CCN meeting | Milena Čukić (University of Belgrade, Serbia)

11-05-2017 from 15:00 to 16:00
Henri Dunantlaan 2, room 4.3 (4C)
Add to calendar

Nonlinear analysis of electrophysiological signals and machine learning application in clinical practice

From last four decades of research it is well-established that all electrophysiological signals are nonlinear, irregular and aperiodic (and usually very far from equilibrium state). The physiological meaning behind it could be the need of an organism to adapt to external changes of conditions. Since those signals are used in everyday clinical practice as diagnostic tools (electromyograms EMG, electrocardiograms ECG, electroencephalograms EEG), a huge progress in using it in making diagnostic more precise and even personally oriented medicine, could be made.

One of the obstacles in doing so is deeply rooted application of classical (spectral) methods of analysis which are present even in all the machine software for recording of mentioned signals, namely Fourier’s analysis and its derivatives. Health practitioners are used to it. But they are not used to nonlinear methods of analysis of electrophysiological data, although it is proved many times that in comparison to widely accepted FFT, other methods stemming from Fractal analysis and Complex system’s theory (Chaos) are much more efficient in detecting subtle changes in those signals.

Many researchers used nonlinear measures and methods to describe and quantify the changes in a living system in different fields of neuroscience and biomedical applications. And with success. Similar to this contrast between classical (spectral) and nonlinear methods of analysis of signals, we can also make a comparison between classical statistical tools and data mining. When one combine the power of nonlinear measures for analysis of electrophysiological data with Machine learning methods many important applications in clinical practice can be yielded.

One of examples (a research by M. Čukić et al, submitted to publication) is very highly accurate method for classification of recurrent depressive disorder by employing nonlinear characterization of EEG and several machine learning methods in order to differentiate not only patients from controls but also the difference between those in exacerbation and remission. Clinicians are aware of close to impossible task for present diagnostic tools to make precise distinction between the phases of the disease, so they can safely change the medication. With the help of nonlinear characterization of EEG and machine learning this could be changed for the better.

Another successful example is early detection of Parkinson’s disease (PD) based on EMG or EEG, employing nonlinear measures of analysis of data combined with Machine learning (ML). Since this is another study in a process of a publication it is not possible at the moment to describe a Method in full detail, but high accuracy combined with very promising goal of early detection could be of potential screening test for larger population. In both cases described, most importantly is the possibility of increasing of a life quality for those who suffer from those debilitating disorders. It is possible, based on this approach to detect early onset of a disease, and to offer to clinicians a new low cost tool which can help them treat their patients in more effective way. In addition to this detection approach and powerful classification, a combination with Big Data is also possible. By connecting with clinical medicine specialists further study of possible applications could yield many benefits.