Abstract
Electroencephalography (EEG) modeling has traditionally struggled to reconcile the complex temporal dynamics of brain signals with their inherent spatial topology. While large-scale pre-trained models have advanced the field, most treat multi-channel EEG as simple temporal sequences, neglecting the hierarchical relationship between global brain states and local channel activity. This paper introduces Topology Hierarchical Derived Brain Autoregressive (THD-BAR), a framework that organizes brain activity into a five-level spatial hierarchy. By redefining autoregressive learning as a "next-scale-time" prediction problem, THD-BAR captures physiological dynamics across multiple resolutions. Evaluated on a massive corpus of 17 pre-training datasets and 10 downstream tasks, THD-BAR achieves new state-of-the-art results in emotion recognition, sleep staging, and epilepsy detection, demonstrating significant performance gains through hierarchical spatial-temporal integration.
Introduction
Electroencephalography (EEG) serves as a critical, non-invasive window into human cognitive states, forming the foundation for modern Brain-Computer Interfaces (BCIs). However, the high dimensionality, low signal-to-noise ratio, and non-stationarity of EEG data make it notoriously difficult to model. Recent efforts have attempted to apply the "foundation model" paradigm—successful in Natural Language Processing (NLP)—to EEG. Yet, a fundamental gap remains: standard autoregressive models typically treat multi-channel EEG as a flat temporal sequence, predicting the next time step while ignoring the rich spatial topology of the brain.
Neural patterns are not merely sequences of numbers; they are spatially distributed phenomena. Whether processing an emotion or executing a motor command, the brain operates across nested scales, from global oscillations to localized channel-specific activity. THD-BAR addresses this by moving beyond "next-token" prediction to a "next-scale-time" approach, ensuring the model learns the physiological structure of the signal alongside its temporal progression.
The THD-BAR Framework
The THD-BAR framework is built upon three core innovations designed to bridge the gap between spatial topology and temporal modeling.
1. Brain Topology Hierarchy (BTH)
To move away from treating channels as a flat list, the researchers defined a Brain Topology Hierarchy (BTH) consisting of five nested spatial scales:
- S1 (Whole Brain): Represents global, aggregate brain activity.
- S2–S4 (Regions & Clusters): Intermediate groupings based on anatomical proximity and functional regions.
- S5 (Individual Channels): The finest granular resolution, representing specific electrode data.
This hierarchy allows the model to view the brain as a structured system rather than a collection of independent sensors.
2. THVQ-VAE: The Multi-Scale Tokenizer
Standard Vector Quantized-Variational Autoencoders (VQ-VAEs) compress data into a single set of discrete tokens. In contrast, the Topology-Hierarchical VQ-VAE (THVQ-VAE) acts as a specialized neural tokenizer that generates discrete tokens across all five spatial scales. It utilizes "downscale" operations to aggregate features for a global view and "upscale" operations to distribute features for local views. This ensures that the latent space explicitly represents the brain's hierarchical structure.
3. The Brain Autoregressive (BAR) Module
The BAR module redefines the objective of pre-training. Instead of predicting the next moment in time for all channels simultaneously, it performs "next-scale-time prediction." Within a single time step, the model predicts tokens hierarchically from the coarsest scale (S1) to the finest (S5). This process is governed by a "Scale-Time-wise" causal mask, forcing the model to understand the global context before refining its prediction for individual channels. Once the spatial hierarchy for time $t$ is resolved, the model moves to time $t+1$.
Experimental Setup and Scaling
The model underwent a rigorous two-stage evaluation process:
- Pre-training: The authors utilized a massive corpus of 17 diverse EEG datasets (including SEED-IV, SEED-V, EmoBrain, and various TUH subsets), encompassing hundreds of hours of multi-channel data to build a robust foundation.
- Downstream Validation: The model was fine-tuned on 10 distinct datasets across five major BCI domains: Emotion Recognition (DEAP, SEED), Motor Imagery (MIBCI, BCIC4-1), Mental Workload (EEGMat, STEW), Sleep Staging (EDF, HMC), and Epilepsy Detection (TUAB, TUEV).
To explore the impact of model capacity, the researchers developed three versions: Base, Large, and Huge. The results demonstrated clear "scaling laws" similar to those seen in LLMs; performance improved consistently as the model scaled from 124 million to 1.5 billion parameters.
Key Results and Insights
THD-BAR consistently outperformed existing state-of-the-art (SOTA) models, including NeuroLM and LaBraM, across nearly all benchmarks:
- Emotion Recognition: On the DEAP dataset, THD-BAR achieved 43.9% balanced accuracy, and 73.9% on SEED, outperforming NeuroLM by approximately 3.8% and 3.7%, respectively.
- Epilepsy Detection: The THD-BAR-Huge model reached 65.3% accuracy on the TUEV dataset, representing a significant leap over previous benchmarks.
- Spatial Fidelity: Qualitative analysis of the THVQ-VAE reconstructions showed that the hierarchical approach preserves the spatial heatmaps of brain activity much more accurately than standard VQ-VAEs, which tend to "blur" spatial transitions.
An essential finding from the ablation studies was that spatial order is as vital as temporal order. Models that used only the coarsest or only the finest scales performed significantly worse than the full hierarchical model. This confirms that for EEG, "context" is defined both by what happened previously and what is happening simultaneously across the rest of the scalp.
Limitations and Future Directions
Despite its performance, THD-BAR faces challenges regarding computational intensity. The multi-scale prediction and high parameter count (1.5B) may limit its current use in real-time, low-power BCI applications. Furthermore, the current hierarchy relies on fixed anatomical templates (the 10-10 system). Future research may explore data-driven topology, where the hierarchy is learned dynamically to accommodate different electrode montages or individual physiological differences. Finally, expanding the framework to include multi-modal data—such as fNIRS or behavioral metrics—could lead to a more holistic foundation model of human cognitive states.
Conclusion
THD-BAR represents a significant shift in EEG modeling by treating the brain's spatial topology as a first-class citizen in the autoregressive process. By organizing data into a hierarchical structure and predicting across scales and time, the model captures the complex, multi-resolution nature of neural activity. With SOTA performance across five major domains and proven scalability, THD-BAR provides a powerful new foundation for the next generation of brain-computer interfaces.
Summary updated