EEG-based unsupervised learning uncovers an insomnia subtype with sleep-state misperception and associated brain and mental health risks

Abstract

Insomnia with sleep-state misperception (SSM), defined by a mismatch between subjective complaints and objective polysomnography, lacks a clear neurophysiological explanation despite its substantial clinical burden. Using an unsupervised autoencoder approach, we extracted latent EEG microstructure features and identified two reproducible insomnia subtypes across multiple datasets: an objective sleep disruption (OSD) phenotype marked by macrostructural abnormalities and an SSM phenotype presenting with near-normal polysomnography. Individuals with SSM showed reduced delta activity and elevated alpha activity during early N3 sleep, indicating shallow deep sleep and alpha intrusion. These microstructural alterations were strongly associated with clinically significant outcomes, including accelerated brain aging, impairments in attention and visual memory, and elevated depressive symptoms. Conventional SSM classifications based solely on subjective–objective discrepancy did not observe these pathophysiological abnormalities or their clinical consequences. Because consumer wearables quantify only macrostructural sleep metrics, they overlook these clinically relevant EEG features. Integrating microstructure-based analysis into portable sleep technologies may allow earlier identification of high-risk insomnia phenotypes that remain undetectable with standard approaches.