Physiology as Language: Translating Respiration to Sleep EEG

TL;DR

This work introduces a cross-physiology translation task that synthesizes sleep EEG directly from respiration signals using a waveform-conditional generative framework. By adopting asymmetric embeddings—preserving raw respiratory context while tokenizing EEG into a discrete spectrogram vocabulary—the model employs a transformer with a masked generation objective to translate breathing dynamics into EEG tokens. Trained on 28,394 individuals across 14 sleep datasets, the method achieves a mean reconstruction MAE of about 0.068 (external) with SNR near 14 dB, and demonstrates competitive downstream performance for age estimation, sex prediction, and sleep staging compared to ground-truth EEG. The approach extends to contactless sensing by successfully synthesizing EEG from wireless RF reflections, highlighting a pathway to remote, non-contact neurological assessment during sleep, while acknowledging privacy and clinical-use considerations.

Abstract

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.

Figures (6)

• Figure 1: Model Pipeline. The model synthesizes sleep EEG from nocturnal breathing signals using an asymmetric embedding strategy to bridge the gap between modalities. The source breathing signal is processed as a raw waveform with a linear projection, while the target EEG is converted into discrete tokens by spectral transformation and vector quantization. A transformer-based model learns to translate the continuous respiratory context into the discrete neurological states using a masked generative modeling objective. During inference, the model predicts the full sequence of EEG tokens from breathing alone, which are then decoded into an EEG spectrogram.
• Figure 2: Vector Quantization for EEG. The EEG spectrogram is discretized into tokens via a codebook of distinct EEG patterns. The resolution of the token ($4$ Hz $\times$$4$ minutes per token) is chosen to align with the physiological semantics of sleep EEG.
• Figure 3: An example of sleep EEG spectrogram. Key sleep patterns are annotated, including the posterior dominant rhythm (low-frequency activity around 8-12 Hz), sleep spindles (short bursts of 12-16 Hz activity), and slow-wave activity (prominent low-frequency oscillations below 4 Hz).
• Figure 4: Visualization of EEG Reconstruction Results. In each panel, top row shows the ground-truth EEG, and bottom row shows the generated counterpart. The boxes highlight fine features in different EEG bands. These examples underscore the ability of the model to capture and replicate essential EEG features while eliminating some artifacts like the stripes in (d).
• Figure 5: Performance across Demographics and Pre-existing Conditions. Results are reported as Mean Absolute Error (MAE); lower is better. The figure shows that the MAE stays low across demographics and health conditions.