EEGInceptionERP: Inception-Based ERP Detection

• EEGInceptionERP is a deep learning framework that adapts inception architectures to extract multi-scale spatiotemporal features for robust ERP detection in EEG signals.
• It integrates transfer learning from large-scale seizure datasets and employs specialized data augmentation to handle class imbalance and low signal-to-noise ratios.
• The approach demonstrates significant performance gains under leave-one-subject-out validation, supporting clinical applications and real-time brain-computer interfaces.

EEGInceptionERP refers to deep learning methodologies for event-related potential (ERP) detection in EEG signals, leveraging inception architectures to extract multi-scale spatiotemporal features. This approach targets the challenge of robust ERP and related pathological response (e.g., Photoparoxysmal Response, PPR) detection, especially under high inter-subject variability and class imbalance scenarios common in clinical neurophysiological applications. Key frameworks include the InceptionTime architecture for time-series analysis, data augmentations tailored for clinical ERP imbalance, and transfer learning from large-scale seizure datasets to ERP targets.

1. Foundations of Inception-Based EEG and ERP Detection

The inception architecture, originally developed for image analysis, has been adapted into InceptionTime for time-series tasks. For EEG, a typical input is a 1 s window segment comprising C channels and T time points—for example, C = 18, T = 500 at 500 Hz for photosensitivity-evoked EEG (Martins et al., 31 Jan 2025). The architecture enacts a sequence of 1D inception modules, each integrating parallel branches with different convolution kernel sizes (e.g., 10, 20, 40 samples), a max-pooling branch, and channel mixing via $1\times1$ convolutions. These multi-scale branches extract discriminative temporal patterns, crucial for ERP detection which is characterized by brief, variable spatiotemporal features.

In a distinct approach for ERP-based brain-computer interfaces, three-branch inception modules are employed with temporal kernels tuned for band-specific ERP components (kernel sizes proportional to the sampling rate) (Cui et al., 2024). Following temporal convolution, linear spatial filters (akin to FBCSP) are learned, and features across branches are concatenated, enhancing frequency and spatial diversity in the representation.

2. Transfer Learning and Cross-Task Adaptation

A notable challenge in clinical ERP/PPR detection is limited labeled data for rare events. EEGInceptionERP frameworks exploit transfer learning by pre-training all layers of inception models on large, heterogeneous epilepsy datasets (e.g., CHB-MIT, 671k 1 s windows labeled SEIZURE/NORMAL, resampled to 500 Hz). Fine-tuning occurs on photosensitivity data using a leave-one-subject-out (LOSO) cross-validation scheme. During fine-tuning, early inception modules are frozen, and only the final two inception modules, the global pooling, and the output FC layer are retrained (Martins et al., 31 Jan 2025).

This domain-adaptive pipeline enables models to transfer generalizable EEG time-series features while adapting final representations to the ERP/PPR detection task, as encoded by the binary cross-entropy loss:

$$\mathcal{L} = -\frac{1}{N} \sum_{i=1}^{N} \bigl[y_i \log \hat y_i + (1-y_i)\log(1-\hat y_i)\bigr]$$

3. Data Augmentation and Class Imbalance Handling

Class imbalance poses a severe bottleneck in photosensitivity and ERP detection datasets. In the PPR context, the positive class is typically an order of magnitude rarer than the background. EEGInceptionERP uses ad-hoc data augmentation designed specifically for PPR/ERP classes (Martins et al., 31 Jan 2025):

• Four PPR window types are defined (onset-only, offset-only, full-PPR, very short-PPR).
• Synthetic windows are generated by segment-wise crossover: two real PPR epochs of the same type are split into S segments, alternately concatenated to produce a new window. Smoothing is applied at cut-points (using weighted blends over five samples), preventing high-frequency artifacts.
• The minority class is inflated to a fixed number (e.g., 3,000 windows), with majority class undersampled to achieve a 60:40 balance.

In contrast, ERP detection in BCI leverages trial-averaging augmentation during pre-training to address low SNR, with three (K=3) single trials of the same class averaged per augmentation (Cui et al., 2024).

4. Experimental Frameworks and Performance Metrics

InceptionTime-based EEG ERP detection is benchmarked across multiple pipelines:

• Preprocessing: Channel harmonization (e.g., dropping non-universal channels FT9/FT10), bipolar montage application, cubic spline resampling to target frequency, segmentation into 1 s windows (no overlap for pre-training, high overlap in fine-tuning) (Martins et al., 31 Jan 2025).
• Model trainings:
  • EXP1: InceptionTime pre-trained/fine-tuned without data augmentation.
  • EXP2: Identical to EXP1 but with augmentation and undersampling.
  • EXP3: Baseline dense-layer NN trained on 12 PCA features.
• Metrics: Accuracy, sensitivity, and specificity are computed for all models.

Table: LOSO Performance on Photosensitivity Dataset (Martins et al., 31 Jan 2025)

Model	Acc.	Sens.	Spec.
InceptionTime (no DA)	0.704±0.274	0.432±0.309	0.715±0.296
InceptionTime + DA	0.987±0.010	0.873±0.111	0.994±0.007
Dense-Layer NN (best)	0.665±0.053	0.879±0.150	0.656±0.048

Application of data augmentation not only increases sensitivity by approximately 44 percentage points but substantially reduces variance across subjects.

For single-trial ERP classification on the P300 speller task (subject-independent), an inception-based contrastive model achieved AUC 0.7233 ± 0.0750, outperforming both prior inception models without contrastive pre-training and classic architectures like EEGNet, with statistical significance at $p<0.01$ (Cui et al., 2024).

5. Training Protocols and Computational Requirements

Training an InceptionTime ensemble (five independent networks) requires substantial GPU resources—training time is on the order of hours per ensemble (fivefold cost compared to a single model). Each network is typically optimized with Adam (lr ≈ 10⁻³), batch size 64, and 150 epochs per fold (Martins et al., 31 Jan 2025). At inference, a single 1 s window is processed by each network, and the output probabilities are averaged.

In the ERP contrastive learning framework, pre-training and classifier fine-tuning are performed sequentially, with the encoder and projector frozen in classifier phase. Hyperparameters include Adam optimization (lr=1e–3/5e–4 for pre-training/fine-tuning, weight decay 0.015), epochs=100, and early stopping (patience=30) (Cui et al., 2024).

6. Clinical and Practical Deployment

EEGInceptionERP has demonstrated promising generalizability under LOSO CV in clinical settings (e.g., deployment at Burgos University Hospital). The InceptionTime ensemble, equipped with tailored data augmentation, enables real-time inference on modern edge or clinical hardware due to the moderate computational cost of forward passes (Martins et al., 31 Jan 2025).

Nevertheless, there are considerations about potential subtle domain shifts or overfitting to synthetic data introduced by heavy data augmentation. Future directions include investigation of generative data augmentation schemes and unsupervised or anomaly detection paradigms to further improve robustness without amplifying overfitting risk.

7. Relationships to Broader ERP Research

The inception-based design within EEGInceptionERP directly addresses challenges common in ERP analysis: low SNR, temporally brief events, multi-scale structure, and inter-subject variability. A plausible implication is the extension of such architectures to other time-critical or rare neurophysiological phenomena beyond PPR and classical ERP paradigms. Furthermore, contrastive representation learning combined with multi-scale temporal convolutions suggests a viable path toward domain-invariant ERP encoding, supporting both traditional classification and BCI applications (Cui et al., 2024).