State space models outperform transformers for decoding brain activity in motor control

Researchers have developed a novel neural network architecture designed to better extract meaningful patterns from brain signals during motor imagery—the act of imagining movement without actually moving. The approach, called Cortical-SSM, uses state space models to analyze electroencephalogram (EEG) recordings and could enable more reliable brain-computer interfaces for patients with motor disabilities.

The challenge with brain signal decoding#

Classifying EEG data during motor imagery has promising real-world applications, from communication aids to physical rehabilitation systems for individuals with paralysis or motor impairment. However, the signals are noisy and prone to contamination from physiological artifacts such as eye blinks and swallowing, which can degrade accuracy.

Recent advances have favored Transformer models—attention-based architectures widely used in modern machine learning—for EEG classification. Yet these models often fail to capture the nuanced, multi-layered patterns embedded in brain signals, which unfold simultaneously across time, space (different electrode locations), and frequency bands.

How Cortical-SSM works differently#

Rather than relying solely on attention mechanisms, the new architecture extends deep state space models to simultaneously capture dependencies across all three dimensions of EEG data. State space models excel at tracking how systems evolve over time, making them well-suited to the inherently temporal nature of neural activity.

The researchers validated Cortical-SSM on two large public datasets containing over 50 subjects. The method outperformed existing baseline approaches on both benchmarks. Critically, the model's learned representations align with known neurophysiology—the visual explanations generated by the system highlight brain regions and patterns that are biologically meaningful, not arbitrary artifacts of the training process.

Implications for clinical deployment#

By combining strong classification performance with interpretability grounded in neuroscience, Cortical-SSM addresses a key barrier to translating AI into clinical practice: the "black box" problem. When doctors and patients can understand which brain signals the system relies on, confidence in the technology increases.

The authors argue that this approach advances the feasibility of —building models that generalize across different people without requiring extensive personalized recalibration. This capability is essential for practical, deployable brain-computer interface systems that serve diverse user populations.