Tapio Grönfors

Epileptic seizure detection: A nonlinear viewpoint

This study concerns the detection of epileptic seizures from electroencephalogram (EEG) data using computational methods. Using short sliding time windows, a set of features is computed from the data. The feature set includes time domain, frequency domain and nonlinear features. Discriminant analysis is used to determine the best seizure-detecting features among them. The findings suggest that the best results can be achieved by using a combination of features from the linear and nonlinear realms alike.

Latency estimation of auditory brainstem response by neural networks

In the clinical application of auditory brainstem responses (ABRs), the latencies of five to seven main peaks are extremely important parameters for diagnosis. In practice, the latencies have mainly been done by manual measurement so far. In recent years, some new techniques have been developed involving automatic computer recognition. Computer recognition is difficult, however, since some peaks are complicated and vary a lot individually. In this paper, we introduce an artificial neural network method for ABR research. The detection of ABR is performed by using artificial neural networks. A proper bandpass filter is designed for peak extraction. Moreover, a new approach to estimate the latencies of the peaks by artificial neural networks is presented. The neural networks are studied in relation to the selection of model, number of layers and number of neurons in each hidden layer. Experimental results are described showing that artificial neural networks are a promising method in the study of ABR.

Peak identification of auditory brainstem responses with multifilters and attributed automaton

An attributed automaton, a special case of attribute grammar, is a flexible tool in pattern recognition. It allows the utilization of contextual information from previously analyzed patterns in the analysis of the current pattern, and offers the possibility of describing those structural characteristics of patterns which cannot be described by classic methods of syntactic pattern recognition. Auditory brainstem responses are routinely used in audiology and otoneurology. Many studies on using the spectral analysis of averaged auditory brainstem responses have described at least two frequency bands, corresponding to the slow and fast components. Selective non-recursive digital filters for each frequency band in the spectrum of the auditory brainstem response have revealed enhancement or attenuation of components, depending on the band. In this study, multi-filters and an attributed automaton were combined for the identification of peaks.

AR parameter estimation by a feedback neural network

Abstract As known, estimation of autoregressive (AR) parameters is fundamental in signal processing and time-series analysis. It can be widely used in many applications. Several sequential algorithms have been developed, such as the well-known Levinson-Durbin algorithm, Burg algorithm and Marple algorithm. Much research is still ongoing in this field. In this paper, we presesnt a new approach that takes advantage of a feedback recurrent neural network to estimate the AR parameters in a parallel way. It is a kind of dynamic mechanism that implement calculation by dynamic evolution. The algorithm was provided and expressed by applying the neural network technique. It also gives the strict proof of stability of the dynamic system and condition of convergence. Besides, we contribute the practical and simple hardware. The algorithm can be easily and simply converted into hardware so that it runs very quickly. The dynamic system is expressed by the first-order ordinary differential system, so we use a numerical method to approximate the solution. The simulation has been done by Runge-Kutta method with respect to simulated signal and auditory brainstem response (ABR) which is commonly used in medical research. Results showed that the method is effective to estimate AR parameters.

Identification of auditory brainstem responses

Auditory brainstem evoked responses are routinely used in audiology and otoneurology. An automatic method can be used to roughly classify the responses into probably normal, probably abnormal, and uncertain cases. Interpretation of the auditory brainstem response is a multistage process. Essential tasks are to detect individual peaks in the waveform and to choose representatives examples as medically interesting Jewett components. Our method is based on the comparison of detected peaks and normal values by means of an evaluation function, choosing representatives so that the values of this function will be maximized. We have studied the effects of some evaluation functions on the ability to correctly classify an evoked response. It turns out that the choice of an appropriate evaluation function is crucial in some problematic cases.

Evaluation of some nonrecursive digital filters for signals of auditory evoked responses

Auditory brain stem evoked responses are routinely used in audiology and otoneurology. Because recordings include more or less noise, the signals of evoked responses need digital filtering to suppress the noise. Nonrecursive digital filters are often the best since they can be organized to have no phase shift, which is essential in order not to distort sensitive parameters, as latency, in evoked responses. We have studied effects of some nonrecursive digital filters on the latency parameters of evoked responses. It turned out that digital filtering can have considerable influence on latencies, and thus the choice of appropriate filters is crucial.

Experiments and comparison of inference methods of regular grammars

Some common algorithms for regular grammatical inference with different regular grammars have been tested in order to clarify how they can consider those languages. The methods of so-called successor, canonical derivative, k-tails, tail-clustering, and skeleton have been examined experimentally. The two last-mentioned methods were shown to be the best and the most general when inferring regular grammars that were close to minimal initial grammars used to generate input strings for the inference process. It was noticed that the quality of the inferred grammars depends to a considerable extent on the properties of input strings. >

A scheme of inference of regular grammars for the syntactic pattern recognition of saccadic eye movements

Saccadic eye movements are measured as digital signals which are then transformed to symbol strings of formal grammars. At first, saccades are identified from a digital signal by a syntactic pattern recognition technique developed earlier. Using strings we automatically infer regular grammars which can be applied to reason different types of saccades as normal, or abnormal and affected by otoneurological disorders. We have tested this scheme with some pathological saccades and found such fully automatic recognition possible at least in principle.

Lossy compression of eye movement and auditory brainstem response signals.

Eye movement and auditory brainstem response signals recorded for balance and hearing investigations were used as a medical test battery for several types of lossy compression techniques. These signals are associated with the function of the ears. The former signals are used to assess the balance problems (especially vertigo) of a subject and the latter his or her hearing problems. New technique is also presented based on successive approximation quantization. The effect of information loss on medical parameters computed from the signals in the course of compression was evaluated for brainstem response signals. It is important to ensure that lossy compression techniques of these biomedical signals do not impair medical parameter values computed from the signals.

Segmentation of auditory brainstem response signals.

Auditory brainstem responses are used to detect hearing defects in audiology and otoneurology. The use of computer programs for the analysis of such recordings is increasing. To identify their detailed properties a pattern recognition algorithm implemented in an analysis program must be highly reliable. For the recognition process, some preprocessing phases after recording the necessary, such as filtering and often also segmentation. In the following, we will explore segmentation, which can be used in preprocessing of biomedical signals after filtering. We studied linear segmentation, where slopes of short signal segments are computed and divided into different classes according to their values. A segment length of 8 samples for a sampling frequency of 50 kHz employed was best according to our tests and error criteria. Using clustering, we found that less than 10 segment classes is suitable for pattern recognition.