EEGForceMap: EEG-Based Grasp Force Decoding

• EEGForceMap is an EEG-based methodology that decodes continuous grasp force by transforming scalp recordings into neurophysiologically aligned representations.
• It employs ICA cleaning, task-specific channel and component selection, and constructs ERP, PSD, and ERDS features to capture force-related signals.
• Coupled with various regressors, EEGForceMap achieves up to 0.815 R², outperforming prior methods by emphasizing simpler, optimized signal representations.

EEGForceMap is an EEG-based methodology for continuous grasp force decoding that was introduced to estimate the time-varying force exerted by the fingers from noninvasive scalp recordings. Its defining claim is methodological rather than architectural: force decoding should not rely only on increasingly complex regressors, but should first transform EEG into a representation that is more closely aligned with the neurophysiology of force control. In the reported formulation, EEGForceMap isolates signals from the premotor-parietal network, extracts task-specific components, constructs three distinct time-frequency feature sets, and couples those features to linear, latent-variable, and neural regressors for prediction of the continuous grasp/load force signal (Dangi et al., 11 Aug 2025).

1. Definition and conceptual basis

EEGForceMap addresses the problem of learning a mapping from multichannel EEG to a continuous force target. The task is formalized as estimating a force trajectory $y(t)$ from a recent temporal window of EEG, $\mathbf{x}_{t-W:t}$, through a regressor $f$, yielding $\hat y(t)=f(\mathbf{x}_{t-W:t})$. The motivation is that scalp EEG is noisy, has low spatial resolution, and mixes force-related activity with eye, muscle, cardiac, and instrumentation artifacts. Within this framing, the methodological novelty lies in reshaping the signal before regression rather than relying primarily on more complex decoders (Dangi et al., 11 Aug 2025).

The method is explicitly positioned against prior force decoders that emphasized model complexity, including a multimodal EEG+EMG CNN-LSTM decoder, an EEG-only classifier-decoder ensemble using LDA + CNN-LSTM, and a spike-inspired SNN. Those methods achieved subject-specific coefficients of determination around $0.457$, $0.487$, and $0.504$, respectively. EEGForceMap was proposed to show that improved EEG selection and feature construction can make simpler decoders substantially stronger. In that sense, EEGForceMap is best understood as a neurophysiology-guided signal representation strategy rather than a single end-to-end deep architecture (Dangi et al., 11 Aug 2025).

The neurophysiological rationale centers on the claim that force is represented within a premotor-parietal network rather than only in a generic motor amplitude signal. This assumption governs both channel selection and feature engineering. The selected representation is intended to retain channels and latent components whose fluctuations covary with grasp force, thereby privileging task-relevant cortical dynamics over unrelated variance.

2. Signal curation and preprocessing pipeline

The reported experiments use the WAY-EEG-GAL dataset, comprising 12 healthy right-handed subjects with EEG recorded from 64 channels at 2048 Hz, alongside EMG, kinematics, and kinetic measurements. For the force-decoding experiments, the study uses only trials with constant surface and varying force levels. The kinetic channels provide the force labels, and the regression target is the total grasp/load force, obtained by summing the force sensor outputs from both fingers at each timestamp (Dangi et al., 11 Aug 2025).

Preprocessing begins with what the manuscript describes as a notch filter spanning 0.5–50 Hz, which it calls “line noise filtering.” Because that specification is unusual, it is most accurately treated as reported rather than normalized to a more conventional narrowband mains filter. After filtering, EEGForceMap applies independent component analysis (ICA) and uses ICLabel to tag components corresponding to line noise, muscle activity, eye movement, and cardiac artifacts. These non-neural components are removed, and the cleaned EEG is reconstructed from the retained neural components (Dangi et al., 11 Aug 2025).

A second preprocessing stage performs task-relevant channel selection. For a small subset of data from each subject, the covariance between each channel $x_c(t)$ and force $y(t)$ is computed as

$$\operatorname{Cov}(x_c,y)=\frac{1}{N-1}\sum_{t=1}^{N}\big(x_c(t)-\bar x_c\big)\big(y(t)-\bar y\big).$$

Channels with the highest sensitivity to force changes are retained. The selected set is

$\mathbf{x}_{t-W:t}$0

This subset is intended to capture activity spanning central, premotor, and parietal territories associated with movement planning and force control (Dangi et al., 11 Aug 2025).

A third block performs task-correlated component selection. ICA-derived components are again ranked by covariance with force, and the highest-covariance components are used to reconstruct an EEG signal that is more sensitive to force fluctuations. The paper does not specify the exact number of retained components. The resulting pipeline therefore combines artifact-aware ICA cleaning, covariance-based channel restriction, and covariance-based component reconstruction before feature extraction (Dangi et al., 11 Aug 2025).

Feature extraction is carried out in a sliding-window pseudo-online manner with a 100 ms window and a 50 ms step size, yielding 50% overlap. At $\mathbf{x}_{t-W:t}$1 Hz, each window contains approximately $\mathbf{x}_{t-W:t}$2 samples, and successive windows are offset by approximately $\mathbf{x}_{t-W:t}$3 samples. This stage is intended both to reduce sensitivity to samplewise fluctuations and to mimic real-time decoding (Dangi et al., 11 Aug 2025).

3. Feature construction and force-sensitive representation

EEGForceMap constructs three distinct time-frequency feature sets: ERP, PSD, and ERDS. These features are compared empirically and are motivated by prior studies of motor-force control (Dangi et al., 11 Aug 2025).

The ERP representation is not the raw waveform. For each selected signal and each window, the method computes the following statistics: mean, mean absolute value, area under the curve, skewness, kurtosis, and variance. Because this representation becomes high-dimensional, the paper applies principal component analysis (PCA) and retains enough principal components to explain 95% of variance. In the reported framework, ERP serves as a compact summary of short-window time-domain structure rather than as a traditional event-locked waveform analysis (Dangi et al., 11 Aug 2025).

The PSD representation targets oscillatory activation in force-relevant bands, especially theta (4–8 Hz) and mu (9–11 Hz). The paper gives the classical PSD definition,

$\mathbf{x}_{t-W:t}$4

and uses theta and mu band power as features. The spectral estimation details beyond this equation are not specified. As reported, PSD captures absolute oscillatory magnitude in bands hypothesized to participate in force planning and control (Dangi et al., 11 Aug 2025).

The most important feature family is event-related desynchronization/synchronization (ERDS). ERDS measures the relative change in spectral power during the task relative to a baseline, with the baseline defined from time points where the total load force is below 0.5 N. The intended formula is

$\mathbf{x}_{t-W:t}$5

ERDS is computed in the same theta and mu bands used for PSD. By normalizing against baseline, ERDS suppresses redundant absolute-amplitude variation and emphasizes task-locked changes in sensorimotor rhythms. This baseline normalization is the step the paper credits with making ERDS markedly more sensitive to grasp-force changes than raw PSD (Dangi et al., 11 Aug 2025).

The paper reports a clear empirical hierarchy among these feature families. ERP features show the weakest linear association with force, with Pearson correlation $\mathbf{x}_{t-W:t}$6. PSD features show moderate correlation, with $\mathbf{x}_{t-W:t}$7 for mu and $\mathbf{x}_{t-W:t}$8 for theta, but substantial redundancy, with coefficient of variation $\mathbf{x}_{t-W:t}$9 reported as 32.41. ERDS features show similar Pearson correlations—mu $f$0, theta $f$1—but much stronger sensitivity, with $f$2 and a reported comparison $f$3, $f$4. The paper interprets this as evidence that baseline normalization makes force-related modulations much more distinct (Dangi et al., 11 Aug 2025).

4. Regression models and experimental protocol

EEGForceMap is paired with four regressors. The first is a single-feature linear regressor (SFLR) that uses only ERP statistics and standard linear regression,

$f$5

The second is multiple linear regression (MLR), which uses all three feature sets together under the same linear form. The third is partial least squares regression (PLSR), described in the paper as a non-linear method, with hyperparameters tuned using grid search with cross-validation over 20 components. The fourth is a neural network regressor (NNR), a multilayer perceptron with an input layer, three hidden layers with 8, 16, and 32 neurons, and an output layer producing a continuous force value. The NNR maps ERDS features to continuous force values, minimizes mean squared error, uses the Adam optimizer, and all models are trained for 100 epochs (Dangi et al., 11 Aug 2025).

Performance is measured primarily with the coefficient of determination, termed CoD in the paper and equivalent to $f$6: $f$7 Feature-force association is assessed with the Pearson correlation coefficient, and coefficient of variation is used when correlation alone is insufficient. The evaluation includes both subject-specific and subject-independent protocols. In the subject-specific case, each participant’s data are randomly shuffled and split into 70% train, 20% test, 10% validation. In the subject-independent case, one subject is held out for testing and the remaining 11 subjects are used for training; this is repeated across all 12 leave-one-subject-out configurations (Dangi et al., 11 Aug 2025).

The model operates in a pseudo-online regime through the 100 ms / 50 ms sliding window. The experiments were run on a Dell Inspiron 15 3000, Intel i3-1115G4, 16 GB RAM, which the paper presents as evidence of a comparatively low computational burden relative to larger deep temporal models (Dangi et al., 11 Aug 2025).

Regressor	Subject-specific CoD	Subject-independent CoD
EEGForceMap + SFLR	0.261	0.156
EEGForceMap + MLR	0.425	0.356
EEGForceMap + PLSR	0.536	0.405
EEGForceMap + NNR	0.815	0.785

These values summarize the central empirical comparison within the EEGForceMap framework (Dangi et al., 11 Aug 2025).

5. Empirical performance, ablations, and statistical interpretation

The best-performing configuration is EEGForceMap + NNR, which reaches $f$8 under subject-specific conditions and $f$9 under subject-independent conditions. The paper states that this corresponds to a 61.7% improvement in subject-specific decoding over the strongest prior subject-specific baseline, the SNN at $\hat y(t)=f(\mathbf{x}_{t-W:t})$0. That percentage is directly verifiable from the reported numbers: $\hat y(t)=f(\mathbf{x}_{t-W:t})$1 The paper also reports a 55.7% improvement in subject-independent conditions; however, the manuscript’s tabulated values do not make that percentage fully reconstructable, so it is most accurately presented as the authors’ claim rather than as a directly reproducible computation (Dangi et al., 11 Aug 2025).

The statistical analysis reports a significant difference in CoD among the proposed models using two-way ANOVA,

$\hat y(t)=f(\mathbf{x}_{t-W:t})$2

It reports no significant inter-subject variation in that analysis,

$\hat y(t)=f(\mathbf{x}_{t-W:t})$3

and it reports that subject-specific models are significantly more accurate than subject-independent ones,

$\hat y(t)=f(\mathbf{x}_{t-W:t})$4

with a small effect size,

$\hat y(t)=f(\mathbf{x}_{t-W:t})$5

Within the model hierarchy, MLR improves over SFLR ($\hat y(t)=f(\mathbf{x}_{t-W:t})$6, $\hat y(t)=f(\mathbf{x}_{t-W:t})$7), PLSR outperforms the linear regressors ($\hat y(t)=f(\mathbf{x}_{t-W:t})$8, $\hat y(t)=f(\mathbf{x}_{t-W:t})$9), and NNR is clearly strongest. The paper notes one characteristic NNR failure mode: small random fluctuations during rest can be confused with the initial rise of force at movement onset (Dangi et al., 11 Aug 2025).

The ablation study is central because it supports the claim that EEGForceMap is a layered methodology rather than merely “ERDS + neural net.” Using the best-performing regressor setup, the paper reports a progressive CoD increase across the pipeline: 0.236 for raw data, 0.265 after line noise filtering, 0.333 after task-related component selection, 0.407 after task-related channel selection, and 0.785 after feature extraction from selected EEG data. The largest increase occurs at the final feature-extraction stage, especially ERDS. Removing feature extraction causes a very large degradation, with $0.457$0, $0.457$1. Component selection produces only a slight improvement ($0.457$2, $0.457$3), whereas channel selection becomes strongly beneficial when combined with feature extraction, with $0.457$4, $0.457$5 in that setting (Dangi et al., 11 Aug 2025).

Taken together, these results support a specific interpretation of the pipeline. Filtering and component selection clean the signal, channel selection increases relevance and efficiency, and ERDS-based feature construction is the decisive representation change that makes continuous grasp force highly decodable from EEG. The regressors differ substantially in how much of that curated representation they can exploit, with the modest multilayer perceptron yielding the best performance.

6. Relation to adjacent force-decoding and mapping paradigms

EEGForceMap belongs to a broader literature on noninvasive force decoding, but its role within that literature is specific. The most direct neighboring line is continuous force regression from cortical signals. For example, “CNNATT: Deep EEG & fNIRS Real-Time Decoding of bimanual forces” reconstructs continuous bimanual grip force trajectories from EEG + fNIRS using an 800 ms causal window and reports 55.2 FVAF[%] for the multimodal deep decoder, with population-level values of 33.5 FVAF[%] for EEG only, 40.0 FVAF[%] for fNIRS only, and 48.5 FVAF[%] for multimodal fusion (Ortega et al., 2021). Relative to that work, EEGForceMap is distinctive in operating from EEG alone, in centering the premotor-parietal network, and in making the preprocessing and feature-engineering stages the primary source of performance gains rather than multimodal fusion.

The term “map” in EEGForceMap can be misread if it is conflated with other EEG mapping traditions. EEGForceMap does not denote an anatomically grounded FEM field map, a scalp potential interpolation method, or a latent 3D manifold trajectory. By contrast, “Autoencoding Coordinate Sequences from Psychophysiologic Signals” maps multimodal psychophysiologic streams into trackable 3D coordinate sequences interpreted as displacement from a learned resting-state manifold (Hutcheson et al., 8 Oct 2025). “Automated modeling of brain bioelectric activity within the 3D Slicer environment” focuses on patient-specific modeling of brain bioelectric fields for iEEG/ECoG forward problems on deformed postoperative anatomy (Safdar et al., 2022). “Zeffiro user interface for electromagnetic brain imaging” provides GPU accelerated FEM forward and inverse computations in Matlab, including lead-field generation and hierarchical Bayesian inversion (He et al., 2018). “FEM-based Scalp-to-Cortex data mapping via the solution of the Cauchy problem” addresses scalp-to-cortex potential continuation through a regularized Cauchy problem for Laplace’s equation (Koshev et al., 2019). EEGForceMap is therefore more accurately classified as a force-decoding representation pipeline than as a physical field-mapping framework.

The paper’s limitations are explicit and consequential. The dataset is based on motor execution, not motor imagery or attempted movement, so transfer to stroke rehabilitation or imagery-driven BCI use is unresolved. The regressors make larger errors during rest periods, and the authors suggest that a future classifier-decoder ensemble could detect task-versus-rest state before force regression. Additional limitations, either stated or directly inferable from the reported protocol, include the small sample size of 12 subjects, the partly manual and subject-dependent covariance-based channel/component selection, and the absence of detailed implementation parameters for spectral estimation and ICA component thresholds. This suggests that the strongest contribution of EEGForceMap is conceptual and methodological: it demonstrates that a neurophysiology-guided EEG representation, especially one built around ERDS in theta and mu bands from a premotor-parietal channel subset, can materially outperform prior EEG-only and even some multimodal reference decoders for continuous grasp-force estimation (Dangi et al., 11 Aug 2025).