Personal homepage of Prof. Tapio Seppänen
Dr. Tapio Seppänen
Professor of biomedical engineering
Computer engineering laboratory
My research activities concentrate on pattern recognition and digital signal processing applications. I mostly work with a problem-based approach where I try to find suitable methods and algorithms for solving particular problems. These problems may come from technical applications as well as from biosciences. Activities can be grouped in two categories:
The complexity of the signals in our target applications provides big challenges for research. For example, adaptive noise filtering methods are often needed due to the non-stationarity of the signals, accurate segmentation methods are necessary for detecting sudden changes in signal characteristics, and the dynamic nature and non-linearity of the system in question should be taken into account in signal characterization. The fusion of information from many sensors must also be considered in some applications
Adaptive noise filtering is usually the first step in signal interpretation. Non-stationary signals create challenges for filtering because the noise model must be promptly adapted as the noise characteristics change. With technical systems the non-stationarity can sometimes be explained with the drifting properties of electronic devices. Other reasons may be due to dynamically changing operation of a target system or its environment. Reactive systems, for example, change their immediate operation as a response to changes in their environmental conditions. With biological systems it may even be difficult to state whether a baseline fluctuation, for example, is signal or noise. We shall solve adaptive filtering problems case-by-case in our applications. Various approaches have been and will be investigated, like various frequency-domain and time-domain methods, but also signal decomposition-based methods (wavelets, projection pursuit). A new method for adaptive filtering of periodic signal components was developed by us which has already led to an application for an international patent. Enhancements are being presently developed to this algorithm and new applications will investigated.
In many of our applications, signal characteristics change suddenly as the underlying system collapses or changes operational mode. Signal segmentation methods will be researched which accurately find out time instants where this occurs. This will enable a detailed analysis of each segment in order to explain system behaviour. For example, we have recently been working on a new adaptive segmentation method for electrical brain signal (EEG) classification purposes.
Signal characterization is the next step towards signal interpretation. In our research we have explored various methods to compute signal features for event recognition and classification. Among these approaches are the first and second order statistics, parametric modeling based on the theory of stochastic processes with applications in biosignal analysis, signal decomposition into meaningful components, and phase synchronization dynamics of cortical neural populations by multi-channel EEG analysis. Multidimensional scaling has been explored to develop useful tools for visualizing the dynamics of complex systems, and fractal dimensions, symbolic dynamics, morphological analysis and phase-space analysis have been applied to characterize ECG signals. In addition to exploring theses issues further, the application of new methods, like independent component analysis, multi-channel signal analysis and neural networks, is among our current research topics. One research problem that is gaining more importance is the characterization of dynamic nature of systems. Non-linearity creates further challenges; in some cases it has led, for example, for a need to apply chaos theory. Signal decomposition, statistical pattern recognition and clustering analysis will be applied for mining signal characteristics that are useful for signal classification and event recognition. A major focus in forthcoming years will be the exploration of algorithms for finding regularities in event-related responses of EEG that can help diagnosing patient condition.
Some of our applications include many different sensors. In order to derive useful information of the target system, sensor fusion is needed. During forthcoming we will explore methods for fusing data at various levels of abstraction, like raw data level, feature level, and concept level. Bayesian networks will be studied for interpreting uncertain sensor measurements. Multi-sensor intra-cardial measurements of electrical heart behaviour will be performed and signal analysis methods will be developed for diagnosing localized heart failures.
Most important application cases in biosignal analysis are heart monitors and automatic measurement of anesthesia depth:
Our group has studied heart signal analysis and applications more than 10 years in cooperation with the University Hospital of Oulu and local companies. The overall goal of the research is to find methods for advanced clinical diagnostic systems, self-management of health and the quality of life. Our role in this cooperation is to develop new algorithms and software for digital signal analysis and classification while the hospital runs extensive analyses with the clinical data. The most important application has been the detection of heart diseases. All the methods have based on computing various parameters of heart rate variability (HRV) and using advanced classifiers. Our approach is non-invasive measurement of these parameters using commercial heart rate monitors or so-called Holter devices of hospitals.
Non-invasive measurement devices with intelligent signal interpretation algorithms enable new kind of products. For example, citizens of the future can carry with them a tiny health monitor that gives indications of suspected heart failures, stress or other potentially dangerous conditions. The device can even instruct them to have a short break during a stressful physical exercise. Doctors can monitor critical patients that are outside the hospital doing their business if wireless Internet connections have been activated in the measurement device. Our work contains two main branches: one for people with a good medical record and the other for the detection of different health risks among the elderly people.
Our HRV research team was nominated as one of the Quality Research Groups of University of Oulu in June 2001. The nomination includes research financing for 3 years from the university.
Automatic measurement of anesthesia depth:
Determining the depth of anesthesia in operating rooms of hospitals is extremely important in order to inject a proper amount of anesthetics into a patient. With too small an amount a patient may not be entirely unconscious and painless during the operation, which may lead to a devastating post-surgery trauma due to the unnecessary pain and psychological effects. An excessive amount of anesthetics is stressful to the body. In the difficult conditions of intensive care units this is also of utmost importance because the patients may be in a bad physical condition and normal doses may not be applicable but must be used with additional care.
We have cooperated with the Cognitive Laboratory of the University Hospital of Oulu in developing new methods of determining the depth of anesthesia. Our role in this cooperation is to develop new algorithms and software for digital signal analysis and classification while the hospital runs extensive analyses with the clinical data. Our model-based methods analyze multi-electrode EEG signals to find out signal sources in the brain, coherence of neural populations and variability of electrical brain activity. We study how these changes in signal dynamics correlate with the depth of consciousness. Signals from measurement devices like ECG Holters, fMRI and isotope devices are being fused in order to develop multi-modal approaches to analysis of brain activity.
The results of this research will potentially yield next generation hospital devices that assist doctors by giving accurate statements of the patient state during operation and giving suggestions of proper amount of anesthetics.
Another important application area is context-aware computation. There I develop algorithms for interpreting various sensor signals of intelligent systems, which may be autonomous robots or humans carrying small special devices including various sensors for measuring the environment. The smart software of the device then reacts to some changes in the environment or the behaviour of the person itself, in order to adapt its operation and the service it provides.
Furher links to our research activities:
In this research area I have activities in the following topics:
- prosodic analysis of speech
We investigate existing and new speech features that can differentiate between various global states of speakers. For example, emotional state, sex, and age can be recognized with some accuracy with these features. We presently perform statistical clustering analysis and classification experiments in this research topic.
- text prediction
This technology will enable new writing-aids to speed up writing on computer keyboard. Our methods are based on word frequency tables and syntactic analysis of written text.
- digital watermarking
Digital watermarking is a technology where hidden data is embedded to a host data in order to prove the identity of the data. For example, copyright information can be hidden in digital images such that it can be extracted from there only with a special algorithm. Watermark should also be resistant to malicious attacks trying to destroy it. We investigate methods to embedded watermarks in digital images, videos and audio recordings.
- content-based information retrieval
Much of the above algorithms can be applied to information retrieval of databases. The basic research of the MSP team yields results that are utilized in multimedia retrieval engines. The multimedia servers of information networks contain massive amounts of heterogeneous data, such as audio and video samples. The users of the retrieval engines of the future will have to define exact search criteria in order to find the information they are looking for. However, a user might be looking for a document whose name he/she does not know, in which case the retrieval engine must be capable of understanding criteria that describe the content of the document, rather than its name. When this is the case, the retrieval engine will have to analyze the documents in the database to recognize their content and to be able to locate the document the user is looking for. This analysis requires evolved techniques for the automatic interpreting of audio, visual and textual information. A cross-disciplinary approach is taken here in which researchers from information technology, applied mathematics and linguistics work together in the same team to solve these problems.
Furher links to our research activities:
Prof. Tapio Seppänen
E-mail: email@example.com (with ‘ä’ -> ‘a’)
Phone: +358-8-553 2897
Last updated: 28th March, 2001