SpellerSSL: Self-Supervised Learning with P300 Aggregation for Speller BCIs

The gist: SpellerSSL, a framework combining self-supervised learning with P300 aggregation, addresses challenges in EEG-based P300 spellers by enhancing signal quality and improving generalization through pretraining and aggregation strategies.

Introduction to Challenges

Electroencephalogram (EEG)-based P300 spellers face three main challenges: low signal-to-noise ratio (SNR), poor generalization, and time-consuming calibration The low SNR of EEG signals makes reliable single-trial P300 detection difficult Current systems often require multiple repetitions of each stimulus to achieve an accurate character recognition rate (CRR), which slows down spelling speed and reduces the information transfer rate (ITR) Existing methods used in P300 spellers include conventional machine learning algorithms such as stepwise linear discriminant analysis (SWLDA) [5] and support vector machines (SVMs) [6], as well as deep learning architectures such as convolutional neural networks (CNNs) [7], CapsNet with attention [8, 4], stacked Transformers [3], and image-based representations such TimeSformer [9]. However, these approaches are task- or subject-specific and fail to provide generalizable representations of the P300 response. Large interindividual differences require a time-consuming and subject-specific calibration session before the speller can be used in real-time and online [10, 11].

SpellerSSL Framework Components

The proposed SpellerSSL framework integrates self-supervised learning (SSL) with P300 aggregation strategies. The contributions include:

"We introduce a P300 aggregation strategy that enhances signal quality in the training set, enabling the model to learn more reliable single-trial responses and achieve higher character recognition rates (CRRs) with fewer repetitions"

We customize a 1D U-Net backbone as the SSL model and show that an SSL model pretrained on both cross-domain and in-domain EEG data can enhance the generalization of learned representations

"We evaluate calibration efficiency under reduced training set sizes and demonstrate that combining the aggregation strategy with a pretrained model enhances robustness, reducing calibration time without compromising accuracy in P300 speller BCIs"

Pretraining Strategy

The pretraining phase involves using a customized 1D U-Net architecture as the backbone SSL model. The model is trained for 200 epochs with a batch size of 64, using AdamW (initial learning rate 2.5 × 10−4, weight decay 1 × 10−2) and a OneCycle schedule [13]. The optimization objective is to reconstruct randomly masked EEG sequences (50% time masking, 0% channel masking). A time–frequency consistency objective is adopted, defined as LSSL = ∥x − xˆ∥1 + λ FFT(x) − FFT(ˆx). The model is pretrained on both cross-domain and in-domain EEG data to enhance generalization.

Downstream Classification and Aggregation

The downstream P300 detection is performed by fine-tuning the learned representations at the U-Net bottleneck layer with an ERP-Head classifier. The ERP-Head operates on the encoder bottleneck sequence b ∈ R1024×L16, discarding the decoder. The aggregation step is applied during fine-tuning to enhance the quality of single-trial signals, defined as averaging G consecutive repetitions. This aggregation involves Step 1: Code-aligned canonical reordering within each repetition, and Step 2: Sliding-window aggregation with group size G=2 (with overlap), where every two consecutive repetitions are averaged to form aggregated trials.

Experimental Results

The experiments evaluate three approaches: Scratch, Cross-domain pretraining, and In-domain pretraining. Evaluation metrics include Character Recognition Rate (CRR), Accuracy, Binary F1-Score, Fisher’s Discriminant Ratio (FDR), and Information Transfer Rate (ITR). In-domain SSL with moderate aggregation achieves the best overall performance, reaching 94% CRR at repetition 7 and the highest single-trial accuracy (76.18%). Cross-domain pretraining also performs competitively, reaching 93% CRR at repetition 7 and FDR = 0.85. Moderate aggregation (G=2) offers the best trade-off, effectively denoising while retaining discriminative ERP features.

Calibration Reduction Analysis

The analysis on calibration reduction demonstrates that combining the aggregation strategy with a pretrained model enhances robustness, reducing calibration time without compromising accuracy in P300 speller BCIs. For example, at 60% calibration, CRR reaches 91% at repetition 7 with FDR = 0.78, compared to only 0.41 for the scratch baseline. Overall, moderate aggregation (G=2) provides the best trade-off.

Conclusion

SpellerSSL outperforms or matches state-of-the-art methods in information transfer rate (ITR). In-domain SSL with moderate aggregation achieves the best overall results, reducing calibration requirements by up to 60% while maintaining the performance of prior models. Cross-domain SSL further demonstrates strong transferability, highlighting the potential of self-supervised EEG pretraining. These results establish SSL and aggregation as effective strategies for improving efficiency, robustness, and generalization in practical P300 speller BCIs. The paper paves the way toward future EEG foundation models.

How it works

The framework combines reconstruction-based pretraining with a lightweight ERP-Head classifier, together with a P300 aggregation strategy to improve ERP signal quality. The pretraining involves zero-padding input sequences so that output length is a multiple of 16, matching the encoder’s four successive downsamplings (each with stride 2). The aggregation step applies a sliding-window aggregation with group size G=2 (with overlap), where every two consecutive repetitions are averaged to form aggregated trials. The ERP-Head classifier operates on the encoder bottleneck sequence b ∈ R1024×L16, applying pointwise projection, depthwise temporal convolution, dilated depthwise convolution, and a final pointwise fusion. The downstream P300 detection uses the supervised cross-entropy loss Lsup = CE(z, y).

Evaluation Metrics

The evaluation metrics include Character Recognition Rate (CRR), Single-trial Accuracy, Binary F1-Score, Fisher’s Discriminant Ratio (FDR), and Information Transfer Rate (ITR). CRRs above 90% are underlined, and the highest values are highlighted in bold. The Fisher’s Discriminant Ratio (FDR) is defined as FDR = (µP300 − µNon)2 / (σ2P300 + σ2Non). Information transfer rate quantifies communication efficiency in bits/min.

Model Architectures

The backbone SSL model is a customized 1D U-Net architecture designed such that input sequences are zero-padded in time so that output length is a multiple of 16, matching the encoder’s four successive downsamplings (each with stride 2). The bottleneck feature map extracted for downstream classification is used by the ERP-Head. The ERP-Head applies a pointwise projection, followed by a depthwise temporal convolution, a dilated depthwise convolution, and a final pointwise fusion.

1. textbfImprove Signal Quality via Aggregation Strategy: The proposed P300 aggregation strategy enhances signal quality in training set by averaging consecutive repetitions, enabling the model to learn more reliable single-trial responses and achieve higher character recognition rates (CRRs) with fewer repetitions.

2. textbfEnhance Generalization with Cross-Domain Pretraining: Employing a customized 1D U-Net backbone pretrained on both cross-domain and in-domain EEG data can enhance the generalization of learned representations, allowing the model to learn effective EEG representations in both in-domain and cross-domain.

3. textbfMinimize Calibration Burden: The combination of the aggregation strategy with SSL significantly reduces calibration burden per subject, as demonstrated by a reduction by 60% while maintaining a comparable character recognition rate.