Domain-Adversarial Transformer for R-Peak Detection

Problem Formulation

EEG signals are routinely contaminated by cardiac artifacts originating from the high-energy P-QRS-T waveforms of the heart. Rather than removing these artifacts to clean the EEG, we exploit them: we aim to precisely localize R-peaks in time directly from EEG, even when cardiac interference is not visually apparent. This is motivated by research into the functional coupling between the central and autonomic nervous systems (brain-heart interplay), where accurate R-peak timing is a prerequisite for downstream analysis.

The key challenge is cross-subject generalization: cardiac artifact morphology varies substantially across individuals due to differences in electrode impedance, head anatomy, and heart position. A model trained on one subject’s EEG may fail entirely on another’s.

Dataset and Preprocessing

The study uses simultaneous 128-channel EEG and single-lead ECG recordings from 54 subjects (University of Padova). EEG signals are band-pass filtered in the [5, 22.5] Hz range — the band that maximises SNR for QRS complex detection — segmented into 10-second non-overlapping windows, and Z-scaled per channel. Ground-truth R-peak positions are extracted from the ECG using Pan-Tompkins and validated by experts. Cross-subject generalization is assessed using Leave-One-Out Cross-Validation (LOOCV).

Architecture Benchmarking

We systematically benchmarked six architectures from related domains, none of which were originally designed for this exact task:

• RPNet — a U-Net with multi-scale Inception-Residual blocks, optimized with Wing Loss for precise peak alignment
• SeizureTransformer — a hybrid CNN–Transformer U-Net combining local convolutions with global self-attention
• LaBraM — a large EEG foundation model (5.8M parameters) pretrained on 2,500+ hours of EEG data, fine-tuned with a learnable EEG adapter
• Seq2Seq Transformer — an autoregressive encoder-decoder model framing the problem as EEG-to-DT translation
• EEGDfus — a conditional diffusion model generating the distance transform conditioned on the raw EEG
• EEGNet — a compact depthwise-separable CNN used as a lightweight baseline

RPNet and SeizureTransformer achieved the strongest validation F1-scores (~63–64%), with SeizureTransformer also showing competitive timing accuracy (MAE ≈ 0.14 s).

Generalization and Pretraining Strategies

Building on the two best-performing architectures, we investigated several strategies to improve cross-subject robustness:

CLIP-style Contrastive Pretraining. We aligned EEG and ECG representations in a shared latent space using a contrastive objective, pairing a trainable EEG encoder (SeizureTransformer) with a frozen β-VAE ECG encoder. A subject-masking strategy prevents biologically related pairs from being treated as negatives. CLIP pretraining with the ECG2DT encoder improved validation F1 by 3.3% and reduced MAE by 1.6% relative to training from scratch.

Attention-based Skip Connections. We replaced standard additive skip connections in the U-Net decoder with Efficient Channel Attention (ECA) modules, yielding the best overall validation performance (F1: 64.40%, MAE: 0.1347 s).

Multi-scale Feature Extraction. We evaluated Inception, SE-Inception, and Selective Kernel (SK) modules in the encoder. The SK module — which adaptively selects receptive field sizes based on input content — achieved the best results (F1: 64.85%, MAE: 0.1353 s).

Domain Generalization. We investigated domain-adversarial training to explicitly suppress subject-specific features from the learned representations, encouraging the encoder to learn cardiac signatures that generalize across individuals.

Status

This work is the subject of my ongoing MSc thesis (Academic Year 2025/2026). Sections on domain generalization, spatial-temporal transformers, and extended MetaFormer architectures are currently under development and evaluation.