[Paper Review] HCFT: Hierarchical Convolutional Fusion Transformer for EEG Decoding

Haodong Zhang, Jiapeng Zhu+2 more|arXiv (Cornell University)|Jan 18, 2026

EEG and Brain-Computer Interfaces0 citations

TL;DR

HCFT introduces a lightweight dual-branch convolutional encoder with cross-attention and hierarchical Transformer fusion for EEG decoding, achieving state-of-the-art results on MI classification (BCI IV-2b) and seizure prediction (CHB-MIT).

ABSTRACT

Electroencephalography (EEG) decoding requires models that can effectively extract and integrate complex temporal, spectral, and spatial features from multichannel signals. To address this challenge, we propose a lightweight and generalizable decoding framework named Hierarchical Convolutional Fusion Transformer (HCFT), which combines dual-branch convolutional encoders and hierarchical Transformer blocks for multi-scale EEG representation learning. Specifically, the model first captures local temporal and spatiotemporal dynamics through time-domain and time-space convolutional branches, and then aligns these features via a cross-attention mechanism that enables interaction between branches at each stage. Subsequently, a hierarchical Transformer fusion structure is employed to encode global dependencies across all feature stages, while a customized Dynamic Tanh normalization module is introduced to replace traditional Layer Normalization in order to enhance training stability and reduce redundancy. Extensive experiments are conducted on two representative benchmark datasets, BCI Competition IV-2b and CHB-MIT, covering both event-related cross-subject classification and continuous seizure prediction tasks. Results show that HCFT achieves 80.83% average accuracy and a Cohen's kappa of 0.6165 on BCI IV-2b, as well as 99.10% sensitivity, 0.0236 false positives per hour, and 98.82% specificity on CHB-MIT, consistently outperforming over ten state-of-the-art baseline methods. Ablation studies confirm that each core component of the proposed framework contributes significantly to the overall decoding performance, demonstrating HCFT's effectiveness in capturing EEG dynamics and its potential for real-world BCI applications.

Motivation & Objective

• Motivate robust EEG decoding that captures fine-grained temporal rhythms, spatial electrode patterns, and multi-scale global dependencies.
• Propose HCFT to fuse dual-branch CNN encoders with hierarchical Transformer blocks.
• Improve training stability with Dynamic Tanh normalization and cross-attention-based feature alignment.

Proposed method

• Dual-branch depthwise separable convolutional encoder extracts temporal and spatiotemporal features.
• Cross-attention mechanism aligns temporal and spatiotemporal features at each stage.
• Hierarchical Convolutional Fusion Transformer blocks fuse features across multiple scales.
• Dynamic Tanh normalization (DyT) as an optional replacement for LayerNorm to stabilize training.
• Pyramidal multi-stage encoder with stage-wise pooling and final global attention before classification.
• Classification via final multi-head attention, LayerNorm or DyT, global average pooling, and fully connected head.

Experimental results

Research questions

• RQ1 How can temporal and spatiotemporal EEG features be effectively aligned and fused across multiple scales?
• RQ2 Can a lightweight dual-branch CNN with Transformer fusion achieve strong cross-subject generalization in MI and robust seizure prediction?
• RQ3 Does Dynamic Tanh normalization improve training stability and generalization across EEG tasks?
• RQ4 What is the contribution of each core HCFT component to decoding performance?

Key findings

• HCFT achieves 80.83% average accuracy and 0.6165 Cohen’s kappa on BCI IV-2b (MI classification) under LOSO, outperforming 15 baselines.
• On CHB-MIT seizure prediction, HCFT attains 99.10% sensitivity, 0.0236 false positives per hour, and 98.82% specificity.
• Ablation studies show cross-attention, self-attention, stage-wise concatenation, and final MHSA all contribute to performance gains.
• DyT normalization improves MI-task performance over LayerNorm, while LayerNorm performs better for seizure prediction, with DyT offering smaller model size and FLOPs.
• Embedding dimension and head counts (D=32, H=2) balance accuracy and efficiency; deeper Stage 3 improves performance.