Signals

Independent Component Analysis ICA and related meth ods like Adaptive Factor Analysis AFA are promising novel approaches for elimination of artifacts and noise from bio medical signals especially EEG MEG data However most of the methods require manual detection and classi cation of interference components Main objective of this paper is to detect and eliminate noise and some artifacts automat ically by computer using criteria for classi cation order ing and detection of noisy and random signals The auto matic detection and on line elimination of noise and other interferences is especially important for long recordings e g EEG MEG recording during sleep In this paper we fo cus mainly on the problem of cleaning or enhancement of noisy EEG MEG data from noise and undesired interfer ences using several techniques ICA and HOS measure of Gaussianity to detect and eliminate Gaussian noise linear predictor to detect i i d sources and classify temporally structured sources and Hurst exponent to detect random ness in independent components and classify independent signals Preliminary extensive computer simulation con rmed potential usefulness of proposed methods for wide class of applications especially in area of analysis and pro cessing of EEG MEG data INTRODUCTION AND PROBLEM DETAILED ELABORATION The nervous systems of humans and animals must encode and process sensory information in the context of noise and interference and the signals which are encoded the images sounds etc have very speci c statistical properties One of the challenging task is how to reliably detect enhance and localize very weak non stationary and corrupted by noise brain source signals e g evoked and event related potentials EP ERP using EEG MEG data Independent Component Analysis ICA and related methods like Adaptive Factor Analysis AFA are promis ing approaches for elimination of artifacts and noise from EEG MEG data However most of the methods require manual detection and classi cation of interference components and or estimation of cross correla tion between each independent components and reference signals corresponding to speci c artifacts Main objective of this paper is to propose some rela tively simple techniques to automatically detect and elimi nate noise and some artifacts and classify independent brain sources Evoked potentials EPs of the brain are meaningful for clinical diagnosis and they are important factors to under stand higher order mechanism in the brain The EPs are usually embedded in the ongoing EEG MEG with signal to noise ratio SNR less than dB making them very di cult to extract using single trial The traditional method of EPs extraction is by using ensemble averaging to improve the SNR This often requires hundreds or thousands of trails to obtain a usable noiseless waveform Therefore it is impor tant to develop novel techniques that can rapidly improve the SNR and reduce to minimum the number of ensembles trials Traditional signal processing techniques such as Wiener ltering adaptive noise canceler latency corrected averaging and invertible wavelets transform ltering have been recently proposed for SNR improvements and en semble reduction However these methods require a priori knowledge pertaining to the nature of the signal Since EPs signals are known to be non stationary sparse and chang ing their characteristic from trial to trail it will be essential in the future to develop novel algorithms for enhancement of single trail EEG MEG noisy data The formulation of the problem could be given in the fol lowing form Denote by x t x t x t xn t the observed n dimensional vector of noisy signals that must be cleaned from the noise and interferences Here we have two types of noise The rst is so called inner noise generated by some primary sources which cannot be observed directly but contained in the vector of observations that is mixture of useful signals and random noise signals or other undesir able sources and second type of noise is the sensor additive noise observation errors at the output of measurement sys tem This noise is not measurable directly also Formally we can write that observed n dimensional vector of sensor signals x t x t x t xn t is mixture of source signals plus observation errors