Physiology as Language: Translating Respiration to Sleep EEG

This paper introduces a novel cross-physiology translation task: synthesizing sleep electroencephalography (EEG) from respiration signals. To address the significant complexity gap between the two modalities, we propose a waveform-conditional generative framework that preserves fine-grained respiratory dynamics while constraining the EEG target space through discrete tokenization. Trained on over 28,000 individuals, our model achieves a 7% Mean Absolute Error in EEG spectrogram reconstruction. Beyond reconstruction, the synthesized EEG supports downstream tasks with performance comparable to ground truth EEG on age estimation (MAE 5.0 vs. 5.1 years), sex detection (AUROC 0.81 vs. 0.82), and sleep staging (Accuracy 0.84 vs. 0.88), significantly outperforming baselines trained directly on breathing. Finally, we demonstrate that the framework generalizes to contactless sensing by synthesizing EEG from wireless radio-frequency reflections, highlighting the feasibility of remote, non-contact neurological assessment during sleep.

💡 Research Summary

This paper introduces a novel cross‑physiology translation task: generating sleep electroencephalography (EEG) from respiration signals. The authors argue that breathing and brain activity share latent information, analogous to different “languages” that encode overlapping health cues. To bridge the substantial complexity gap between the low‑frequency, mechanically driven respiratory waveform and the high‑frequency, stochastic EEG, they design an asymmetric waveform‑conditional generative framework.

Input representation – The raw respiration signal is segmented into non‑overlapping 4‑minute windows and linearly projected into continuous embeddings. No deep encoder is applied, preserving subtle morphological cues that might be lost by aggressive compression.

Target representation – EEG is first transformed into a multitaper power spectral density (PSD) spectrogram using 30‑second windows, aligning with standard sleep‑stage epoch lengths. A VQ‑GAN (vector‑quantized generative adversarial network) then discretizes the spectrogram into a sequence of tokens. Each token corresponds to a 4 Hz frequency band by 4‑minute time patch, yielding a vocabulary of 512 codebook vectors that capture canonical EEG patterns (delta, theta, alpha, sigma, beta). This tokenization reduces the high‑dimensional regression problem to a classification over a finite set of physiological “words”.

Translation model – A 12‑layer Transformer receives a concatenated sequence of continuous respiration embeddings and partially masked EEG tokens. Positional embeddings encode temporal order for respiration and a 2‑D (time‑frequency) grid for EEG tokens. During training, a random masking ratio (mean 0.55, clipped to 0.5–1.0) is applied to the EEG tokens, and the model learns to reconstruct the masked positions via a cross‑entropy loss. This masked‑generative objective, inspired by BERT, MASS, and MaskGIT, forces the network to infer EEG semantics from the surrounding respiratory context. At inference time, the EEG side is fully masked; the Transformer predicts the entire token sequence, which the frozen VQ‑GAN decoder converts back into a continuous spectrogram.

Dataset and scale – The authors aggregate 14 sleep datasets, covering 28 394 participants and 33 919 nights (≈ 1 year of recordings). This scale far exceeds prior physiological translation studies, providing a robust testbed for generalization.

Evaluation – Reconstruction quality is measured by mean absolute error (MAE) on normalized spectrogram values and signal‑to‑noise ratio (SNR). The model achieves an overall MAE of 7 % and an average SNR of ~12 dB, with particularly low errors in the delta and theta bands. No prior work exists for direct breathing‑to‑EEG reconstruction, so these metrics serve as primary baselines.

Downstream utility – The authors assess whether the synthesized EEG can replace real EEG for three clinically relevant tasks: age regression, sex classification, and sleep‑stage labeling. Models trained on generated EEG attain performance nearly identical to those trained on ground‑truth EEG (age MAE 5.0 vs 5.1 years, sex AUROC 0.81 vs 0.82, sleep‑stage accuracy 0.84 vs 0.88). In contrast, models that operate directly on respiration alone perform substantially worse (age MAE ≈ 7.2 years, sex AUROC ≈ 0.68, sleep‑stage accuracy ≈ 0.62). This demonstrates that the translation model extracts and amplifies brain‑specific information latent in breathing dynamics.

Contactless extension – Using a subset of the MGH dataset that includes wireless radio‑frequency (RF) reflections, belt‑based respiration, and EEG, the authors replace the belt signal with RF‑derived breathing estimates. The same pipeline yields an MAE of 8 % (only 1 % higher than belt‑based input), confirming that non‑contact RF sensing can feed the model and produce meaningful EEG spectrograms. This opens the possibility of remote, wearable‑free neurophysiological monitoring during sleep.

Strengths – The paper’s key contributions are (1) defining a new cross‑domain translation task, (2) introducing an asymmetric embedding and tokenization scheme that makes the problem tractable, (3) demonstrating large‑scale clinical validation, and (4) providing the first evidence that contactless RF signals can be mapped to EEG. The methodological design is well‑justified, the experiments are extensive, and the results convincingly show both reconstruction fidelity and functional utility.

Limitations and future work – The current implementation uses a single EEG channel and a relatively coarse token granularity (4‑minute patches), which may miss brief events such as K‑complexes or micro‑arousals. Scaling to multi‑channel EEG, finer token resolutions, and real‑time inference would be necessary for broader clinical adoption. Moreover, the codebook size and masking schedule could be further optimized to balance reconstruction detail against computational cost. Finally, validation on pathological populations (e.g., sleep apnea, epilepsy) would clarify the method’s diagnostic robustness.

Conclusion – By treating physiological signals as languages with distinct vocabularies, the authors successfully translate breathing waveforms into realistic sleep EEG spectrograms. The approach achieves high reconstruction accuracy, supports downstream neuro‑diagnostic tasks, and works with contactless RF sensing, suggesting a viable pathway toward scalable, comfortable, and potentially wearable‑free brain monitoring during sleep.