Overview

Despite spending one-third of our lives asleep, how the brain transitions from wakefulness to sleep remains one of neuroscience's enduring mysteries. Utilising a novel computational framework, we discovered that falling asleep follows a predictable bifurcation dynamic with a distinct tipping point, validated these findings in two independent datasets, and demonstrated real-time prediction of sleep progression.

What was the approach to solving the problem?

We developed a computational framework that transforms EEG brain activity during falling asleep into a trajectory in a normalised feature space. By tracking the instantaneous distance to sleep onset, we tested whether this transition follows a bifurcation dynamic—a mathematical pattern predicted by theoretical models but never experimentally validated in the human brain.

What NSRR data were used?

We used EEG data from 1,011 participants (524 women, age 69.4 ± 9.07 years) from the Multi-Ethnic Study of Atherosclerosis (MESA) dataset available through NSRR. The cohort included participants with sleep latencies ranging from 3 to 90 minutes, providing a diverse representation of normal sleep onset patterns.

What were the results?

We discovered that falling asleep follows a predictable fold-bifurcation dynamic with a distinct tipping point occurring approximately 4.5 minutes before sleep onset. The transition shows critical slowing down—increased autocorrelation and variance before the bifurcation—similar to other critical transitions in nature. Importantly, we validated these findings in a second independent dataset and demonstrated that individual sleep transitions can be predicted in real-time with over 95% accuracy.

What were the conclusions and implications of this work?

This work provides the first experimental evidence that the human brain's transition to sleep displays bifurcation dynamics, validating decades of theoretical predictions. The framework enables real-time prediction of sleep progression and identifies a physiologically precise "tipping point" that challenges conventional sleep onset definitions. These findings could transform how we diagnose and treat sleep disorders, develop strategies to prevent unwanted sleep during critical activities like driving, and fundamentally understand one of life's most essential yet mysterious processes.

Are there any tools available?

The computational framework and analysis pipeline are openly available on GitHub with links provided in the open-access paper.

Paper Summary

Sleep is critical for physical and mental health, yet our understanding of how the brain transitions from wakefulness to sleep has remained limited. Traditional approaches classify this continuous process into discrete stages—a method that has remained largely unchanged for 60 years, despite growing recognition that falling asleep is a continuous process. Theoretical neural circuit models have long suggested that sleep-wake transitions should follow bifurcation dynamics—mathematical patterns where systems undergo rapid, nonlinear shifts between stable states. However, until now, no experimental evidence existed to validate whether the human brain actually displays such dynamics during sleep onset.

We developed a novel computational framework that represents brain EEG activity changes during sleep onset as a trajectory in a normalised, multi-dimensional feature space. The approach computes 50 comprehensive EEG features every 6 seconds and calculates "sleep distance"—the instantaneous Euclidean distance from any point to the sleep onset location. This transformation converts one-dimensional EEG voltage timeseries into a trajectory through feature space, where distance to sleep becomes the key variable.

Analysing the MESA dataset (n=1,011), we found that sleep distance remains relatively stable during early wakefulness but drops abruptly in the final minutes before sleep onset. This pattern matched a fold-bifurcation function with remarkable accuracy (R² = 0.96 at the group level), revealing a distinct tipping point at 4.5 minutes before sleep onset on average. Remarkably, 94% of participants were still classified as "awake" by traditional sleep staging when they crossed this physiological tipping point. Before the bifurcation, the sleep distance showed increased autocorrelation and variance—hallmark signs of critical slowing down that appeared approximately 4 minutes before sleep onset. The bifurcation pattern also held at the individual level across independent of sleep latency, age, or gender.

Functional principal component analysis revealed that 96% of the variance in EEG features could be explained by a single component displaying bifurcation dynamics. Features such as theta band power and spectral slope increased toward sleep, while peak beta frequency dropped from ~21 Hz to ~15.5 Hz, aligning with the known emergence of slow oscillatory activity and sleep spindles during the transition to sleep.

We validated our framework in a second independent cohort of 36 participants recorded across 267 nights. Each person's sleep onset location in feature space remained remarkably consistent across nights, with 83% of participants showing statistically indistinguishable sleep onset coordinates regardless of where they started at bedtime. Using just one night's data, we could predict the sleep distance trajectory on subsequent nights with 95% accuracy. We successfully predicted when individuals would cross their sleep tipping point with a mean error of only 0.82 minutes, demonstrating that the falling asleep process is not only mathematically predictable but also highly personalised and reproducible within individuals.

The discovery of bifurcation dynamics in human sleep onset validates decades of theoretical predictions from computational neuroscience models, showing that the "flip-flop switch" mechanism proposed from studies of sleep-wake regulatory circuits manifests in macroscopic brain dynamics measurable via EEG. This framework has several potential applications: the bifurcation tipping point provides an objective, physiologically precise definition of when sleep truly begins, which could improve diagnosis of sleep-onset disorders like insomnia and narcolepsy. Real-time monitoring of approaching tipping points could warn of dangerous drowsiness during driving or other critical activities, as the critical slowing down signals appear minutes before sleep onset. The preservation of individual sleep onset coordinates across nights enables personalised prediction and potentially personalised intervention strategies, while the framework provides a continuous, quantitative measure of sleep progression that could reveal subtle abnormalities in clinical populations or responses to interventions.

Li, J., Ilina, A., Peach, R., Wei, T., Rhodes, E., Jaramillo, V., Violante, I. R., Barahona, M., Dijk, D.-J., & Grossman, N. (2025). Falling asleep follows a predictable bifurcation dynamic. Nature Neuroscience, 1–11. https://doi.org/10.1038/s41593-025-02091-1

Blogpost Author: Dr. Junheng Li