Estimating the Event-Related Potential from Few EEG Trials

TL;DR

ERPs are conventionally estimated from many trials, limiting study speed and applicability across populations. EEG2ERP introduces an uncertainty-aware autoencoder with bootstrapped ERP targets and a variance decoder to recover ERPs from few trials, while disentangling subject and task factors in a split latent space for zero-shot generalization. Across ERP CORE, Wakeman-Henson, and P300 Speller datasets, the approach consistently outperforms conventional and robust averaging in low-data regimes and provides calibrated uncertainty estimates. This work reduces data collection burden, supports reliable ERP analysis in clinical and BCI contexts, and establishes a data-efficient benchmark for future EEG representation learning.

Abstract

Event-related potentials (ERP) are measurements of brain activity with wide applications in basic and clinical neuroscience, that are typically estimated using the average of many trials of electroencephalography signals (EEG) to sufficiently reduce noise and signal variability. We introduce EEG2ERP, a novel uncertainty-aware autoencoder approach that maps an arbitrary number of EEG trials to their associated ERP. To account for the ERP uncertainty we use bootstrapped training targets and introduce a separate variance decoder to model the uncertainty of the estimated ERP. We evaluate our approach in the challenging zero-shot scenario of generalizing to new subjects considering three different publicly available data sources; i) the comprehensive ERP CORE dataset that includes over 50,000 EEG trials across six ERP paradigms from 40 subjects, ii) the large P300 Speller BCI dataset, and iii) a neuroimaging dataset on face perception consisting of both EEG and magnetoencephalography (MEG) data. We consistently find that our method in the few trial regime provides substantially better ERP estimates than commonly used conventional and robust averaging procedures. EEG2ERP is the first deep learning approach to map EEG signals to their associated ERP, moving toward reducing the number of trials necessary for ERP research. Code is available at https://github.com/andersxa/EEG2ERP

Figures (12)

• Figure 1: The EEG2ERP model framework. In the top-right graph, the dashed line is the target ERP, and the blue line and area is the model's estimated ERP and confidence respectively. The model maps an input averaged over $K_{b}$ EEG trials from one set to an estimated ERP averaged over $N_b$ trials from a separate set. It is conditioned on $K_b$ and includes a decoder branch that estimates uncertainty. This visualization simplifies the prediction process to a single channel and demonstrates how bootstrapped ERPs replace the single-trial autoencoder targets of CSLP-AE while accounting for ERP uncertainty including a variance decoder.
• Figure 2: Left panel: Boxplots comparing $R^2$ values of the obtained ERPs against the test ERP. Comparisons include the simple average of all trials (100%), simple average over $K=5$ trials, and EEG2ERP applied at $K=5$ trials. The missing boxes of simple average at $K=5$ for the MMN paradigm appear outside the bounds (below $R^2=-3$). Right panel: Variance estimates given by standard deviation as a function of root mean squared error (RMSE) to the ground truth test ERP. Points represent the RMSE of denoised ERPs using $K=5$ trials to the test ERP for specific subject-task pairs, alongside the noise standard deviation estimates of the denoised trials.
• Figure 3: Left panel: Root Mean Squared Error (RMSE) as a function of the number of trials included, comparing conventional averaging, Dynamic Time Warped averaging molina2024enhanced, and the EEG2ERP procedure. Results are shown for three test subjects across one specific task from each of the six paradigms. Right panel: Estimated ERP signals based on $K=5$ trials, comparing conventional averaging, robust averaging, and EEG2ERP. Confidence bands show bootstrapped estimates of the true ERP, with variance estimates from EEG2ERP predictions displayed as $\pm 2$ times the standard deviation.
• Figure 4: Topographic comparison of EEG2ERP and simple averaging for subject 22, which was not seen during training, under the N170 Faces paradigm at 170 ms and 320 ms. Rows show results using one trial, five trials, all available trials on the model input half, and all available trials on the unseen target half. EEG2ERP recovers the characteristic N170 pattern with very limited data, while simple averaging degrades under low trial counts. Color scale is the same for all plots and indicates scalp potentials in microvolts.
• Figure 5: Confusion matrix on subject, task and paradigm labels using XGBoost Classifier with 5-fold Cross-Validation splits as in CSLP-AE. Latents are encoded using single-trial samples only.