Cortical-SSM: A Deep State Space Model for EEG and ECoG Motor Imagery Decoding

TL;DR

This paper tackles MI decoding from EEG and ECoG by addressing limitations of Transformer-based approaches in preserving fine-grained temporal structure and frequency information. It introduces Cortical-SSM, a deep state space model extension comprising three modules: Wavelet-Convolution for interpretable frequency features, Frequency-SSM for per-frequency temporal-spatial dynamics, and Channel-SSM for electrode-specific temporal-frequency modeling. The approach yields state-of-the-art results across three MI benchmarks (OpenBMI, Stieger, and ECoG-ALS) and provides neurophysiologically meaningful, sample-agnostic visual explanations in temporal, spatial, and frequency domains. The method enhances MI-BCI applicability by delivering both high accuracy and interpretable insights into the neural substrates of motor imagery, while acknowledging domain-shift challenges and suggesting directions for future domain-adaptive extensions.

Abstract

Classification of electroencephalogram (EEG) and electrocorticogram (ECoG) signals obtained during motor imagery (MI) has substantial application potential, including for communication assistance and rehabilitation support for patients with motor impairments. These signals remain inherently susceptible to physiological artifacts (e.g., eye blinking, swallowing), which pose persistent challenges. Although Transformer-based approaches for classifying EEG and ECoG signals have been widely adopted, they often struggle to capture fine-grained dependencies within them. To overcome these limitations, we propose Cortical-SSM, a novel architecture that extends deep state space models to capture integrated dependencies of EEG and ECoG signals across temporal, spatial, and frequency domains. We validated our method across three benchmarks: 1) two large-scale public MI EEG datasets containing more than 50 subjects, and 2) a clinical MI ECoG dataset recorded from a patient with amyotrophic lateral sclerosis. Our method outperformed baseline methods on the three benchmarks. Furthermore, visual explanations derived from our model indicate that it effectively captures neurophysiologically relevant regions of both EEG and ECoG signals.

Figures (11)

• Figure 1: Task overview. The input is EEG or ECoG signals recorded while the subject imagines actions (e.g., elbow extension), and the model's output are predictions of the corresponding actions. The model also provides explicit interpretability across temporal, spatial, and frequency domains.
• Figure 2: Overview of the proposed Cortical-SSM. Given an EEG or ECoG signal, the Wavelet-Convolution module extracts interpretable frequency features by combining deterministic and adaptive features (left). The Frequency-SSM and Channel-SSM then model spatio-temporal features for each frequency component and temporal-frequency features for each electrode, respectively (middle). Finally, the resulting features $\mathbf{V}^{(L)}$ and $\mathbf{U}^{(L)}$ are fused for action prediction, while providing visual explanations in spatio-temporal and temporal-frequency domains, respectively (right).
• Figure 3: Sample-agnostic visual explanations of our proposed method on OpenBMI OpenBMI. Rows (a) and (b) display the temporal-frequency and spatio-temporal visual explanations for Subject 02, respectively. Columns (i) and (ii) correspond to left and right hand grasp, respectively.
• Figure 4: Sample-agnostic visual explanations of our proposed method on ECoG-ALS. Rows (a) and (b) display the temporal-frequency and spatio-temporal visual explanations for Session 07, respectively. Columns (i)–(iv) correspond to the output classes: (i) elbow extension, (ii) elbow flexion, (iii) hand extension, and (iv) hand grasp.
• Figure 5: Electrode distribution in ECoG-ALS. Electrodes 9--25 and 49--60, positioned within the precentral and postcentral gyrus, respectively, are located in the Hand Knob Area HandKnobArea.