Acronym:

A software system for parallel analysis of multiple heterogeneous time series with application in biomedicine 

From: 2015-10-01 To: 2018-09-30 (Completed)

HRZZ 

- 

dr.sc. Alan Jović

- 

The task of time series analysis is to discover and classify significant patterns in data that contain a temporal component. Temporal analysis is conducted in various areas of science, such as: econometrics, meteorology, and biomedicine. This project deals with development of an integrated software system that includes general and domain specific time series features, with application in biomedicine. The goal of the project is to develop an efficient and upgradeable system for automatic classification of human body disorders based on the analysis of multiple heterogeneous biomedical signals (heart rhythm, ECG, EEG, etc.). Besides the construction of classification models, the project will pursue visualization of disorders using computer graphics. The system will be implemented in the Java 8 programming language, which will enable both independence from underlying operating system as well as high calculation efficiency. In order to increase calculation speed, multithreading will be used. The system will encompass subsystems for: 1) selection, display, and pre-processing of one or more signals from input records, 2) parallel analysis and extraction of multiple domain specific and general signal features, 3) visualization of signals and disorders by using computer graphics, and 4) automatic construction and evaluation of the models. For evaluation purposes, the project will use referential databases of biomedical signals from the PhysioNet web portal and, if possible, anonymous records from local hospitals. One of the significant contributions of the project will be development of an expert subsystem for automatic recommendation of the set of features that should be extracted from the signals at hand, depending on the type of analysis. The general signal features which will be implemented will include relevant nonlinear dynamics features (phase space, fractal properties, entropies, etc.) for characterization of signals. Specialized, domain specific features will be implemented for each type of biomedical signal individually using the existing standards, medical guidelines, referential scientific literature in the areas of biomedical engineering and medicine, and in collaboration with medical experts. When constructing disorder models, dimensionality reduction of the feature space will be pursued by using various filter methods. Disorders will be modeled based on clear description machine learning algorithms such as classification rules as well as maximum accuracy algorithms such as decision tree ensembles. Within the scope of this interdisciplinary project, several contributions in the areas of computer science, biomedical engineering, and medicine are expected.