A0507
Title: MIDAST as a novel approach to multivariate data segmentation - smart brain monitoring
Authors: Justyna Witulska - Wroclaw University of Science and Technology (Poland) [presenting]
Marta Hendler - Wroclaw University of Science and Technology (Poland)
Magdalena Kasprowicz - Wroclaw University of Science and Technology (Poland)
Marek Czosnyka - University of Cambridge (United Kingdom)
Ireneusz Jablonski - Fraunhofer Institute for Photonic Microsystems (Germany)
Agnieszka Wylomanska - Wroclaw University of Science and Technology (Poland)
Abstract: The research addresses the problem of identifying distributional changes in multivariate non-Gaussian data. A novel and general methodology, called MIDAST, is introduced for fusion-based segmentation of multivariate data using statistical tests (e.g., the Kolmogorov-Smirnov test, a maximum mean discrepancy-based test, and a kernel-based test). The proposed approach is evaluated against baseline methods, including E-Divisive and Kernel Change-point Analysis (KCPA), with segmentation accuracy and computational complexity as key performance indicators. The methodology is tested through computer simulations, using multivariate sub-Gaussian and multivariate Student's t distributions, under varying degrees of correlation, degrees of freedom, and stability indices. MIDAST, enhanced by a windowing mechanism, enables the detection of one or multiple change points that indicate shifts in the statistical properties of the system. A real-world application demonstrates the ability of the method to reduce invasiveness in intracranial hypertension event detection by identifying structural changes in multivariate temporal patterns. The proposed approach facilitates more accurate and adaptive monitoring of complex systems governed by evolving statistical dynamics.
