Clinical MI ECoG Dataset Overview

Updated 24 October 2025

The Clinical MI ECoG dataset comprises intracranial recordings during imagined motor tasks, supporting BCI research, neurorehabilitation, and presurgical mapping.
Signal preprocessing involves artifact correction, spectral decomposition, and windowing to extract time-frequency features essential for accurate movement decoding.
Advanced analytical approaches using both linear and deep learning models leverage the dataset to enhance interpretability and enable real-time clinical applications.

A Clinical Motor Imagery (MI) Electrocorticography (ECoG) Dataset comprises intracranial neural recordings obtained during motor imagery tasks, chiefly from patients with motor impairments or undergoing neurosurgical evaluation. These datasets serve as benchmarks and development resources for brain-computer interface (BCI) research, rehabilitation technology, and neurophysiological exploration of movement-related cortical activity. MI ECoG datasets extend traditional overt-movement paradigms to imagined movements, thereby supporting applications in populations unable to perform overt motion and facilitating presurgical “prehabilitation” or BMI development.

1. Dataset Structure, Recording Protocols, and Subject Cohorts

Clinical MI ECoG datasets are typically obtained from human patients implanted with subdural macroelectrode arrays for clinical purposes (e.g., epilepsy monitoring, BCI trials, or neurosurgical mapping). Recordings are performed during prescribed MI tasks—such as imagined hand/arm/finger movements—sometimes along with real movement (RM) as a control or calibration condition (Korostenskaja et al., 2017, Śliwowski et al., 2021, Suzuki et al., 17 Oct 2025). Datasets vary in subject composition (single vs. multi-subject), clinical context, and recording duration. For example:

In BCI clinical trials (e.g., NCT02550522), a single subject with tetraplegia performed repeated 3D virtual hand translation tasks by means of MI, across 43–44 sessions spanning hundreds of minutes. Two 8×8 electrode grids (with 32 chessboard-sampled channels per hemisphere) were used, and signals sampled at 586 Hz (Śliwowski et al., 2021, Śliwowski et al., 2022, Śliwowski et al., 2022).
For neurorehabilitation or presurgical mapping, patients with epilepsy or ALS underwent experimental MI protocols entailing several sessions; a paradigmatic dataset may feature ~8 sessions with 94 electrodes (e.g., ECoG-ALS), with each session containing 40–50 trials per task (Korostenskaja et al., 2017, Suzuki et al., 17 Oct 2025).
Typical MI tasks involve cue-based attempts or imagination of discrete actions: hand/arm grasp, extension, flexion, or reaching to targets (Suzuki et al., 17 Oct 2025). Each trial is temporally structured into rest/pre-cue, cue, and imagery windows (often 2 s before and 2 s after onset) to facilitate analysis of task-locked neural responses.

Dataset organization generally includes multi-session, multi-trial matrices of shape (sessions × trials × channels × time), with labels indicating task type, timing, and often movement kinematics for calibration or ground truth purposes.

2. Signal Preprocessing and Feature Extraction

Raw ECoG signals undergo extensive preprocessing and transformation before statistical analysis or model training:

Artifact Correction and Referencing: Poor-quality or artifactual channels are excluded. Common average referencing (CAR) is commonly employed.
Spectral Decomposition: Continuous wavelet transform (CWT; Morlet basis, 10–150 Hz in 10 Hz steps) or bandpass filtering is used to extract time-frequency representations, yielding tensors of size (channels × frequencies × time) (Śliwowski et al., 2021, Śliwowski et al., 2022, Geyer et al., 2020).
Windowing: 1 s windows (often 90% overlap) are conventionally used, mapping to task epochs. The resulting tensor may be further processed into four modes—trials, electrodes, frequency, and time (Geyer et al., 2020).
Feature Engineering: For movement decoding, hand-crafted features include AR coefficients per channel (with temporal shifts), signal power in canonical bands (e.g., alpha: 8–13 Hz, beta: 18–28 Hz, high gamma: 70–170 Hz), or temporal/frequency statistics. Alternatively, end-to-end deep learning may learn optimized spatial and spectral filters directly (Śliwowski et al., 2022).

Further, task-specific preprocessing is conducted, e.g., mapping to anatomical parcellations for connectivity/network analyses (Betzel et al., 2017), or spatial filtering via common spatial patterns for event classification (Mendoza-Cardenas et al., 2021).

3. Analytical Approaches and Modeling Paradigms

A spectrum of analytical models is applied to MI ECoG datasets, reflecting methodological advances.

Linear and Sparse Models

Switching Linear Models: Used for decoding flexion trajectories (e.g., finger movements). These combine AR-based features, group-lasso channel selection, and mutually exclusive classifiers per finger state, followed by state-dependent ridge regression for kinematic prediction (Flamary et al., 2011).
Sparse and Functional Decomposition: Regularized higher-order principal components analysis (ρ‐PCA, ρ‐PLS) employs CP decomposition with ℓ₁ (sparsity) and functional (smoothness) penalties on electrodes and frequency factors, enabling interpretable extraction of spatial and spectral modes (Geyer et al., 2020).

Deep Learning-based Models

Convolutional (CNN), Recurrent (LSTM), and Hybrid Architectures: CNNs leverage the spatial arrangement of electrodes, while LSTMs capture sequential dependencies. For MI hand translation, models such as CNN2D+LSTM+MT process time-frequency tensors into translation vectors, outperforming multilinear baselines (e.g., cosine similarity increases from ~0.189 to ~0.302) (Śliwowski et al., 2021).
End-to-End vs. Hand-crafted Features: Fully optimized deep architectures may provide modest performance gains relative to models based on hand-engineered time-frequency features, with trade-offs in computational cost and explainability (Śliwowski et al., 2022). Central frequency-optimized variants constrain initial layers to Morlet-family filters for interpretable feature extraction.
State Space Models and Explainability: The Cortical-SSM model fuses deterministic wavelet features with learned Conv1D filters and leverages deep state space layers separately along frequency (Frequency-SSM) and channel (Channel-SSM) axes, preserving fine temporal and spatial structure (Suzuki et al., 17 Oct 2025). MST-ECoGNet combines signal processing (modified S Transform) with lightweight deep models, explicitly disentangling time, frequency, and spatial features, enhancing interpretability (Ji, 25 Nov 2024).

Alternative and Complementary Methods

Shift-Invariant Waveform Learning: Spherical shift-invariant k-means learns dictionaries of non-sinusoidal prototypical waveforms; bag-of-words encodings are then supplied to classifiers for seizure prediction (Mendoza-Cardenas et al., 2021).
Self-Supervised Pretraining: Reengineered wav2vec models perform self-supervised contrastive learning on unlabeled ECoG to discover latent representations, which, when used as input for supervised decoders, halve word error rates in speech-BCI tasks (Yuan et al., 28 May 2024).
Network and Systems Neuroscience Analytics: Statistical relational models incorporating inter-regional distances, white matter structural connectivity, and transcriptomic co-expression predict ECoG functional connectivity, enhancing clinical mapping and outcome prediction (Betzel et al., 2017).

4. Neurophysiological Signatures and Clinical Task Design

Clinical MI ECoG datasets enable precise delineation of physiological processes underpinning imagined and real movement.

Localization: For RM, ECoG changes reliably localize to perirolandic motor cortex (M1), while MI activation is less universal (occurring in both sensorimotor and frontal areas in a subset of subjects) (Korostenskaja et al., 2017, Suzuki et al., 17 Oct 2025).
Frequency Signatures: RM robustly engages alpha (8–13 Hz), beta (18–28 Hz), and high gamma (70–170 Hz) bands. MI activations are more restricted—typically beta and high gamma, lacking broad low-frequency engagement (Korostenskaja et al., 2017). Synchronization/desynchronization patterns also differ: RM produces both low-frequency desynchronization and high-gamma synchronization; MI predominantly engenders desynchronization.
Spatial Structure: Task-relevant electrodes cluster in known functional areas (e.g., the Hand Knob Area), a finding consistently visualized in attention maps from state-of-the-art deep models (Suzuki et al., 17 Oct 2025). MST-based approaches confirm the discriminative value of spatial and low-frequency features (Ji, 25 Nov 2024).
Temporal Dynamics: High temporal resolution allows for detection of task-locked spectral changes and even processing delays (e.g., ~50 ms lag post stimulus in visual decoding) (Ji, 25 Nov 2024).

A plausible implication is that subject-specific approaches—with individualized feature extraction and adaptable model structure—are beneficial, given MI responsivity varies across patients and sessions.

5. Performance Metrics, Adaptation, and Data Efficiency

Evaluation frameworks in clinical MI ECoG research span regression, classification, and information-theoretic metrics:

Decoding Performance: Cosine similarity in continuous regression (e.g., 3D hand translation) and accuracy/AUROC for classification (e.g., MI task discrimination) (Śliwowski et al., 2021, Suzuki et al., 17 Oct 2025). For state classification, average cross-correlation between predicted and ground truth trajectories provides model comparison (Flamary et al., 2011).
Data Size Effects: Model performance (e.g., cosine similarity) escalates rapidly with initial training data (≤40 min), then saturates—with only ~5% additional gain by extending training >60–90 min (Śliwowski et al., 2022). Deep and multilinear models exhibit similar sample-efficiency learning curves, but deep architectures achieve higher asymptotic performance.
Longitudinal Adaptation: Chronic BCI use promotes improved MI patterns—later sessions yield higher-quality signals, as revealed by UMAP embedding clustering and rising SVM separability (Śliwowski et al., 2022). Higher intrinsic dimensionality of the neural feature manifold correlates with improved decoding, indicating that richer neural representations support superior BCI performance.
Model Robustness: End-to-end deep learning may be more prone to overfitting or label noise under limited data regimes; pipelines leveraging established domain knowledge (e.g., wavelet hand-crafted features) offer more robust and explainable alternatives (Śliwowski et al., 2022).

6. Interpretability, Visualization, and Clinical Implications

Interpretability is essential in clinical translation, influencing trust, safety, and neurophysiological insight:

Automated Feature Selection and Visualization: Sparse regularization in tensor decomposition and model architecture (e.g., in ρ‐PCA, MST-ECoGNet) directly reveals the most informative channels and spectral bands (Geyer et al., 2020, Ji, 25 Nov 2024).
Neurophysiological Mapping: Attention maps generated by deep state space models identify classical motor regions during MI, providing validation and guiding further model refinement (Suzuki et al., 17 Oct 2025).
Clinical Usage: MI ECoG datasets advance applications including BCI-driven neuroprosthetics for motor-impaired patients, personalized presurgical mapping, seizure prediction, and prehabilitation strategies leveraging interhemispheric transfer (Śliwowski et al., 2021, Korostenskaja et al., 2017, Mendoza-Cardenas et al., 2021).
Model Efficiency and Real-Time Deployment: Lightweight models such as MST-ECoGNet achieve state-of-the-art performance with <10% the parameters of conventional architectures, facilitating real-time and embedded clinical use (Ji, 25 Nov 2024).
Data-Driven Clinical Decision Support: The integration of network neuroscience, transcriptomics, and spatiotemporal signal analytics provides multidimensional insight for surgical prognosis and brain mapping (Betzel et al., 2017).

7. Challenges, Limitations, and Future Directions

Several methodological and translational challenges persist in clinical MI ECoG research:

Signal Nonstationarity and Inter-Session Variability: ECoG signals are nonstationary and susceptible to inter-session and inter-patient domain shifts, necessitating robust cross-session validation and domain-adaptive models (Suzuki et al., 17 Oct 2025).
Model Complexity vs. Explainability: The marginal gains of deep end-to-end models are offset by increased computational burden and reduced interpretability in clinical contexts (Śliwowski et al., 2022).
Feature Robustness: RM/MI activity is variably observed across subjects (e.g., only 60% showing significant MI effect), supporting adaptive and individualized modeling strategies (Korostenskaja et al., 2017).
Extension to Larger Cohorts: Most clinical MI ECoG datasets are single-patient or small-cohort, and further work is required to robustly generalize findings and methods to broader clinical populations (Śliwowski et al., 2022, Suzuki et al., 17 Oct 2025).
Artifact Management: Preprocessing must robustly remove non-physiological artifacts to avoid model contamination, particularly in long-term recordings or high-noise clinical settings (Mendoza-Cardenas et al., 2021).
Advanced Temporal Modeling and Transfer Learning: Attention mechanisms, self-supervised pretraining, and cross-subject transfer can further mitigate data scarcity and improve decoding performance, especially for speech and limb control in motor/communication prostheses (Yuan et al., 28 May 2024, Śliwowski et al., 2021).

Future research avenues include scalable multisubject curation, real-time adaptation and closed-loop BCI deployment, comparative studies across feature engineering paradigms, and integration with complementary neuroimaging and omics data for holistic brain-machine interface development.