Arquitetura neuro-simbólica com floresta de caminhos ótimos para classificação de EEG através dos macroestados com GRPO

Abstract

Neurodegenerative dementias represent a growing challenge for early diagnosis, requiring computational methods capable of capturing dynamic changes in the functional organization of brain networks. This work proposes a neuro-symbolic artificial intelligence

architecture that combines unsupervised Optimum-Path Forest (OPF) with Deep Neural Network (DNN) and a decision strategy grounded in Group Relative Policy Optimization (GRPO), applied to the classification of electroencephalography (EEG) signals through microstates. OPF is employed to symbolically model the structure of microstate temporal parameters, while the DNN captures non-linear relationships and complex discriminative patterns. The combination of predictions is performed using a decision mechanism inspired by GRPO, which dynamically adjusts the relative contributions of each model based on their performance. The study used a dataset from AHEPA University General Hospital, consisting of resting-state EEG records from Cognitively Normal (CN), patients with Alzheimer’s Disease (AD), and patients with Frontotemporal Dementia (FTD). The methodology includes preprocessing stages, microstate extraction through modified k-means, a

systematic evaluation of 37 distance measures in the context of OPF, and stratified cross- validation. Results demonstrate that the proposed architecture achieves an average

accuracy of 97.31%±0.58 and an F1-score of 97.34%±0.58, even under class imbalance conditions, outperforming traditional approaches such as k-means, Self-Organizing Map (SOM), and isolated neural classifiers. The use of a relative decision policy improves performance stability and consistency in distance-measure rankings. The main contributions of this work include: (i) a neuro-symbolic architecture for EEG microstate classification; (ii) an extensive comparative analysis of distance measures in the OPF context; and (iii) the application of a GRPO-based decision strategy for adaptive model fusion. The results indicate relevant translational potential for clinical applications supporting the differential diagnosis of neurodegenerative dementias.