Balancing Interpretability and Performance in Motor Imagery EEG Classification: A Comparative Study of ANFIS-FBCSP-PSO and EEGNet

TL;DR

This study tackles MI-EEG classification by explicitly balancing interpretability and performance. It compares an interpretable ANFIS–FBCSP–PSO pipeline with a deep-learning EEGNet model on the BCICIV-2a dataset, evaluated under within-subject and LOSO protocols. Within-subject, ANFIS–FBCSP–PSO achieves higher accuracy ($68.58\% \pm 13.76$) and $\kappa$ ($58.04 \pm 18.43$) than EEGNet ($63.79\% \pm 8.49$, $51.54 \pm 11.67$), while cross-subject generalization favors EEGNet ($68.20\% \pm 12.13$; $\kappa = 57.33 \pm 16.22$) over ANFIS ($65.71\% \pm 14.89$, $53.66 \pm 20.52$). The results highlight a practical trade-off: interpretable, subject-specific models for personalized BCI vs scalable, robust generalization from end-to-end deep learning. The authors suggest future directions in hybrid and transformer-based neuro-symbolic frameworks to achieve transparent yet scalable EEG decoding for MI-BCIs.

Abstract

Achieving both accurate and interpretable classification of motor imagery EEG remains a key challenge in brain computer interface (BCI) research. This paper compares a transparent fuzzy reasoning approach (ANFIS-FBCSP-PSO) with a deep learning benchmark (EEGNet) using the BCI Competition IV-2a dataset. The ANFIS pipeline combines filter bank common spatial pattern feature extraction with fuzzy IF-THEN rules optimized via particle swarm optimization, while EEGNet learns hierarchical spatial temporal representations directly from raw EEG data. In within-subject experiments, the fuzzy neural model performed better (68.58 percent +/- 13.76 percent accuracy, kappa = 58.04 percent +/- 18.43), while in cross-subject (LOSO) tests, the deep model exhibited stronger generalization (68.20 percent +/- 12.13 percent accuracy, kappa = 57.33 percent +/- 16.22). The study provides practical guidance for selecting MI-BCI systems according to design goals: interpretability or robustness across users. Future investigations into transformer based and hybrid neuro symbolic frameworks are expected to advance transparent EEG decoding.

Figures (3)

• Figure 1: The EEGNet Architecture for Motor Imagery dataset. A compact and generalizable Convolutional Neural Network (CNN) for EEG classification, featuring Temporal, Depthwise, and Separable convolution blocks optimized for extracting frequency, spatial, and temporal features from raw EEG time series data.
• Figure 2: Our Proposed Methodology for EEG Classification on BCI IV 2A dataset
• Figure 3: The ANFIS-FBCSP-PSO Hybrid Architecture. This model integrates the FBCSP module for robust EEG feature extraction, an ANFIS for interpretable, rule-based classification, and Particle Swarm Optimization (PSO) for the hybrid tuning of the ANFIS parameters.