Sleep apnea polysomnography can now be synthesized: PSG-GAN generates four 32-minute signals at once, and a scorer trained only on synthetic data still reads real recordings
Synopsis
PSG-GAN moves polysomnography synthesis from short single-channel segments to long multi-signal stretches: it jointly generates 32-minute EEG, EOG, nasal pressure and photoplethysmography signals and reproduces the heart-rate changes during and after breathing obstructions. A scoring model trained from scratch on automatically scored synthetic data alone reached sleep-staging kappa of 0.71 and 0.63 and respiratory-event segmentation kappa of 0.57 and 0.60 on the healthy-volunteer and OSA-patient test sets.
Interpretation
Synthetic polysomnography is no longer confined to short segments of a single signal; four signals, EEG, EOG, nasal pressure and photoplethysmography, can be generated together at the timescale of a night recording. Earlier synthetic sleep-signal work focused on individual signals of short duration and did not cover the inter-signal dynamics that sleep-disorder scoring depends on. Real STAGES data and synthetic data had similarly distributed relative delta, theta, alpha and beta band powers in EEG and EOG across sleep stages, and automatically scored apneas had median durations of 24 s versus 22 s for real versus synthetic data.
The model learned coupling between signals, not just the waveform of each channel. Apnea scoring depends on the timing between a breathing obstruction and the heart-rate response, which single-channel generation cannot supply. Inter-beat intervals extracted from synthetic photoplethysmography followed the same slowing-then-speeding trend during and after respiratory events as real data, although median inter-beat intervals were consistently higher in synthetic data during and after events, except for apneas longer than 30 s where the medians were similar during the events.
Synthetic signals can serve directly as training data for a scorer that then works on real clinical recordings. Scoring models normally depend on manually annotated real recordings, and manual scoring is time-consuming, expensive and limited by inter-scorer agreement. A model trained from scratch only on automatically scored synthetic data reached sleep-staging kappa of 0.71 and 0.63 and one-second respiratory-event segmentation kappa of 0.57 and 0.60 on the STAR-HV and STAR-SA test sets.
Perspective
The demonstrated scope is synthesis of four signals at 64 Hz in fixed 32-minute segments, trained on the STAGES dataset and evaluated on the MESA, STAR-SA and STAR-HV real datasets; the authors propose extending the generator to variable-length segments to analyze full-night recordings and raising the sampling rate to cover the EEG gamma band up to 100 Hz.
Synthetic data differ noticeably from real data in high-frequency pulse-rate variability during REM sleep, with median relative high-frequency power of 20.7% versus 30.3%; synthetic nasal pressure inherits flat or noisy stretches from the training data, which the authors suggest may weaken respiratory-event segmentation by the synthetic model; the level of anonymity has not been quantified.
