Bayesian Complete-Pooling in Cross-Subject Classification for Motor Imagery Electroencephalogram

arXiv cs.LG Papers

Summary

This paper presents a large-scale study comparing Bayesian complete-pooling models to frequentist baselines for cross-subject motor imagery EEG classification, finding that Bayesian methods improve reliability modestly but at a higher computational cost, with limited practical benefit.

arXiv:2607.22980v1 Announce Type: new Abstract: Brain-computer interfaces (BCIs) have long sought calibration-free operation, but classifiers are typically benchmarked by discrimination alone, blind to whether predicted probabilities are well calibrated - a meaningful gap given nonstationary electroencephalogram (EEG) signals and the risk of overconfident point-estimate classifiers under distribution shift. We conducted a large-scale study contrasting Bayesian complete-pooling models against frequentist baselines for cross-subject, left-hand versus right-hand motor imagery EEG classification across 20 datasets. Six frequentist pipelines were each paired with an analogous Bayesian pipeline sharing identical feature engineering, fit via Markov chain Monte Carlo posterior sampling. Our primary metric was the Brier score, decomposed into reliability and resolution, alongside AUROC for discrimination and Shannon entropy for sharpness. Each metric was analyzed via random-effects meta-analysis (REML, Knapp-Hartung adjustment), verified by leave-one-out influence analysis. Bayesian complete-pooling produced statistically but not practically significant improvements in reliability and increases in predictive uncertainty (lower sharpness); Brier score, resolution, and discrimination showed no significant differences. Between-study heterogeneity was low across all metrics, though the reliability result was sensitive to leave-one-out removal. We additionally profiled computational cost, finding that Bayesian pipelines consumed roughly thirteen times more energy than their frequentist counterparts, a cost that remains modest relative to common household appliances. These results suggest that Bayesian complete-pooling alone offers limited practical benefit for cross-subject motor imagery classification, and that partial-pooling across subjects and sessions is a more promising direction for future work.
Original Article
View Cached Full Text

Cached at: 07/28/26, 06:23 AM

# Bayesian Complete-Pooling in Cross-Subject Classification for Motor Imagery Electroencephalogram
Source: [https://arxiv.org/html/2607.22980](https://arxiv.org/html/2607.22980)
###### Abstract

Brain\-computer interfaces \(BCIs\) have long sought calibration\-free operation, but classifiers are typically benchmarked by discrimination alone, blind to whether predicted probabilities are well calibrated — a meaningful gap given nonstationary electroencephalogram \(EEG\) signals and the risk of overconfident point\-estimate classifiers under distribution shift\. We conducted a large\-scale study contrasting Bayesian complete\-pooling models against frequentist baselines for cross\-subject, left\-hand versus right\-hand motor imagery EEG classification across 20 datasets\. Six frequentist pipelines were each paired with an analogous Bayesian pipeline sharing identical feature engineering, fit via Markov chain Monte Carlo posterior sampling\. Our primary metric was the Brier score, decomposed into reliability and resolution, alongside AUROC for discrimination and Shannon entropy for sharpness\. Each metric was analyzed via random\-effects meta\-analysis \(REML, Knapp\-Hartung adjustment\), verified by leave\-one\-out influence analysis\. Bayesian complete\-pooling produced statistically but not practically significant improvements in reliability and increases in predictive uncertainty \(lower sharpness\); Brier score, resolution, and discrimination showed no significant differences\. Between\-study heterogeneity was low across all metrics, though the reliability result was sensitive to leave\-one\-out removal\. We additionally profiled computational cost, finding that Bayesian pipelines consumed roughly thirteen times more energy than their frequentist counterparts, a cost that remains modest relative to common household appliances\. These results suggest that Bayesian complete\-pooling alone offers limited practical benefit for cross\-subject motor imagery classification, and that partial\-pooling across subjects and sessions is a more promising direction for future work\.

Keywords:brain\-computer interface, motor imagery, Bayesian machine learning, probabilistic calibration, meta\-analysis, electroencephalogram

## 1Introduction

A long\-standing goal of brain\-computer interface \(BCI\) research has been to develop calibration\-free devices\. Whether due to fatigue, hunger, or a drying out of sensor gel, the signal that comes out of current noninvasive brain measurement techniques changes with time\. Worse, while human brains share a generic pattern of functionality, individuals vary significantly\. The oldest, most reliable strategy to deal with session\-to\-session variation is training\. Marginally better than months of training, we can simply recalibrate the BCI for each new usage\. While somewhat tiresome for the subject, this can require as little as 10 minutes\.

Small\-scale studies based on maximum a posteriori \(MAP\) estimation for EEG classification have shown that partial\-pooling across subjects and sessions is a promising direction of research\[[23](https://arxiv.org/html/2607.22980#bib.bib8)\]\. Even for within\-session BCIs, fully Bayesian multilevel/hierarchical modes have been shown to improve discrimination\[[49](https://arxiv.org/html/2607.22980#bib.bib9)\]\. Complementary to these results, large\-scale benchmarks of established \(no\-pooling and complete\-pooling\) BCI models based on maximum likelihood estimation \(MLE\) have uncovered ordered rankings of classifiers by name and by category in terms of discrimination\[[7](https://arxiv.org/html/2607.22980#bib.bib21)\]\.

Towards a large\-scale experiment of the effects of partial\-pooling on BCI prediction performance, we conducted an empirical study of Bayesian complete\-pooling pipelines for left hand and right hand motor imagery electroencephalogram \(MI\-EEG\) classification contrasted with frequentist baselines\. Discrimination metrics such as area under the receiver operating characteristic \(AUROC\) curve are by far the most common way to rank BCIs\. However, discrimination is independent of calibration\[[36](https://arxiv.org/html/2607.22980#bib.bib22)\], and can become badly miscalibrated under the distribution shift of nonstationary EEG\.

In addition to discrimination, we contrasted frequentist pipeline controls to Bayesian pipeline treatments in terms of calibration\. We used a proper scoring rule as our primary metric of comparison\. As secondary metrics we analyzed calibration/reliability, resolution, discrimination, and sharpness\[[46](https://arxiv.org/html/2607.22980#bib.bib10)\]\. Our experiment was a large\-scale study that pooled binary classification results across 20 left hand and right hand MI\-EEG datasets\. The summary is that with Bayesian complete\-pooling we found statistically significant differences in reliability and sharpness yet results were not practically significant\.

In[Section2](https://arxiv.org/html/2607.22980#S2)we briefly overview preliminary concepts of Bayesian ML pointing to key characteristics that may benefit BCIs and recent transfer learning research\. Additionally, we delineate the metrics of our experiment and why they are relevant for building trustworthy ML\. Since our benchmark analysis includes discussions of effect sizes, uncertainty, and between\-study heterogeneity measurements in addition topp\-values, we explain these too\.[Section3](https://arxiv.org/html/2607.22980#S3)outlines our experiment design before[Section4](https://arxiv.org/html/2607.22980#S4)revealing the results and[Section5](https://arxiv.org/html/2607.22980#S5)where we interpret the outcomes\. We finish by describing how our work can support future research\.

## 2Background

### 2\.1Model Building

In supervised learning, a taskTTlearns a mappingfffrom inputs𝒙∈𝒳\\bm\{x\}\\in\\mathcal\{X\}to outputsy∈𝒴y\\in\\mathcal\{Y\}\. The inputs𝒙\\bm\{x\}are often a fixed dimensional vector𝒙∈ℝD\\bm\{x\}\\in\\mathbb\{R\}^\{D\}whereDDis the dimensionality of the vector\. An experienceEEis given in the form of a set ofNNinput\-output pairs𝒟=\{\(𝒙n,yn\)\}n=1N\\mathcal\{D\}=\\\{\(\\bm\{x\}\_\{n\},y\_\{n\}\)\\\}^\{N\}\_\{n=1\}, known as the training set, whereNNis the sample size\. In classification problems, the output space is a set ofCCunordered and mutually exclusive labels known as classes𝒴=\{𝒞1,𝒞2,…,𝒞C\}\\mathcal\{Y\}=\\\{\\mathcal\{C\}\_\{1\},\\mathcal\{C\}\_\{2\},\\ldots,\\mathcal\{C\}\_\{C\}\\\}\. The problem of training is often to find a set of parameters that solves the maximum likelihood estimation of the model as in[Equation1](https://arxiv.org/html/2607.22980#S2.E1)\[[34](https://arxiv.org/html/2607.22980#bib.bib3)\]\.

p​\(y\|𝒙,𝜽^\)such that𝜽^=argmax𝜽​∑n=1Nlog⁡p​\(yn\|𝒙n,𝜽\)\.p\(y\\,\|\\,\\bm\{x\},\\hat\{\\bm\{\\theta\}\}\)\\quad\\text\{such that\}\\quad\\hat\{\\bm\{\\theta\}\}=\\operatorname\*\{argmax\}\_\{\\bm\{\\theta\}\}\\sum\_\{n=1\}^\{N\}\\log p\(y\_\{n\}\\,\|\\,\\bm\{x\}\_\{n\},\\bm\{\\theta\}\)\.\(1\)
Bayesian ML lets us model a probability distribution over weights allowing us to represent uncertainty in our prediction\. For a training set𝒟\\mathcal\{D\}, outputyy, input𝒙\\bm\{x\}, and model weights𝜽\\bm\{\\theta\}, we have the Bayesian treatment of ML found in[Equation2](https://arxiv.org/html/2607.22980#S2.E2)\[[3](https://arxiv.org/html/2607.22980#bib.bib5)\]\. However, there is a major drawback with the Bayesian framework which involves the integration over the space of parameters\[[2](https://arxiv.org/html/2607.22980#bib.bib6)\]\. This is familiar from introductory calculus: It is easier to take a derivative, e\.g\. gradient descent, than it is to find an integral\. Given a limited compute budget and ample training data, it is often better to use point estimation\[[2](https://arxiv.org/html/2607.22980#bib.bib6)\], though, typically Bayesian ML corresponds to better generalization\[[35](https://arxiv.org/html/2607.22980#bib.bib4)\]\.

p​\(y\|𝒙,𝒟\)=∫p​\(y\|𝒙,𝜽\)​p​\(𝜽\|𝒟\)​d𝜽wherep​\(𝜽\|𝒟\)∝p​\(𝜽\)​p​\(𝒟\|𝜽\)\.p\(y\\,\|\\,\\bm\{x\},\\mathcal\{D\}\)=\\int p\(y\\,\|\\,\\bm\{x\},\\bm\{\\theta\}\)p\(\\bm\{\\theta\}\\,\|\\,\\mathcal\{D\}\)\\mathrm\{d\}\\bm\{\\theta\}\\quad\\text\{where\}\\quad p\(\\bm\{\\theta\}\\,\|\\,\\mathcal\{D\}\)\\propto p\(\\bm\{\\theta\}\)p\(\\mathcal\{D\}\\,\|\\,\\bm\{\\theta\}\)\.\(2\)
Uncertainty in parameters is naturally and coherently modeled with Bayesian ML via the posterior probability distributionp​\(𝜽\|𝒟\)p\(\\bm\{\\theta\}\\,\|\\,\\mathcal\{D\}\)\. Furthermore, this uncertainty estimation generalizes to multilevel/hierarchical models over datasets with group\-level structure\[[32](https://arxiv.org/html/2607.22980#bib.bib11)\]\. Indeed, multilevel modeling is a frequent motivation for using Bayesian methods\. For BCIs, it is intuitive to see how multilevel modeling can be applied across subjects and sessions\. Given the population\-level nonstationarity of EEG signals, Bayesian ML offers useful estimation capabilities through principled uncertainty quantification approaches\.

### 2\.2Prediction Verification

When building predictors our goal is to maximize sharpness subject to reliability\[[14](https://arxiv.org/html/2607.22980#bib.bib23)\]\. This concept intuitively refers to the accuracy and precision of a bullseye diagram where predictions may be tight/scattered and centered/off\-centered\. Formally, lety^m\\hat\{y\}\_\{m\}be a prediction which can take on any of theMMvaluesy^1,y^2,…,y^M\\hat\{y\}\_\{1\},\\hat\{y\}\_\{2\},\\ldots,\\hat\{y\}\_\{M\}, and letyny\_\{n\}be the corresponding target which may be any of theNNvaluesy1,y2,…,yNy\_\{1\},y\_\{2\},\\ldots,y\_\{N\}\. Accuracy generally refers to the average correspondence between individual predictions and the events they predict\. Reliability, resolution, discrimination, and sharpness are components of accuracy\[[46](https://arxiv.org/html/2607.22980#bib.bib10)\]\.

Reliability, or calibration, corresponds to the average observationypy\_\{p\}given a predictiony^r\\hat\{y\}\_\{r\}summarizing the conditional distributionp​\(yp\|y^r\)p\(y\_\{p\}\\,\|\\,\\hat\{y\}\_\{r\}\)\. Resolution is also concerned withp​\(yp\|y^r\)p\(y\_\{p\}\\,\|\\,\\hat\{y\}\_\{r\}\)and measures the difference in conditional averages of an observationypy\_\{p\}for different predictionsy^r\\hat\{y\}\_\{r\}andy^s\\hat\{y\}\_\{s\}\. Discrimination is the converse of resolution, in that it pertains to differences between the conditional averages of a predictiony^r\\hat\{y\}\_\{r\}conditional on different targetsypy\_\{p\}andyqy\_\{q\}\. Sharpness is an attribute of predictions alone without regard to their corresponding observations, characterizing the unconditional distribution of a predictionp​\(y^r\)p\(\\hat\{y\}\_\{r\}\)\[[46](https://arxiv.org/html/2607.22980#bib.bib10)\]\.

Prediction verification is easiest to understand with nonprobabilistic predictions of discrete predictands, such as accuracy calculated as the proportion of true positives and true negatives over all predictions\. Verification of probabilistic predictions is more nuanced\. Since nonprobabilistic predictions contain no expression of uncertainty, it is clear whether an individual prediction is correct or not\. The Brier score \(BS\) is a common probabilistic prediction that is essentially the mean squared error of the predictions\. BS averages the squared differences between pairs of predictions and observations shown in[Equation3](https://arxiv.org/html/2607.22980#S2.E3), whereKKis the number of pairs\[[46](https://arxiv.org/html/2607.22980#bib.bib10)\]\.

BS=1K​∑k=1K\(yk−y^k\)2\.\\mathrm\{BS\}=\\frac\{1\}\{K\}\\sum\_\{k=1\}^\{K\}\(y\_\{k\}\-\\hat\{y\}\_\{k\}\)^\{2\}\.\(3\)
Prediction verification has perhaps been most developed by the atmospheric sciences, which have taken a calibration\-first approach to research\[[46](https://arxiv.org/html/2607.22980#bib.bib10)\]\. On the other hand, machine learning has largely developed through a discrimination\-first approach\[[11](https://arxiv.org/html/2607.22980#bib.bib12)\]\. In the last several years, though, ML researchers have increasingly built calibration\-aware models\[[17](https://arxiv.org/html/2607.22980#bib.bib24)\]\. Bayesian ML generally offers better calibrated models because they average over parameter space rather than optimizing point estimates which may result in overconfident \(overly sharp\) estimates\[[35](https://arxiv.org/html/2607.22980#bib.bib4)\]\. This has motivated the use of Bayesian methods in high\-risk scenarios for the purpose of trustworthy ML\.

### 2\.3Evidence Synthesis

The goal of meta\-analyses is to quantitatively synthesize evidence from multiple studies\[[19](https://arxiv.org/html/2607.22980#bib.bib13)\]\. Meta\-analyses of effect sizes are generally preferred since they address the issue of interest \(practical significance\) rather than the null hypothesis \(statistical significance\)\[[4](https://arxiv.org/html/2607.22980#bib.bib14)\]\. An effect size captures the direction and magnitude of the relationship between two entities\. For some studykk, we denote the true effectθk\\theta\_\{k\}, the observed effectθ^k\\hat\{\\theta\}\_\{k\}, and the sampling errorϵk\\epsilon\_\{k\}, such that we have[Equation4](https://arxiv.org/html/2607.22980#S2.E4)\. When pooling results, meta\-analyses weight studies by their standard error \(SE\), defined by the standard deviation of the sampling distribution\[[19](https://arxiv.org/html/2607.22980#bib.bib13)\]\.

θ^k=θk\+ϵk\.\\hat\{\\theta\}\_\{k\}=\\theta\_\{k\}\+\\epsilon\_\{k\}\.\(4\)
[Equation4](https://arxiv.org/html/2607.22980#S2.E4)is known as the random\-effects model which assumes that there is not only one true effect size \(fixed\-effect model\) but a distribution of true effect sizes\. The goal of the random\-effects model is to estimate the mean of the distribution of true effects\. It stipulates that there is a second source of errorζk\\zeta\_\{k\}introduced by the fact that the true effect sizeθk\\theta\_\{k\}is only part of an over\-arching distribution of true effect sizes with meanμ\\mugiving us[Equation5](https://arxiv.org/html/2607.22980#S2.E5)\. A crucial assumption of the random\-effects model is that the size ofζk\\zeta\_\{k\}is independent ofkk\. This is the exchangeability assumption of the random\-effects model\[[19](https://arxiv.org/html/2607.22980#bib.bib13)\]\.

θ^k=μ\+ζk\+ϵk\.\\hat\{\\theta\}\_\{k\}=\\mu\+\\zeta\_\{k\}\+\\epsilon\_\{k\}\.\(5\)
The challenge associated with random\-effects models is that we have to take the errorζk\\zeta\_\{k\}into account\. To do this, we have to estimate the varianceτ2\\tau^\{2\}of the distribution of true effect sizes\. It is ongoing research to find which estimator ofτ2\\tau^\{2\}performs best for different kinds of data\. The DerSimionian\-Laird estimator is based on closed\-form expressions, while restricted maximum likelihood finds the optimal value ofτ2\\tau^\{2\}through an iterative algorithm\. While significance tests of the pooled effect usually assume a normal distribution known as Wald tests, the Knapp\-Hartung method is based on att\-distribution and used when the number of studies is small\[[19](https://arxiv.org/html/2607.22980#bib.bib13)\]\.

τ2\\tau^\{2\}is used to estimate the extent to which true effect sizes vary within a meta\-analysis known as between\-study heterogeneity\. Extreme heterogeneity can mean that the studies have nothing in common, and that it makes no sense to interpret the pooled effect\. The difficulty of quantifying between\-study heterogeneity is to identify how much of the variation can be attributed to the sampling error, and how much to true effect size differences\. TheI2I^\{2\}statistic is the percentage of variability in the effect sizes that is not caused by sampling error\. Prediction intervals \(PIs\) give us a range into which we expect the effects of future studies to fall based on present evidence\[[19](https://arxiv.org/html/2607.22980#bib.bib13)\]\.

Between\-study heterogeneity can be caused by one or more studies with extreme effect sizes\. When conducting a meta\-analysis, we want to know if the pooled effect estimate is robust, meaning that it does not depend heavily on one single study\. Such studies are called influential cases\. For a meta\-analysis withKKstudies, the leave one out \(LOO\) method recalculates the results of the meta\-analysisKKtimes, each time leaving out one study\. Influence diagnostics allow us to detect the studies which influence the overall estimate of our meta\-analysis the most, and let us assess if this large influence distorts our pooled effect\[[19](https://arxiv.org/html/2607.22980#bib.bib13)\]\.

## 3Methods

### 3\.1Datasets

We used the Mother of all BCI Benchmarks \(MOABB\) v1\.5 to access all left hand and right hand MI\-EEG datasets used in our experiment\[[22](https://arxiv.org/html/2607.22980#bib.bib20)\]\. For the purpose of unbiased data selection, we used all 20 available datasets under the conditions that, first, a dataset had at least 8 subjects and, second, it was not a subset of another dataset\. There were 10 datasets that did not meet these criteria\.[Table1](https://arxiv.org/html/2607.22980#S3.T1)shows an overview of the datasets used for our evaluation\. Certain datasets were collected for the purpose of recording high\-quality MI\-EEG signals\. Others were acquired such that specific experimental interventions could be tested\.

Table 1:Sample characteristics of the 20 datasets\.
### 3\.2Paradigm

Our experiment was concerned with left hand and right hand motor imagery and filtered any other classes from the datasets\. Additional filtering to the raw datasets was the inclusion of only sessions 1–6 for Stieger2021 and the inclusion of only subjects 13–20 with Zhou2020\. In the case of Stieger2021, this session reduction was done in response to GPU memory bottlenecks specifically in our Bayesian neural networks \(BNNs\)\. Regarding Zhou2020, only A\-subjects whose electrode information could be parsed were included\.[Figure1](https://arxiv.org/html/2607.22980#S3.F1)visualizes our datasets after signal preprocessing was completed\.

Sensors were filtered for those belonging to the sensorimotor cortex of the international 10\-05 montage\. This subset of 33 channels spans the FC, FCC, C, CCP, and CP indexes\. The primary motivation for filtering electrodes was dimensionality reduction in response to device memory exhaustion during training\. However, this simultaneously reduced extraneous noise from other brain regions\. Additional dimensionality reduction we used was to downsample the frequency of all datasets to 128 Hz\. Combinations of our largest datasets and pipelines exhausted available GPU memory without downsampling, though, this also homogenized our datasets\.

![Refer to caption](https://arxiv.org/html/2607.22980v1/assets/montages.png)\(a\)Intersection of electrodes with the international 10\-05 montage covering the sensorimotor cortex\.
![Refer to caption](https://arxiv.org/html/2607.22980v1/assets/bubble-plot.png)\(b\)Scale comparison in terms of number of sessions and total trials\.

Figure 1:The MI\-EEG datasets after preprocessing\.
### 3\.3Pipelines

Our work can be seen as an extension of previous large\-scale BCI benchmarking research byChevallieret al\.\[[7](https://arxiv.org/html/2607.22980#bib.bib21)\]\. In their study, three categories of BCI classifiers were identified: Raw, Riemannian, and Deep Learning\. Respectively, these are pipelines based on raw signals, algorithms relying on Riemannian geometric statistics, and deep learning approaches\. For their benchmark, 19 classifiers evaluated across 10 left/right hand MI\-EEG datasets were used\. Binary classification performance was ranked by the area under the receiver operating characteristic curve \(AUROC\) from within\-session evaluation\.

For our benchmark, we selected two of the top performing pipelines from each category ofChevallieret al\.\[[7](https://arxiv.org/html/2607.22980#bib.bib21)\]\. From within both the Raw and Deep Learning categories we picked the first and second best performing pipelines out of nine and seven choices, respectively\. With the Riemannian category, we chose the third and fourth highest performing pipelines out of seven options\. We did not select either the first or second highest performing Riemannian pipelines because they did not map cleanly to a Bayesian ML treatment\.[Table2](https://arxiv.org/html/2607.22980#S3.T2)shows the pipelines we selected that were drawn fromChevallieret al\.\[[7](https://arxiv.org/html/2607.22980#bib.bib21)\]\.

Table 2:Baseline pipelines used in our experiment ordered by mean AUROC across 10 left/right hand MI\-EEG datasets fromChevallieret al\.\[[7](https://arxiv.org/html/2607.22980#bib.bib21)\]: BNCI2014\_001, BNCI2014\_004, Cho2017, GrosseWentrup2009, Lee2019\_MI, PhysionetMI, Schirrmeister2017, Shin2017A, Weibo2014, and Zhou2016\. From this list, only Zhou2016 \(4 subjects\) was not used by our evaluation\. The original rank among all 19 frequentist pipelines is shown\. Also the paired Bayesian pipelines from our experiment are displayed\.#### 3\.3\.1Frequentist Baselines

Both the CSP and TS based classifiers depended on the covariance matrix of EEG channels across time points of each trial as input features\. For covariance estimation, we used the oracle approximating shrinkage \(OAS\) estimator\. Although its closed\-form solution is sensitive to outliers, the low latency of OAS is ideal for real\-time BCIs\. With CSP we used the Euclidean metric for estimation of mean covariance matrices\. Additionally, we decomposed 6 components along orthogonal discriminative axes\. For TS we applied the Riemannian metric for reference matrix estimation and tangent space mapping\.

For our LDA classifiers we used the singular value decomposition \(SVD\) to fit Gaussian densities to left/right hand trials\. This method avoids directly computing the covariance matrix which is numerically unstable\. Our logistic regression configuration used the Limited\-memory Broyden\-Fletcher\-Goldfarb\-Shanno \(L\-BFGS\) solver with default regularization\. This is a second\-order optimization algorithm, but uses limited memory to approximate the Hessian matrix\. To accommodate convergence with complex datasets, we set the maximum number of iterations to 1000\.

SVMs are generally not interpreted with probability calibration since their decision boundary is geometrically optimized\. A common post\-hoc method for measuring probabilistic metrics with SVMs is to use Platt scaling involving internal calibration cross\-validation\[[9](https://arxiv.org/html/2607.22980#bib.bib1)\]\. We used Platt scaling with 5\-fold cross\-validation when fitting our SVMs\. Given CSP features, we trained SVMs with radial basis function \(RBF\) kernels\. In contrast, for TS pipelines we applied a linear kernel\. Geometrically, a linear decision boundary in the tangent space of the reference matrix approximates the best fit geodesic separating both classes\.

To train ShallowConvNet \(SCNN\) and DeepConvNet \(DCNN\)\[[41](https://arxiv.org/html/2607.22980#bib.bib37)\], we used a uniform evaluation loop for both models\. The loss function used was cross\-entropy\. The gradient update algorithm we used was AdamW\. We configured the learning rate and weight decay to be10−310^\{\-3\}and10−210^\{\-2\}, respectively\. A maximum of 300 epochs with batch size of 64 and 80–20 validation split was used\. Early stopping based on validation loss was applied with 150 epochs of patience and a10−410^\{\-4\}threshold for improvement\. A learning rate scheduler was also used, with 50 epochs of patience before halving the learning rate, but not below10−610^\{\-6\}, if the validation loss did not decrease\.

#### 3\.3\.2Bayesian Models

The pipelines we selected fromChevallieret al\.\[[7](https://arxiv.org/html/2607.22980#bib.bib21)\]were the control group of our experiment\. Each of these pipelines was paired with a Bayesian ML pipeline from our treatment group\. In all pairs, feature engineering remained fixed\. The pairwise contrast was frequentist point estimate optimizationp​\(y\|x,θ^\)p\(y\\,\|\\,x,\\hat\{\\theta\}\)versus Bayesian parameter space modeling of the posterior weightsp​\(θ\|𝒟\)p\(\\theta\\,\|\\,\\mathcal\{D\}\)with Bayesian model averaging for predictionp​\(y\|x,𝒟\)=∫p​\(y\|x,θ\)​p​\(θ\|𝒟\)​dθp\(y\\,\|\\,x,\\mathcal\{D\}\)=\\int p\(y\\,\|\\,x,\\theta\)p\(\\theta\\,\|\\,\\mathcal\{D\}\)\\mathrm\{d\}\\theta\. Indeed, all feature extraction was deterministic, whereas probabilistic classification was our experimental intervention\.

Centering and scaling input features allowed our non\-hierarchical prior specifications to generalize across datasets with heterogeneous feature scales\. For all Bayesian pipelines, we appliedzz\-score normalization to inputs after feature extraction and before classification\. For the purpose of maintaining consistent feature engineering across all paired ML models,zz\-score transformation was also applied to features of frequentist pipelines\. This experiment design localized the effect of Bayesian non\-hierarchical prediction performance to classification\.

We used Markov chain Monte Carlo \(MCMC\) based posterior sampling to fit our Bayesian pipelines\. Specifically, No\-U\-Turn Sampling \(NUTS\) was applied\. NUTS is an extension of Hamiltonian Monte Carlo \(HMC\) gradient\-informed posterior sampling\. For both warmup draws and tuning samples of MCMC we used 1000 iterations\. To ensure robust sampling, we configured the target acceptance rate to be 0\.95\. The cost of a higher acceptance rate is often smaller steps per iteration and slower convergence, while the benefit is more accurate posterior approximations\. All Bayesian pipelines were fit with 4 chains to validate the correctness of convergence\.

LDA is a type of Gaussian mixture model \(GMM\) that is supervised\[[20](https://arxiv.org/html/2607.22980#bib.bib2)\]\. It was the only generative classifier in our experiment\. To fit Bayesian LDA \(BLDA\), we modeled the likelihood of observations given labelsp​\(x\|y\)p\(x\\,\|\\,y\)via a GMM and used Bayes theorem to estimate the probability of labels given observationsp​\(y\|x\)p\(y\\,\|\\,x\)\. As a prior on the binary labels we used a weakly informativeπ∼Beta⁡\(2,2\)\\pi\\sim\\operatorname\{Beta\}\(2,2\)prior distribution\. For the location priors and shared scale prior we used standard normal distributions𝝁0,𝝁1∼𝒩​\(0,1\)\\bm\{\\mu\}\_\{0\},\\bm\{\\mu\}\_\{1\}\\sim\\mathcal\{N\}\(0,1\)and a half\-normal distribution𝝈∼ℋ​𝒩​\(1\)\\bm\{\\sigma\}\\sim\\mathcal\{HN\}\(1\), respectively\. See[Figure2](https://arxiv.org/html/2607.22980#S3.F2)for the probabilistic graphical model of our BLDA\.

![Refer to caption](https://arxiv.org/html/2607.22980v1/assets/blda.png)Figure 2:The PGM of BLDA wherePPis the number of trials andQQis the number of electrodes\.Bayesian logistic regression \(BLR\) appeared in one of our Riemannian pipelines and in both of our Deep Learning classifiers\. A standard normal prior distribution was used for the mean and variance of the weights and bias parameters𝒘,b∼𝒩​\(0,1\)\\bm\{w\},b\\sim\\mathcal\{N\}\(0,1\)\. With the variance of the weights, though, we relied on LeCun initialization such thatσw=1D\\sigma\_\{w\}=\\frac\{1\}\{\\sqrt\{D\}\}whereDDwas the fan\-in\[[35](https://arxiv.org/html/2607.22980#bib.bib4)\]\. Our Bayesian neural networks \(BNNs\) were Bayesian last layer \(BLL\) models\. In two stages we, first, trained the SCNN and DCNN architectures then, second, dropped the deterministic linear classifier and fit a probabilistic BLR predictor as a replacement\.

SVMs belong to a category of kernel methods\[[3](https://arxiv.org/html/2607.22980#bib.bib5)\]\. As a Bayesian analog to SVMs we used Gaussian processes classifiers \(GPCs\)\. The MAP of GPCs with probit loss yields highly similar solutions to SVMs with its characteristic hinge loss function\[[39](https://arxiv.org/html/2607.22980#bib.bib7)\]\. For CSP features with an RBF kernel, we used a half\-normal amplitude priorη∼ℋ​𝒩​\(1\)\\eta\\sim\\mathcal\{HN\}\(1\)and a log\-normal lengthscale priorℓ∼ℒ​𝒩​\(0,0\.5\)\\ell\\sim\\mathcal\{LN\}\(0,0\.5\)\. Given TS features and a linear kernel, we applied a half\-normal amplitude prior distributionη∼ℋ​𝒩​\(1\)\\eta\\sim\\mathcal\{HN\}\(1\)\. To fit our GPCs we used the deterministic training conditional \(DTC\) method for sparse approximation with 100 inducing points found viakk\-means\.

### 3\.4Evaluation

EEG signals have their lowest signal\-to\-noise ratio \(SNR\) when trials across sessions and subjects are pooled\. We fit Bayesian non\-hierarchical models for cross\-subject classification to test the effect of our treatment in this scenario\. Our use of cross\-subject evaluation as opposed to within\-session evaluation or cross\-session \(within\-subject\) evaluation provided an environment where the benefits of uncertainty quantification in pattern recognition from Bayesian ML would be most visible\. Our experiment used a combination of leave one subject out \(LOSO\) and 10\-fold cross\-validation: LOSO if the number of subjects in a dataset was less than 10, otherwise 10\-fold CV\.

Our primary metric for pairwise comparisons was Brier score\. As a proper scoring rule, BS awards the true beliefs of a predictor, avoiding hedging, or gaming, of predictions\[[46](https://arxiv.org/html/2607.22980#bib.bib10)\]\. BS can be decomposed into reliability and resolution\[[33](https://arxiv.org/html/2607.22980#bib.bib25)\], and we used these components directly as secondary metrics to measure prediction performance\. To decompose BS used the CORP \(Consistent, Optimally binned, and Reproducible\) framework\[[10](https://arxiv.org/html/2607.22980#bib.bib26)\]\. We also measured discrimination and sharpness for prediction verification\. For discrimination we measured AUROC, and for sharpness we measured Shannon entropy\[[30](https://arxiv.org/html/2607.22980#bib.bib15)\]\.

### 3\.5Analysis

The 20 datasets we used had varying hardware, subjects, and samples meaning they were heterogeneous and could not be merged into one large dataset\. For each dataset, we computed mean LOSO or 10\-fold CV scores for all of our pipelines, before pairwise differences\. We did this for each of our 5 metrics\. This generated 6 data points per dataset per metric\. Effect sizes and SEs were computed for matched groups\[[4](https://arxiv.org/html/2607.22980#bib.bib14)\]\. For each metric, we fit a random\-effects meta\-analysis with a restricted maximum likelihood \(REML\) estimator and Knapp\-Hartung adjustment\. We also performed an influence analysis to verify the robustness of our results\.

## 4Results

### 4\.1Posterior Sampling

Table 3:Convergence diagnostics from posterior sampling totaled across 196 runs for each pipeline\.For our Bayesian pipelines, we measured posterior sampling diagnostics to verify the quality of our estimates: bulk and tail effect sample size \(ESS\), and potential scale reduction factor \(R^\\hat\{R\}\)\. ESS takes autocorrelation of our draws into account and provides the number of draws we would have if our sample was iid\.R^\\hat\{R\}estimates convergence of separate chains estimating their similarity at the end of sampling\. The recommended threshold for ESS bulk/tail is above 400, and values ofR^≤1\.01\\hat\{R\}\\leq 1\.01are considered safe\[[31](https://arxiv.org/html/2607.22980#bib.bib16)\]\. Grouping by pipelines, we aggregated worst case diagnostics across datasets and cross\-validation in[Table3](https://arxiv.org/html/2607.22980#S4.T3)\. Our results were validated for continued analysis\.

### 4\.2Meta\-Analysis

The pooled effect of accuracy \(Brier score\) was not significant\. Additionally, both the pooled effect of resolution and discrimination \(AUROC\) were not significant\. However, the aggregated effect size of reliability and sharpness were statistically significant\. For reliability, the direction of the effect was negative meaning that Bayesian complete\-pooling reduced calibration error\. Furthermore, the direction for sharpness \(Shannon entropy\) was positive telling us that Bayesian complete\-pooling increased uncertainty in predictions\. At the same time, though, neither of these statistically significant effects were practically significant\. See[Table4](https://arxiv.org/html/2607.22980#S4.T4)for details\.

Table 4:The pooled effects for all metrics\.The variance components of between\-study heterogeneity measured by our meta\-analyses suggested that our results across left/right hand MI\-EEG datasets were homogenous across datasets\. That is, according to the rule of thumb for how heterogeneity is interpreted throughI2I^\{2\}\(25% = low, 50% = moderate, 75% = substantial\)\[[21](https://arxiv.org/html/2607.22980#bib.bib27)\]\. These outcomes validate the 95% PIs that we computed\. PIs are generally wider than the 95% confidence intervals \(CIs\) of the pooled effect estimate\. Statistically significant estimates with PIs that cross zero are less robust\. This is what we see from reliability, though, the pooled effect of sharpness remains positive in[Table5](https://arxiv.org/html/2607.22980#S4.T5)\.

Table 5:The between\-study heterogeneity\.For each iteration of LOO influence analysis we found that results for Brier score, resolution, and AUROC yielded consistent conclusions: null difference for the pooled estimate and low between\-study heterogeneity\. However, for reliability,11/20=55%11/20=55\\%of iterations yielded null difference pooled effects suggesting our above results were sensitive\. In the remaining 45% of statistically significant cases, pooled effects were not practically significant, and in all cases heterogeneity remained low\. For Shannon entropy, just 1/20 iterations of LOO yielded a null difference pooled effect\. As before, though, results had low practical significance and heterogeneity\.

![Refer to caption](https://arxiv.org/html/2607.22980v1/assets/brier-forest.png)Figure 3:The Brier score forest plot\.[Figure3](https://arxiv.org/html/2607.22980#S4.F3)shows the forest plot from our meta\-analysis of Brier score\. Additionally,[Figure4](https://arxiv.org/html/2607.22980#S4.F4)and[Figure5](https://arxiv.org/html/2607.22980#S4.F5)show the forest plots for reliability and Shannon entropy, respectively\. For each study in a forest plot, the size of the square corresponds to the precision of its point estimate, driving the weight of the study when estimating the pooled effect\. The brackets represent the 95% confidence intervals of a study\. Likewise, the size of the diamond corresponds to the precision of the pooled effect estimate\. However, its brackets represent the 95% prediction intervals, denoting where we would expect to find the effect size of an unseen study\[[19](https://arxiv.org/html/2607.22980#bib.bib13)\]\.

![Refer to caption](https://arxiv.org/html/2607.22980v1/assets/reliability-forest.png)\(a\)Per\-dataset effects, pooled estimate, and prediction interval\.
![Refer to caption](https://arxiv.org/html/2607.22980v1/assets/reliability-baujat.png)\(b\)Contribution to heterogeneity versus influence on the pooled result\.

Figure 4:Forest and Baujat plots for reliability\.[Figure4](https://arxiv.org/html/2607.22980#S4.F4)and[Figure5](https://arxiv.org/html/2607.22980#S4.F5)also show Baujat plots from our influence analysis of reliability and Shannon entropy, respectively\. In a Baujat plot, the horizontal axis corresponds to the contribution of a study to the overall between\-study heterogeneity, and the vertical axis pertains to how much the overall heterogeneity would change if the study were removed\. A point in the upper\-right is a study that is both a poor fit to the model and substantially shifts the pooled estimate when removed, the classic signature of a problem study the plot is designed to surface\[[19](https://arxiv.org/html/2607.22980#bib.bib13)\]\. In our Baujat plots, extreme points in the top quartile on either axis are labeled\.

![Refer to caption](https://arxiv.org/html/2607.22980v1/assets/entropy-forest.png)\(a\)Per\-dataset effects, pooled estimate, and prediction interval\.
![Refer to caption](https://arxiv.org/html/2607.22980v1/assets/entropy-baujat.png)\(b\)Contribution to heterogeneity versus influence on the pooled result\.

Figure 5:Forest and Baujat plots for Shannon entropy\.
### 4\.3Compute Profiling

Table 6:Central tendencies of energy consumption \(Wh\) across datasets\. Prior to central tendency calculation, mean energy consumption was calculated between folds of LOSO/CV within datasets\.[Table6](https://arxiv.org/html/2607.22980#S4.T6)summarizes the energy consumption measured in watt\-hours \(Wh\) from training pipelines across datasets\. As context, consider that the French Agency of Ecological Transition reports that the cost of charging a smartphone once per day is 19\.178 Wh/day\[[27](https://arxiv.org/html/2607.22980#bib.bib28)\]\. Generally, our frequentist pipelines were below this marker, and our Bayesian pipelines were above it\. To gauge the cost of charging a smartphone, consider that a combination refrigerator freezer with 24/7 usage on average consumes 863\.01 Wh/day, and running a washing machine and tumble dryer together has an average combined energy consumption of 1338\.38 Wh/cycle\[[27](https://arxiv.org/html/2607.22980#bib.bib28)\]\.

Computed in log space, the geometric\-mean ratio of increased energy consumption by the Bayesian complete\-pooling pipelines was13\.1×13\.1\\times\. These verifications gave us perspectives to quantitatively understand the computational costs of training Bayesian complete\-pooling models for MI\-EEG classification, both in terms of comparisons to baseline frequentist models, and in absolute terms\. Bayesian ML is substantially more expensive to train than baseline models, however, compared to everyday household appliances it is economical\. Since training costs heavily depend on algorithm and hardware implementations, we kept our analysis of computational costs informal\.

## 5Discussion

### 5\.1Interpretation

Our large\-scale experiment shows that the effects of Bayesian complete\-pooling models for left/right hand MI\-EEG cross\-subject classification are not practically significant\. Still, the statistical significance in reliability and sharpness tells us that even with complete\-pooling models, Bayesian classification is doing real work\. On average, our Bayesian complete\-pooling model predicted classes with frequencies that more closely matched the true frequencies of classes, giving better reliability\. Shannon entropy increased \(sharpness decreased\), meaning that Bayesian complete\-pooling model predictions were more uncertain or more hedged between classes\.

The effects on reliability and sharpness that we empirically observed tell us that our baseline frequentist classifiers were overconfident\. Bayesian model averaging integrates predictions across the full posterior rather than relying on a single point estimate\. In high\-dimensional parameter spaces, the region immediately around the posterior mode contains disproportionately little of the total probability mass compared to the much larger volume of the surrounding region, known as the typical set\[[30](https://arxiv.org/html/2607.22980#bib.bib15)\]\. Therefore, the averaged predictive distribution tends to smooth out the overconfident predictions a point estimate alone would produce\.

At the same time, the null differences we observed for resolution and discrimination tell us that our Bayesian complete\-pooling models found a similar decision boundary to our frequentist baselines\. Like our point estimate models, the Bayesian complete\-pooling models averaged decision boundaries across the population of subjects and sessions for low SNR MI\-EEG\. This population\-level averaging explains why uncertainty quantification would do little to improve classification accuracy\. Our null difference in Brier score tells us that Bayesian complete\-pooling does not improve the ability of classifiers to recognize patterns in left vs\. right hand MI\-EEG\.

### 5\.2Limitations

Our pipeline evaluations were fit without hyperparameter optimization \(HPO\)\. This decision was made out of practical time constraints\. Our experiment fit 12 models with LOSO or 10\-fold cross\-validation, plus Platt scaling with inner 5\-fold CV for probability calibration of the SVMs\. Common grid search is inefficient in high\-dimensional hyperparameter space\[[24](https://arxiv.org/html/2607.22980#bib.bib19)\], however, libraries such as Optuna could be explored to optimize our hyperparameter optimization\[[37](https://arxiv.org/html/2607.22980#bib.bib29)\]\. At the same time, though, we carefully chose our pipeline hyperparameters informed by the literature and our data normalization for our experiment\.

Meta\-analysis was developed for the purpose of accumulating results from multiple previous studies where raw data is not available for fitting regression models\. Statistics reported by previous studies, though, are lossy summaries of the information available in raw data\. Indeed, our analysis of prediction performance was not limited to a two\-stage approach that calculated effect sizes before estimating pooled effects\. Instead, a one\-stage approach to multilevel random\-effects models could have been taken that was directly fit to the raw data\[[40](https://arxiv.org/html/2607.22980#bib.bib17),[13](https://arxiv.org/html/2607.22980#bib.bib18)\]\. This point summarizes the debate between two\-stage and one\-stage individual participant data \(IPD\) meta\-analysis\[[40](https://arxiv.org/html/2607.22980#bib.bib17)\]\.

### 5\.3Future Work

We conducted a large\-scale experiment to estimate the effect of Bayesian complete\-pooling models for cross\-subject left/right hand MI\-EEG classification\. Empirical results showed that the effect of Bayesian complete\-pooling was not practically significant\. This conclusion was also supported by theory in the literature\. We believe that a similar large\-scale experiment with Bayesian partial\-pooling across subjects and sessions would yield a greater effect on prediction accuracy and its subcomponents reliability, resolution, discrimination, and sharpness\. The computational costs we measured tell us that Bayesian ML research is economical for MI\-EEG classification\.

## Funding

This research received no specific grant from any funding agency in the public, commercial, or not\-for\-profit sectors\.

## Conflict of Interest

The author declares no conflict of interest\.

## Ethics Statement

This study did not involve the collection of new data from human participants\. All datasets analyzed were previously collected, publicly available, and gathered under the ethical approvals described in their original publications \(cited in[Table1](https://arxiv.org/html/2607.22980#S3.T1)\)\.

## Data Availability

The datasets analyzed in this study are publicly available through the Mother of All BCI Benchmarks \(MOABB\) framework\. Primary sources for each dataset are cited in[Table1](https://arxiv.org/html/2607.22980#S3.T1)\. A patch exposing predictions from MOABB, along with code used to generate posterior sampling diagnostics and energy measurements and to perform the meta\-analysis reported here, is available athttps://doi\.org/10\.5281/zenodo\.20415453\. Raw per\-fold predictions, posterior sampling diagnostics, and energy measurements are archived athttps://doi\.org/10\.5281/zenodo\.21538705\.

## References

- \[1\]C\. Benaroch, M\. S\. Yamamoto, A\. Roc, P\. Dreyer, C\. Jeunet, and F\. Lotte\(2022\-04\)When should MI\-BCI feature optimization include prior knowledge, and which one?\.Brain\-Computer Interfaces9\(2\),pp\. 115–128\.External Links:[Link](https://inria.hal.science/hal-03920680),[Document](https://dx.doi.org/10.1080/2326263X.2022.2033073)Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.7.6.1)\.
- \[2\]C\. M\. Bishop and H\. Bishop\(2023\)Deep learning: foundations and concepts\.1 edition,Springer Cham\.External Links:ISBN 978\-3\-031\-45467\-7,[Document](https://dx.doi.org/10.1007/978-3-031-45468-4),[Link](https://link.springer.com/book/10.1007/978-3-031-45468-4)Cited by:[§2\.1](https://arxiv.org/html/2607.22980#S2.SS1.p3.4)\.
- \[3\]C\. M\. Bishop\(2006\)Pattern recognition and machine learning\.1 edition,Information Science and Statistics,Springer New York,New York, NY\.External Links:ISBN 978\-0\-387\-31073\-2,ISSN 1613\-9011Cited by:[§2\.1](https://arxiv.org/html/2607.22980#S2.SS1.p3.4),[§3\.3\.2](https://arxiv.org/html/2607.22980#S3.SS3.SSS2.p6.4)\.
- \[4\]M\. Borenstein, L\. V\. Hedges, J\. P\. T\. Higgins, and H\. R\. Rothstein\(2009\)Introduction to meta\-analysis\.Wiley\.External Links:[Document](https://dx.doi.org/10.1002/9780470743386)Cited by:[§2\.3](https://arxiv.org/html/2607.22980#S2.SS3.p1.4),[§3\.5](https://arxiv.org/html/2607.22980#S3.SS5.p1.1)\.
- \[5\]S\. Brandl and B\. Blankertz\(2020\)Motor imagery under distraction— an open access bci dataset\.Frontiers in NeuroscienceVolume 14 \- 2020\.External Links:[Link](https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2020.566147),[Document](https://dx.doi.org/10.3389/fnins.2020.566147),ISSN 1662\-453XCited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.4.3.1)\.
- \[6\]W\. Chang, W\. Kong, G\. Yan, R\. Lv, K\. Du, M\. T\. Sadiq, B\. Guo, R\. Yin, and X\. Liu\(2025\)A multi\-paradigm EEG dataset for studying upper limb rehabilitation exercises\.Scientific Data12\(1\),pp\. 1877\.External Links:[Document](https://dx.doi.org/10.1038/s41597-025-06147-6),[Link](https://doi.org/10.1038/s41597-025-06147-6),ISSN 2052\-4463Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.5.4.1)\.
- \[7\]S\. Chevallier, I\. Carrara, B\. Aristimunha, P\. Guetschel, S\. Sedlar, B\. Junqueira Lopes, S\. Velut, S\. Khazem, and T\. Moreau\(2024\-04\)The largest eeg\-based bci reproducibility study for open science: the moabb benchmark\.Note:working paper or preprintExternal Links:[Link](https://universite-paris-saclay.hal.science/hal-04537061)Cited by:[§1](https://arxiv.org/html/2607.22980#S1.p2.1),[§3\.3\.2](https://arxiv.org/html/2607.22980#S3.SS3.SSS2.p1.3),[§3\.3](https://arxiv.org/html/2607.22980#S3.SS3.p1.1),[§3\.3](https://arxiv.org/html/2607.22980#S3.SS3.p2.1),[Table 2](https://arxiv.org/html/2607.22980#S3.T2),[Table 2](https://arxiv.org/html/2607.22980#S3.T2.3.2)\.
- \[8\]H\. Cho, M\. Ahn, S\. Ahn, M\. Kwon, and S\. C\. Jun\(2017\-05\)EEG datasets for motor imagery brain–computer interface\.GigaScience6\(7\),pp\. gix034\.External Links:ISSN 2047\-217X,[Document](https://dx.doi.org/10.1093/gigascience/gix034),[Link](https://doi.org/10.1093/gigascience/gix034),https://academic\.oup\.com/gigascience/article\-pdf/6/7/gix034/60709569/gigascience\_6\_7\_gix034\.pdfCited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.6.5.1)\.
- \[9\]M\. P\. Deisenroth, A\. Faisal, and C\. S\. Ong\(2020\)Mathematics for machine learning\.Cambridge University Press\.External Links:ISBN 9781108455145,[Link](https://www.cambridge.org/au/universitypress/subjects/computer-science/pattern-recognition-and-machine-learning/mathematics-machine-learning?format=PB)Cited by:[§3\.3\.1](https://arxiv.org/html/2607.22980#S3.SS3.SSS1.p3.1)\.
- \[10\]T\. Dimitriadis, T\. Gneiting, and A\. I\. Jordan\(2021\)Stable reliability diagrams for probabilistic classifiers\.Proceedings of the National Academy of Sciences118\(8\),pp\. e2016191118\.External Links:[Document](https://dx.doi.org/10.1073/pnas.2016191118),[Link](https://www.pnas.org/doi/abs/10.1073/pnas.2016191118),https://www\.pnas\.org/doi/pdf/10\.1073/pnas\.2016191118Cited by:[§3\.4](https://arxiv.org/html/2607.22980#S3.SS4.p2.1)\.
- \[11\]R\. O\. Duda, P\. E\. Hart, and D\. G\. Stork\(2001\)Pattern classification\.2nd edition,Wiley\-Interscience,New York\.External Links:ISBN 0\-471\-05669\-3Cited by:[§2\.2](https://arxiv.org/html/2607.22980#S2.SS2.p5.1)\.
- \[12\]D\. Forenzo, Y\. Liu, J\. Kim, Y\. Ding, T\. Yoon, and B\. He\(2024\)Integrating simultaneous motor imagery and spatial attention for eeg\-bci control\.IEEE Transactions on Biomedical Engineering71\(1\),pp\. 282–294\.External Links:[Document](https://dx.doi.org/10.1109/TBME.2023.3298957)Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.8.7.1)\.
- \[13\]A\. Gelman and J\. Hill\(2006\)Data analysis using regression and multilevel/hierarchical models\.Cambridge University Press\.Cited by:[§5\.2](https://arxiv.org/html/2607.22980#S5.SS2.p2.1)\.
- \[14\]T\. Gneiting, F\. Balabdaoui, and A\. E\. Raftery\(2007\-04\)Probabilistic forecasts, calibration and sharpness\.Journal of the Royal Statistical Society Series B: Statistical Methodology69\(2\),pp\. 243–268\.External Links:ISSN 1369\-7412,[Document](https://dx.doi.org/10.1111/j.1467-9868.2007.00587.x),[Link](https://doi.org/10.1111/j.1467-9868.2007.00587.x),https://academic\.oup\.com/jrsssb/article\-pdf/69/2/243/49794500/jrsssb\_69\_2\_243\.pdfCited by:[§2\.2](https://arxiv.org/html/2607.22980#S2.SS2.p1.6)\.
- \[15\]A\. L\. Goldberger, L\. A\. Amaral, L\. Glass, J\. M\. Hausdorff, P\. C\. Ivanov, R\. G\. Mark, J\. E\. Mietus, G\. B\. Moody, C\. K\. Peng, and H\. E\. Stanley\(2000\)PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals\.Circulation101\(23\),pp\. e215–e220\.External Links:[Document](https://dx.doi.org/10.1161/01.cir.101.23.e215)Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.15.14.1)\.
- \[16\]M\. Grosse\-Wentrup, C\. Liefhold, K\. Gramann, and M\. Buss\(2009\)Beamforming in noninvasive brain–computer interfaces\.IEEE Transactions on Biomedical Engineering56\(4\),pp\. 1209–1219\.External Links:[Document](https://dx.doi.org/10.1109/TBME.2008.2009768)Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.9.8.1)\.
- \[17\]C\. Guo, G\. Pleiss, Y\. Sun, and K\. Q\. Weinberger\(2017\)On calibration of modern neural networks\.InProceedings of the 34th International Conference on Machine Learning,D\. Precup and Y\. W\. Teh \(Eds\.\),Proceedings of Machine Learning Research, Vol\.70,pp\. 1321–1330\.Cited by:[§2\.2](https://arxiv.org/html/2607.22980#S2.SS2.p5.1)\.
- \[18\]E\. Guttmann\-Flury, X\. Sheng, and X\. Zhu\(2025\)Dataset combining EEG, eye\-tracking, and high\-speed video for ocular activity analysis across BCI paradigms\.Scientific Data12\(1\),pp\. 587\.External Links:[Document](https://dx.doi.org/10.1038/s41597-025-04861-9),[Link](https://doi.org/10.1038/s41597-025-04861-9),ISSN 2052\-4463Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.10.9.1)\.
- \[19\]M\. Harrer, P\. Cuijpers, F\. T\. A, and D\. D\. Ebert\(2021\)Doing meta\-analysis with R: a hands\-on guide\.1st edition,Chapman & Hall/CRC Press,Boca Raton, FL and London\.External Links:ISBN 9780367610074Cited by:[§2\.3](https://arxiv.org/html/2607.22980#S2.SS3.p1.4),[§2\.3](https://arxiv.org/html/2607.22980#S2.SS3.p3.5),[§2\.3](https://arxiv.org/html/2607.22980#S2.SS3.p5.5),[§2\.3](https://arxiv.org/html/2607.22980#S2.SS3.p6.2),[§2\.3](https://arxiv.org/html/2607.22980#S2.SS3.p7.2),[§4\.2](https://arxiv.org/html/2607.22980#S4.SS2.p4.1),[§4\.2](https://arxiv.org/html/2607.22980#S4.SS2.p5.1)\.
- \[20\]T\. Hastie, R\. Tibshirani, and J\. Friedman\(2009\-02\)The elements of statistical learning\.2 edition,Springer Series in Statistics,Springer New York,New York, NY\.External Links:[Document](https://dx.doi.org/10.1007/978-0-387-84858-7),ISBN 978\-0\-387\-84857\-0,ISSN 0172\-7397Cited by:[§3\.3\.2](https://arxiv.org/html/2607.22980#S3.SS3.SSS2.p4.5)\.
- \[21\]J\. P\. T\. Higgins and S\. G\. Thompson\(2002\-06\)Quantifying heterogeneity in a meta\-analysis\.Statistics in Medicine21\(11\),pp\. 1539–1558\.External Links:[Document](https://dx.doi.org/10.1002/sim.1186),ISSN 0277\-6715Cited by:[§4\.2](https://arxiv.org/html/2607.22980#S4.SS2.p2.1)\.
- \[22\]V\. Jayaram and A\. Barachant\(2018\-09\)MOABB: trustworthy algorithm benchmarking for bcis\.Journal of Neural Engineering15\(6\),pp\. 066011\.External Links:[Document](https://dx.doi.org/10.1088/1741-2552/aadea0),[Link](https://doi.org/10.1088/1741-2552/aadea0)Cited by:[§3\.1](https://arxiv.org/html/2607.22980#S3.SS1.p1.1)\.
- \[23\]V\. Jayaram, K\. Fiebig, J\. Peters, and M\. Grosse\-Wentrup\(2018\)Transfer learning for BCIs\.InBrain–Computer Interfaces Handbook: Technological and Theoretical Advances,C\. S\. Nam, A\. Nijholt, and F\. Lotte \(Eds\.\),pp\. 425–441\.External Links:ISBN 978\-1\-4987\-7343\-0,[Document](https://dx.doi.org/10.1201/9781351231954-22)Cited by:[§1](https://arxiv.org/html/2607.22980#S1.p2.1)\.
- \[24\]M\. Kuhn and K\. Johnson\(2019\)Feature engineering and selection: a practical approach for predictive models\.1st edition,Chapman and Hall/CRC\.External Links:[Document](https://dx.doi.org/10.1201/9781315108230)Cited by:[§5\.2](https://arxiv.org/html/2607.22980#S5.SS2.p1.1)\.
- \[25\]S\. Kumar, H\. Alawieh, F\. S\. Racz, R\. Fakhreddine, and J\. d\. R\. Millán\(2024\-02\)Transfer learning promotes acquisition of individual bci skills\.PNAS Nexus3\(2\),pp\. pgae076\.External Links:ISSN 2752\-6542,[Document](https://dx.doi.org/10.1093/pnasnexus/pgae076),[Link](https://doi.org/10.1093/pnasnexus/pgae076),https://academic\.oup\.com/pnasnexus/article\-pdf/3/2/pgae076/57145397/pgae076\.pdfCited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.12.11.1)\.
- \[26\]M\. Lee, O\. Kwon, Y\. Kim, H\. Kim, Y\. Lee, J\. Williamson, S\. Fazli, and S\. Lee\(2019\-01\)EEG dataset and openbmi toolbox for three bci paradigms: an investigation into bci illiteracy\.GigaScience8\(5\),pp\. giz002\.External Links:ISSN 2047\-217X,[Document](https://dx.doi.org/10.1093/gigascience/giz002),[Link](https://doi.org/10.1093/gigascience/giz002),https://academic\.oup\.com/gigascience/article\-pdf/8/5/giz002/60716361/gigascience\_8\_5\_giz002\.pdfCited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.13.12.1)\.
- \[27\]L\. librairie ADEME\(2026\-03\)Panel usages électrodomestiques année 6\.External Links:[Link](https://librairie.ademe.fr/batiment/9155-panel-usages-electrodomestiques-annee-6.html)Cited by:[§4\.3](https://arxiv.org/html/2607.22980#S4.SS3.p1.1)\.
- \[28\]H\. Liu and X\. Lv\(2022\-12\)EEG datasets of stroke patients\.External Links:[Link](https://figshare.com/articles/dataset/EEG_datasets_of_stroke_patients/21679035),[Document](https://dx.doi.org/10.6084/m9.figshare.21679035.v5)Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.14.13.1)\.
- \[29\]H\. Liu, P\. Wei, H\. Wang, X\. Lv, W\. Duan, M\. Li, Y\. Zhao, Q\. Wang, X\. Chen, G\. Shi, B\. Han, and J\. Hao\(2024\)An EEG Motor Imagery Dataset for Brain Computer Interface in Acute Stroke Patients\.Scientific Data11\(1\),pp\. 131\.External Links:[Document](https://dx.doi.org/10.1038/s41597-023-02787-8)Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.14.13.1)\.
- \[30\]D\. J\. C\. MacKay\(2003\)Information theory, inference, and learning algorithms\.Cambridge University Press,Cambridge, UK\.External Links:ISBN 0\-521\-64298\-1Cited by:[§3\.4](https://arxiv.org/html/2607.22980#S3.SS4.p2.1),[§5\.1](https://arxiv.org/html/2607.22980#S5.SS1.p2.1)\.
- \[31\]O\. A\. Martin, R\. Kumar, and J\. Lao\(2021\)Bayesian modeling and computation in python\.1 edition,Chapman and Hall/CRC\.External Links:[Document](https://dx.doi.org/10.1201/9781003019169),[Link](https://doi.org/10.1201/9781003019169)Cited by:[§4\.1](https://arxiv.org/html/2607.22980#S4.SS1.p1.3)\.
- \[32\]R\. McElreath\(2020\)Statistical rethinking: a bayesian course with examples in R and STAN\.2 edition,Chapman and Hall/CRC\.External Links:[Document](https://dx.doi.org/10.1201/9780429029608)Cited by:[§2\.1](https://arxiv.org/html/2607.22980#S2.SS1.p5.1)\.
- \[33\]A\. H\. Murphy\(1973\)A new vector partition of the probability score\.Journal of Applied Meteorology12\(4\),pp\. 595–600\.External Links:[Document](https://dx.doi.org/10.1175/1520-0450%281973%29012%3C0595%3AANVPOT%3E2.0.CO%3B2)Cited by:[§3\.4](https://arxiv.org/html/2607.22980#S3.SS4.p2.1)\.
- \[34\]K\. P\. Murphy\(2022\)Probabilistic machine learning: an introduction\.MIT Press\.External Links:[Link](http://probml.github.io/book1)Cited by:[§2\.1](https://arxiv.org/html/2607.22980#S2.SS1.p1.13)\.
- \[35\]K\. P\. Murphy\(2023\)Probabilistic machine learning: advanced topics\.MIT Press\.External Links:[Link](http://probml.github.io/book2)Cited by:[§2\.1](https://arxiv.org/html/2607.22980#S2.SS1.p3.4),[§2\.2](https://arxiv.org/html/2607.22980#S2.SS2.p5.1),[§3\.3\.2](https://arxiv.org/html/2607.22980#S3.SS3.SSS2.p5.3)\.
- \[36\]Y\. Ovadia, E\. Fertig, J\. Ren, Z\. Nado, D\. Sculley, S\. Nowozin, J\. V\. Dillon, B\. Lakshminarayanan, and J\. Snoek\(2019\)Can you trust your model’s uncertainty? evaluating predictive uncertainty under dataset shift\.Curran Associates Inc\.,Red Hook, NY, USA\.Cited by:[§1](https://arxiv.org/html/2607.22980#S1.p3.1)\.
- \[37\]Y\. Ozaki, S\. Watanabe, and T\. Yanase\(2025\)OptunaHub: a platform for black\-box optimization\.arXiv preprint arXiv:2510\.02798\.Cited by:[§5\.2](https://arxiv.org/html/2607.22980#S5.SS2.p1.1)\.
- \[38\]L\. Pillette, A\. Roc, B\. N’Kaoua, and F\. Lotte\(2021\)Experimenters’ influence on mental\-imagery based brain\-computer interface user training\.International Journal of Human\-Computer Studies149,pp\. 102603\.External Links:ISSN 1071\-5819,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/j.ijhcs.2021.102603),[Link](https://www.sciencedirect.com/science/article/pii/S1071581921000215)Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.7.6.1)\.
- \[39\]C\. E\. Rasmussen and C\. K\. I\. Williams\(2005\-11\)Gaussian processes for machine learning\.The MIT Press\.External Links:ISBN 9780262256834,[Document](https://dx.doi.org/10.7551/mitpress/3206.001.0001),[Link](https://doi.org/10.7551/mitpress/3206.001.0001),https://direct\.mit\.edu/book\-pdf/2514321/book\_9780262256834\.pdfCited by:[§3\.3\.2](https://arxiv.org/html/2607.22980#S3.SS3.SSS2.p6.4)\.
- \[40\]R\. D\. Riley, J\. F\. Tierney, and L\. A\. Stewart \(Eds\.\)\(2021\)Individual participant data meta\-analysis: a handbook for healthcare research\.Wiley\.External Links:[Document](https://dx.doi.org/10.1002/9781119333784)Cited by:[§5\.2](https://arxiv.org/html/2607.22980#S5.SS2.p2.1)\.
- \[41\]R\. T\. Schirrmeister, J\. T\. Springenberg, L\. D\. J\. Fiederer, M\. Glasstetter, K\. Eggensperger, M\. Tangermann, F\. Hutter, W\. Burgard, and T\. Ball\(2017\)Deep learning with convolutional neural networks for eeg decoding and visualization\.Human Brain Mapping38\(11\),pp\. 5391–5420\.External Links:[Document](https://dx.doi.org/https%3A//doi.org/10.1002/hbm.23730),[Link](https://onlinelibrary.wiley.com/doi/abs/10.1002/hbm.23730),https://onlinelibrary\.wiley\.com/doi/pdf/10\.1002/hbm\.23730Cited by:[§3\.3\.1](https://arxiv.org/html/2607.22980#S3.SS3.SSS1.p4.4),[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.16.15.1)\.
- \[42\]J\. Shi, D\. Chen, X\. Zhao, Z\. Zhao, S\. Li, Y\. Xu, T\. Ding, Z\. Zhu, P\. Zhang, Q\. Ye, Y\. Tang, P\. Zhang, B\. Tao, and Z\. Tang\(2025\)HEFMI\-ICH: a hybrid EEG\-fNIRS motor imagery dataset for brain\-computer interface in intracerebral hemorrhage\.Scientific Data12\(1\),pp\. 1816\.External Links:[Document](https://dx.doi.org/10.1038/s41597-025-06100-7),[Link](https://doi.org/10.1038/s41597-025-06100-7),ISSN 2052\-4463Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.11.10.1)\.
- \[43\]J\. Shin, A\. von Lühmann, B\. Blankertz, D\. Kim, J\. Jeong, H\. Hwang, and K\. Müller\(2017\)Open access dataset for eeg\+nirs single\-trial classification\.IEEE Transactions on Neural Systems and Rehabilitation Engineering25\(10\),pp\. 1735–1745\.External Links:[Document](https://dx.doi.org/10.1109/TNSRE.2016.2628057)Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.17.16.1)\.
- \[44\]J\. R\. Stieger, S\. A\. Engel, and B\. He\(2021\)Continuous Sensorimotor Rhythm Based Brain Computer Interface Learning in a Large Population\.Scientific Data8\(1\),pp\. 98\.External Links:[Document](https://dx.doi.org/10.1038/s41597-021-00883-1)Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.18.17.1)\.
- \[45\]M\. Tangermann, K\. Müller, A\. Aertsen, N\. Birbaumer, C\. Braun, C\. Brunner, R\. Leeb, C\. Mehring, K\. J\. Miller, G\. Mueller\-Putz, G\. Nolte, G\. Pfurtscheller, H\. Preissl, G\. Schalk, A\. Schlögl, C\. Vidaurre, S\. Waldert, and B\. Blankertz\(2012\)Review of the bci competition iv\.Frontiers in NeuroscienceVolume 6 \- 2012\.External Links:[Link](https://www.frontiersin.org/journals/neuroscience/articles/10.3389/fnins.2012.00055),[Document](https://dx.doi.org/10.3389/fnins.2012.00055)Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.2.1.1),[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.3.2.1)\.
- \[46\]D\. S\. Wilks\(2019\)Chapter 9 \- forecast verification\.InStatistical Methods in the Atmospheric Sciences \(Fourth Edition\),pp\. 369–483\.External Links:ISBN 978\-0\-12\-815823\-4,[Document](https://dx.doi.org/https%3A//doi.org/10.1016/B978-0-12-815823-4.00009-2),[Link](https://www.sciencedirect.com/science/article/pii/B9780128158234000092)Cited by:[§1](https://arxiv.org/html/2607.22980#S1.p4.1),[§2\.2](https://arxiv.org/html/2607.22980#S2.SS2.p1.6),[§2\.2](https://arxiv.org/html/2607.22980#S2.SS2.p2.11),[§2\.2](https://arxiv.org/html/2607.22980#S2.SS2.p3.1),[§2\.2](https://arxiv.org/html/2607.22980#S2.SS2.p5.1),[§3\.4](https://arxiv.org/html/2607.22980#S3.SS4.p2.1)\.
- \[47\]B\. Yang, F\. Rong, Y\. Xie, D\. Li, J\. Zhang, F\. Li, G\. Shi, and X\. Gao\(2025\)A multi\-day and high\-quality EEG dataset for motor imagery brain\-computer interface\.Scientific Data12\(1\),pp\. 488\.External Links:[Document](https://dx.doi.org/10.1038/s41597-025-04826-y),[Link](https://doi.org/10.1038/s41597-025-04826-y),ISSN 2052\-4463Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.20.19.1)\.
- \[48\]W\. Yi, S\. Qiu, K\. Wang, H\. Qi, L\. Zhang, P\. Zhou, F\. He, and D\. Ming\(2014\-12\)Evaluation of eeg oscillatory patterns and cognitive process during simple and compound limb motor imagery\.PLOS ONE9\(12\),pp\. 1–19\.External Links:[Document](https://dx.doi.org/10.1371/journal.pone.0114853),[Link](https://doi.org/10.1371/journal.pone.0114853)Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.19.18.1)\.
- \[49\]Y\. Zhang\(2018\)Bayesian learning for EEG analysis\.InBrain–Computer Interfaces Handbook: Technological and Theoretical Advances,C\. S\. Nam, A\. Nijholt, and F\. Lotte \(Eds\.\),pp\. 407–424\.External Links:ISBN 978\-1\-4987\-7343\-0,[Document](https://dx.doi.org/10.1201/9781351231954-21)Cited by:[§1](https://arxiv.org/html/2607.22980#S1.p2.1)\.
- \[50\]Q\. Zhou, J\. Lin, L\. Yao, Y\. Wang, Y\. Han, and K\. Xu\(2021\)Relative power correlates with the decoding performance of motor imagery both across time and subjects\.Frontiers in Human Neuroscience15\.External Links:[Document](https://dx.doi.org/10.3389/fnhum.2021.701091),[Link](https://www.frontiersin.org/journals/human-neuroscience/articles/10.3389/fnhum.2021.701091),ISSN 1662\-5161Cited by:[Table 1](https://arxiv.org/html/2607.22980#S3.T1.4.21.20.1)\.

Similar Articles

Meta-learning In-Context Enables Training-Free Cross Subject Brain Decoding

Hugging Face Daily Papers

This paper introduces a meta-optimized approach for semantic visual decoding from fMRI signals that generalizes to novel subjects without fine-tuning, using in-context learning to infer unique neural encoding patterns from a small set of image-brain activation examples. The method achieves strong cross-subject and cross-scanner generalization without requiring anatomical alignment or stimulus overlap.

Modeling Shared and Individual Structure for Cross-Subject Continuous Affect Regression from EEG-fNIRS

arXiv cs.AI

The paper studies zero-shot cross-subject continuous valence-arousal regression from synchronized EEG-fNIRS data, decomposing affect into a stimulus-shared component and an individual component calibrated via label-free alpha-band cross-channel synchrony, outperforming EEGNet and ASAC-Net baselines while reporting a systematic negative-result search over alternative architectures and features.