The Deep-Match Framework for Event-Related Potential Detection in EEG

Abstract

Reliable detection of event-related potentials (ERPs) at the single-trial level remains a challenge due to low signal-to-noise ratio and high variability in electroencephalography (EEG) recordings. This work investigates the use of the Deep-Match framework (Deep-MF) for ERP detection. We examine whether incorporating prior knowledge of an ERP template into deep learning models improves detection performance. As a proof-of-concept study, the framework was evaluated on a single dataset with multi-channel EEG recordings during laser stimulation. The model was trained in two stages. First, an encoder–decoder architecture was trained to reconstruct input EEG signals in order to learn compact signal representations. In the second stage, the decoder was replaced with a detection module and the network was fine-tuned for ERP identification. Two model variants were evaluated: a standard model with randomly initialized filters and a Deep-MF model in which input kernels were initialized using ERP templates. Models performance was assessed on a single-trial ERP detection task during leave-one-out validation, and then compared with matched filter detector. The neural network models outperformed the matched filter detector and proposed that the Deep-MF model slightly outperformed the detector with standard kernel initialization for the majority of held-out subjects. Although both approaches exhibited substantial inter-subject variability, Deep-MF achieved a higher average F1-score (0.37) compared to the standard network (0.34), indicating improved robustness to cross-subject differences. Performance varied considerably across participants. The best performance obtained by Deep-MF reached an F1-score of 0.71, exceeding the maximum score achieved by the standard model (0.59). These results showed that ERP-informed kernel initialization provides improvements in single-trial ERP detection under subject-independent evaluation. These findings demonstrate that integrating domain knowledge with deep learning architectures can improve single-trial ERP detection. The proposed approach provides a step towards practical wearable EEG and passive brain–computer interface applications, as well as towards real-time monitoring of cognitive processes.

1. Introduction

An event-related potential (ERP) is a measured brain response that is evoked by a specific sensory, cognitive, or motor event [1]. ERPs have been used in attention studies [2], for cognitive load estimation during N-back tasks [3], and for hearing threshold estimation [4]. Event-related potentials (ERPs) are widely used to aid in the diagnosis and assessment of various neurological disorders. Specifically, ERPs can detect early memory and cognitive declines before significant structural changes become visible on MRI. Furthermore, they help identify cognitive impairment associated with Multiple Sclerosis [5], serve as biomarkers to differentiate patients with mild cognitive impairment from healthy controls [6], and assist in evaluating stroke recovery [7]. Changes in ERP have also been observed in schizophrenia [8].

Since ERP amplitude is similar to spontaneous brain potentials, ERPs are burned in EEG signals. During typical ERP analysis in cognitive neuroscience, the grand average response across multiple trials is used to analyze and compare differences between ERP characteristics (amplitude, latency) across subjects [9]. The requirement of several trials is mainly due to the noise in the signal coming from eye movements, muscular contractions, and ongoing brain activity that is unrelated to the experimental task. Due to averaging ERP characteristics variations across trials may not be captured by the grand average response.

On the other hand, the brain–computer interface (BCI) is required to detect specific event-related potentials. In many real-world scenarios, repeated stimulus presentation is either impractical or impossible. For example, when stimuli appear only once, for example, as an individual frame in a video stream or dynamically presented image, there is no opportunity to aggregate responses across repetitions [10]. Moreover, repetition itself may affect the neural response due to memory or training effects. In such time-constrained and dynamically changing environments, decisions must rely on single-trial EEG responses. Unlike traditional ERP analysis, single-trial detection aims to identify event-related responses within individual recordings, making the task considerably more challenging. Therefore, improving the accuracy and robustness of single-trial ERP detection is essential for enabling practical, high-speed, and gaze-independent BCI systems, as well as for increasing their applicability in naturalistic settings. ERPs can be estimated using iterative template-matching methods [11], adaptive filtering and weighted averaging techniques [12], or matched filtering approaches [13]. Depuydt et al. [14] demonstrated that neural networks outperform conventional single-trial latency estimation methods.

This work investigates the adaptation of the Deep-Match framework (Deep-MF) [15] for single-trial ERP detection. The Deep-MF was originally proposed for the detection of R-peaks in ear-ECG signals [15]. Deep-MF consists of an encoder stage, initialized with an ECG template, and an R-peak classifier stage. Operating as a matched filter, the encoder searches for matches between the ECG template and the input signal. These candidate matches are subsequently processed by convolutional layers, which refine the detections and identify peaks corresponding to the ground-truth ECG. Deep-MF outperforms existing algorithms for R-peak detection in noisy ECG recordings. Moreover, it can be used for online R-peak detection on mobile devices [16] and for robust heart rate estimation [17,18].

In this paper, we evaluate the potential application of Deep-MF for single-trial ERP detection. Our primary methodological contribution lies in an ERP-specific adaptation and template-based initialization strategy for Deep-MF, rather than proposing an entirely new neural network architecture from scratch.

Translating Deep-MF from ECG to EEG presents unique signal-processing challenges. Unlike high-amplitude, highly localized R-peaks in single-lead ECG, ERPs exhibit lower signal-to-noise ratios. As a proof-of-concept study, we assess the performance of Deep-MF on a single dataset of pain-related ERPs evoked by laser stimulation. Laser-evoked potentials serve as good test case for this model because laser pulses evoke distinct N2–P2 wave complexes that yield a high signal-to-noise ratio and correlate strongly with subjective pain perception [19].

To this end, we investigate whether the matched filter-based encoder can effectively capture the characteristic temporal morphology of laser-evoked ERP components in noisy EEG recordings. To benchmark its performance, we compare the Deep-MF approach with a standard convolutional neural network (CNN) model trained directly for ERP detection and with a standard matched filter method.

2. Method

For this study, we used the ds005284 dataset from OpenNeuro repository (https://openneuro.org/datasets/ds005284/ (accessed on 18 July 2026)). The dataset contains a 64-lead EEG recorded from 26 young participants (average age 21 years, 18 Female). The data was recorded by Lu et al. [20]. None of the participants in this study reported any hearing impairment or neurological or psychiatric disorder. This study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee at the Institute of Psychology, Chinese Academy of Sciences. The written informed consent for participation was obtained from all participants prior to testing [20].

Each participant received 16 fixed-intensity laser stimuli delivered at approximately 20 s intervals [20]. During the experiment, participants were instructed to keep their eyes open, focus on a fixation cross displayed on the screen, and attend carefully to each laser stimulus. After each stimulation, a 3–5 s pause was introduced, during which participants orally reported their pain rating within a 3–5 s response window. The subsequent trial began randomly within 1–3 s after the rating was completed.

For the preprocessing of signals, we used EEGLAB 2022.1 and MatLab R2023B. We followed the established pipeline [21]. The EEG data were filtered using high- and low-pass filters, with cut-off frequencies of 1 Hz and 40 Hz. Signals were re-referenced to the average of the P9 and P10 electrodes. Independent component analysis was then performed.

First, we performed event-related potential (ERP) analysis across EEG channels. The data were epoched around the event onset, with time limits set from −0.2 s to 1 s relative to the event. The pre-event interval (−0.2 to 0 s) was used for baseline correction by subtracting the mean baseline activity from each epoch. We used the $pop\_autorej$ function to automatically remove epochs contaminated by artifacts. All epochs for each channel were then averaged. Figure 1 shows sample ERPs. Spatially, ERP signal quality was highest at central electrodes and lower at peripheral locations. Channels were selected through visual inspection of the grand-averaged ERP templates across all participants. Specifically, Cz, C3, C4, and Pz channels were selected as model inputs based on visual inspection. The channels met the predefined criteria of exhibiting a high SNR alongside a prominent, well-defined ERP template with clear visible negative and positive peak complexes. For further analysis, the data were resampled to 250 Hz from 1024 Hz.

Figure 1.

Grand-average event-related potentials (ERPs) for representative EEG leads across subjects. Clear ERP components are visible over central electrodes (C3, Cz, Pz, and C4), whereas more distant electrodes show weaker or absent responses. Only a subset of leads is displayed for clarity; similar symmetrical patterns were observed across contralateral electrodes.

Spatially, ERP signal quality was highest at central electrodes and lower at peripheral locations. Channels were selected through visual inspection of the grand-averaged ERP templates across all participants. Specifically, Cz, C3, C4, and Pz were retained as they met the predefined criteria of exhibiting a high SNR alongside prominent, well-defined negative and positive peak complexes.

To build ERP templates, the averaged ERP across all subjects was computed for each selected channel. The resulting templates were then smoothed using the smoothdata function. In deep convolutional networks, improper weight initialization can cause training gradients to explode or vanish. To prevent these issues, we employed the kernel scaling method proposed by He et al. [22]. First, the mean value of each kernel was subtracted, and the templates were divided by their standard deviations. Finally, each template was scaled using the factor:

$$\sigma =\sqrt{{\displaystyle \frac{2}{\mathrm{kernel}\mathrm{height}\times \mathrm{kernel}\mathrm{width}\times \mathrm{number}\mathrm{of}\mathrm{input}\mathrm{channels}}}}$$ (1)

The resulting templates were used to initialize the input kernels of Deep-MF. Detectors were trained using PyTorch 2.8.0. All models were trained using the Adam optimizer with a learning rate of 1 × 10⁻³. The architecture of the network is shown in Figure 2. The model was trained in two stages. In the first stage, it was trained as an encoder–decoder [23] to reconstruct the input signal. The data were standardized and split into 2 s samples with 80% overlap. Two encoder–decoders were trained initially: one with standard weight initialization and another using Deep-MF, where the input filters were initialized with ERP templates. Mean squared error loss function was used.

Figure 2.

Architecture of the model. The model was trained in two stages. In the first stage, the encoder–decoder network was trained to reconstruct the input signal. Two models were trained: Deep-MF, with input filters initialized using ERP templates, and a standard model with random initialization. In the second stage, the decoder part of the network was replaced with a detector, and the models were fine-tuned for ERP detection.

In the second stage, ERP detectors were trained. The pretrained models were fine-tuned, and the decoder part of each model was replaced with a detector (Figure 2). The resulting models were trained to detect events, specifically the timing of the ERP. In this step, a balanced dataset was used, with an equal number of samples containing events and samples without events. For each event, 12 samples were selected with a 0.1 s step, and the starting position for each window was randomized within the step size. Samples without events were randomly selected from the dataset. As model input, 2 s segments were used. The target outputs were created as 300-sample vectors, where a value of 1 corresponded to the event time.

In the second stage, leave-one-out (LOO) validation was used to fine-tune the encoder–decoders. For each iteration, one participant was held out as the validation set while the remaining participants were used for training. This procedure was repeated until every participant had served as the validation set once. To account for the rarity of positive events, the models were trained using a weighted binary cross-entropy loss with a positive class weight of 3.0.

For evaluation during leave-one-out (LOO) validation, ERP peaks were detected from the model outputs using a peak-finding algorithm. Detected peaks were identified with a minimum height threshold (0.25) and a minimum distance between peaks (30). The detected peaks were then compared to the ground-truth event times using a tolerance window to account for slight timing differences. True positives (TPs), false positives (FPs), and false negatives (FNs) were computed for each participant. We count a TP when a peak is detected in the model output within a tolerance of ±30 ms from a true peak; a FP for all other peaks detected in the model output; and a FN for a lack of detection in a given window. The F1 metric was used to assess the accuracy and reliability of the ERP detection. To establish a baseline, a classical matched filter utilizing generated templates was evaluated under the identical peak-detection pipeline.

All scripts used for the analysis reported in this study are available in the GitHub repository: https://github.com/Marower/DeepMF_ERP_Detection_in_EEG-main (accessed on 18 July 2026).

3. Results

ERPs were analyzed across multiple EEG leads to examine their spatial distribution. As illustrated in Figure 1, clear ERP components are consistently observable over the central electrodes (C3, Cz, and Pz) across subjects. In contrast, electrodes located further from the central region exhibit markedly weaker or indistinct ERP responses. This spatial pattern suggests that task-related neural activity is predominantly localized over the central scalp areas. For clarity, only a subset of EEG leads is shown in the figure; however, similar symmetrical patterns were observed across the corresponding contralateral electrodes.

Based on these observations, four channels (C3, Cz, Pz, and C4) were selected for subsequent analysis. These channels were chosen because they strictly met our criteria of displaying a high signal-to-noise ratio (SNR) with clear, prominent negative (N2) and positive (P2) peaks in the grand-averaged templates across all subjects. Restricting our analysis to these central channels reduces input dimensionality while preserving the most informative signals, thereby reducing the network size and improving computational efficiency. The grand average provides a broad, population-level topological baseline rather than subject-specific fine tuning, making the risk of optimistic performance bias on any individual test-fold minimal.

Leave-one-subject-out cross-validation was performed to evaluate subject-independent generalization. Figure 3 presents the F1-scores obtained for each held-out subject. Deep-MF slightly outperformed the detector trained with standard kernel initialization on the majority of subjects. Although both methods exhibited variability across individuals, Deep-MF achieved a higher average F1-score (0.37) compared to the second detector (0.34). Both methods showed their poorest performance on Subject 10, where the F1-score was 0.01. The highest F1-score achieved by Deep-MF was 0.71, compared to 0.59 for the standard network.

Figure 3.

F1-scores obtained for each subject using leave-one-subject-out cross-validation. Deep-MF consistently outperformed the standard network on the majority of subjects and achieved a higher mean F1-score (0.37 vs. 0.34). The lowest performance for both methods occurred for Subject 10 (F1 = 0.01).

When evaluated as a baseline reference, the traditional matched filter yielded a lower overall average F1-score of 0.11. However, its performance is stable and less dependent on the participant. What is interesting is that the matched filter outperforms CNN models in Subject 10.

We used a paired t-test to compare model performance across validation folds. Statistical comparison across validation folds revealed a higher sensitivity for Deep-MF compared to the baseline (p = 0.0661, Cohen’s d = 0.22), while maintaining comparable precision (p = 0.2293). This translated to an overall upward trend in F1-score (p = 0.1208). These results indicate that the integration of template knowledge in Deep-MF primarily enhances the model’s ability to detect true ERP events (sensitivity) without increasing false positives. Given that this is a single-dataset proof-of-concept study, the sample size limits overall statistical power.

4. Discussion

ERPs exhibit substantial variability across single trials. Differences in amplitude and latency carry important information regarding attention, mental state, fatigue, and habituation [24]. Single-trial EP evaluation remains a challenging problem, with multiple possible applications, inter alia BCI [25], auditory attention [26], or pain evaluation [27].

In our study we evaluate application of Deep-match framework for detection of a single-trial ERP. In this proof-of-concept study, we used the Deep-MF to provide exact time of ERP, not only binary classification. We found that the Deep-MF provides slightly better results to CNN model and outperform traditional matched filter. Initialization of the input kernel using the ERP template improved the sensitivity of the model.

Overall, the model’s performance is surpassed by the Deep-MF method [15] for R-peak detection. This can be explained by the lower SNR of EEG. Amplitudes of evoked potentials are lower than the amplitude of the spontaneous EEG [28], and the ERP template inter-subject variability is higher compared to I-lead ECG morphology.

On the other hand, laser stimulation produces robust, high-amplitude ERP compared to other cognitive ERPs. This robust morphology facilitated reliable grand-average template extraction for initializing the Deep-MF encoder.

However, this high baseline SNR raises an important consideration regarding generalization in paradigms characterized by weaker brain responses. Nevertheless, because the Deep-MF encoder filters remain trainable during backpropagation, the network possesses the capacity to adapt beyond the initial template weights. Future work should evaluate the limits of this initialization scheme across lower-SNR cognitive ERP tasks, exploring adaptive or soft-constrained template initialization strategies to maintain performance under low signal strength.

With regard to the neural processing of pain, the insular cortex is thought to mediate both sensory-discriminative and affective-motivational dimensions [29]. While specific motor-related areas are essential for executing context-sensitive behavioral responses to painful stimuli [30], the bilateral insular cortex appears to maintain a consistent response to pain regardless of whether an action is required. Valeriani et al. [31] found that insular cortex potentials mirror those recorded at the Cz lead. This correlation accounts for the high-quality ERP components observed in our study across the Cz, C3, C4, and Pz leads.

Shalchy et al. [21] found significant differences in behavioral and electrophysiological signatures in response to N-back stimulus type, task structure, preprocessing method, and laboratory equipment. Late latency peak (P300) is more closely related to cognitive processing than to the physical attributes of the stimulus [32]. Brain cognitive state can be affected by multiple factors, such as external ambient sounds or internal arousal level [33]. Consequently, systematic investigation of preprocessing strategies to determine the optimal configuration remains an important task for future work.

Individual differences in ERPs may arise from multiple sources, including biological factors (age, brain anatomy, clinical status), cognitive traits (attention, working memory, executive control), momentary brain states (arousal, fatigue), task and stimulus properties, and technical aspects of EEG acquisition and preprocessing. These individual differences present major challenges for Deep-MF models, which are based on signal templates. One possible solution is the use of individualized or reinforcement learning-based templates that adapt the network to personal characteristics [34]. Development of the method will require interdisciplinary cooperation that recognizes the full range of human cognition, effort, and attention, with the goal of improving the ability of ERP detection models to adapt to changing cognitive states and ensure reliable performance in real-world conditions outside laboratory environments. Such research should include not only laboratory studies but also real-life experiments [35].

Another approach is to improve detection by analyzing the internal structure of ERP signals and designing multiple matched-filter templates that capture their characteristic components. According to Longe et al. [36], for movement-related potential, the ERP can be divided into three main components: a P1-N1-P2-N2 complex, a P300 component and late potentials. Authors used 5 templates for matched filters and achieved a SNR improved by up 10 dB compared to other methods.

Moreover, different components of an evoked potential complex may originate from different functional brain sites and can be distinguished according to their respective amplitudes [37]. In this study, ERP responses were mainly localized in central leads, but different stimulus potentials may originate from different sites, increasing complexity of a proper Deep-MF design.

The performance of deep learning methods critically depends on the quality and quantity of the available training data. This is particularly true for physiological time series, which are often noisy and limited in size, thereby motivating the use of data augmentation techniques to artificially increase dataset size [38]. A data augmentation approach based on epoch averaging was employed by Nguyen et al. to improve the performance of a deep learning model for single-word auditory attention decoding. Such augmentation strategies should be considered in future work to enhance the performance of Deep-MF, since the dataset used in this study contains only 16 events per subject.

The observed accuracy suggests that Deep-MF is a promising approach for reliable event detection in EEG signals. Furthermore, Deep-MF can be integrated into advanced analytical pipelines. Its computationally efficient detection capability may serve as a preliminary stage prior to the application of more complex algorithms, such as speech reception threshold evaluation [39]. By restricting the use of large-scale models, including general foundation models with millions of parameters [40], to contextually relevant signal segments, Deep-MF may reduce the overall computational load. This reduction directly improves energy efficiency, which is critical for wearable systems. Furthermore, such an approach may facilitate the development of next-generation hearing assistive devices designed to reduce sustained listening effort [41] through selective and attention-driven amplification.

Further studies should include the evaluation of the concept on larger datasets, evaluation across different tasks, optimization of the network architecture, and improvement of the processing pipeline. In addition to proper frequency filtering, signal components may be extracted using methods such as dynamic mode decomposition [42] and appropriate baseline correction methods [43].

5. Conclusions

In this work, we used the Deep-Match framework for the detection of event-related potentials in EEG signals. We showed that the Deep-MF model yielded higher sensitivity than the standard model. Initializing the input kernels with ERP templates improved performance and enhanced interpretability.

Single-trial ERP detection may provide a foundation for brain–computer interface applications. Reliable single-trial ERP detection enables adaptive interfaces that dynamically adjust system behavior according to user cognitive responses. Wearable EEG headbands could detect ERP components (e.g., P300) to estimate mental workload in real time.

Further studies should focus on improving the proposed framework. First, systematic optimization of the preprocessing pipeline may further enhance ERP detectability. Second, the use of adaptive ERP templates that dynamically adjust to subject-specific or session-dependent variability could improve robustness across individuals and recording conditions. Additionally, reinforcement learning approaches may allow adaptive optimization in real-time applications. Optimal performance also necessitates tuning the network architecture. Finally, the use of data augmentation techniques may increase generalization and reduce overfitting in low-data EEG scenarios.

Author Contributions

Conceptualization, M.Ż.; methodology, M.Ż.; software, M.Ż.; validation, M.Ż., B.T.Ś. and G.C.; formal analysis, M.Ż.; investigation, M.Ż.; resources, M.Ż.; data curation, M.Ż.; writing—original draft preparation, M.Ż.; writing—review and editing, B.T.Ś.; visualization, M.Ż.; supervision, G.C.; project administration, G.C.; funding acquisition, G.C. All authors have read and agreed to the published version of the manuscript.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

For this study we used the ds005284 public dataset from the OpenNeuro repository (https://openneuro.org/datasets/ds005284/ (accessed on 18 July 2026)). Code is available in the GitHub repository: https://github.com/Marower/DeepMF_ERP_Detection_in_EEG-main (accessed on 18 July 2026).

Conflicts of Interest

The authors declare no conflicts of interest.

Funding Statement

This work was supported by the National Science Centre, Poland. Grant number 2026/10/X/ST6/00033.