FORMALIZATION AND OPTIMIZATION OF SOFTWARE COMPONENT PIPELINES FOR ELECTROENCEPHALOGRAPHIC SIGNAL CLASSIFICATION

Abstract

The article proposes a formalized approach to the construction and optimization of software component pipelines for electroencephalographic (EEG) signal classification. This is caused by the limitations of existing methods, which are usually empirical when defining software component pipeline configurations, as well as by the lack of a formalized procedure for selecting optimal software components that considers not only one specific metric but a set of metrics. The aim of the study is to formalize and optimize software component pipelines as a basis for their automated usage in a software system for EEG signal classification. The article proposes using a tree-like directed graph structure to represent the software component pipeline. The nodes of the structure correspond to software components, while the edges define valid directions of data transfer between processing stages. To ensure acyclicity, a valid sequence of stages, compatibility of software components, and structural completeness of configurations, rules for constructing the structure are defined. The optimization procedure involves formalized multi-metric evaluation. The evaluation metrics include accuracy, F1 score, area under the ROC curve, and normalized processing time per signal sample. All metrics are normalized, brought to a common minimization direction, and combined into an objective function with weight coefficients. The search for the optimal software component pipeline configuration is implemented using mixed-integer linear programming. The experimental verification of the proposed solutions was performed using real EEG data related to the motor activity of human upper limb fingers. Within the constructed tree-like directed graph structure, 8 valid software component pipeline configurations were formed. According to the optimization results, the optimal software component pipeline configuration for the baseline scenario with equal metric weights was “PCA → CNN”. The obtained results confirmed the correctness of the proposed formalizations, their consistency with each other, and their practical suitability for automated construction and optimization of software component pipelines for EEG signal classification.