Predictive Modeling of Sleep Slow Oscillation Emergence on the electrode manifold: Toward Personalized Closed-Loop Brain Stimulation

Abstract

Background: Sleep slow oscillations (SOs), characteristic of NREM sleep, are causally tied to cognitive outcomes and the health-promoting homeostatic functions of sleep. Due to these known benefits, brain stimulation techniques aiming to enhance SOs are being developed, with great potential to contribute to clinical interventions, as they hold promise for improving sleep functions in populations with identified SO deficits (e.g., mild cognitive impairment). SO-targeting closed-loop stimulation protocols currently strive to identify SO occurrences in real time, a computationally intensive step that can lead to reduced precision (compared to post-hoc detection). These approaches are also often limited to focusing on only one electrode location, thus inherently precluding targeting of SOs that is informed by the overall organization of SOs in space-time. Prediction of SO emergence across the electrode manifold would establish an alternative to online detection, thus greatly advancing the development of personalized and flexible brain stimulation paradigms. This study presents a computational model that predicts SO occurrences at multiple locations across a night of sleep. In combination with our previous study on optimizing brain stimulation protocols using the spatiotemporal properties of SOs, this model contributes to increasing the accuracy of SO targeting in brain stimulation applications. Methods: SOs were detected in a dataset of nighttime sleep of 22 subjects (9 females), acquired with polysomnography including 64 EEG channels. Modeling of SO occurrence was achieved for SOs in stage N3, or in a combination of stages N2 and N3 (N2&N3). We study SO emergence at progressively more refined time scales. First, the cumulative SO occurrences in successive sleep cycles were successfully fit with exponentials. Secondly, the SO timing in each individual was modeled with a renewal point process. Using an inverse Gaussian model, we estimated the probability density function of SO timing and its parameters μ (mean) and λ (shape, representing skewness) in successive cycles. Results: We observed a declining trend in the SO count across sleep cycles, which we modeled using a power law relationship. The decay rate per cycle was 1.473 for N3 and 1.139 for N2&N3, with variances of the decay rates across participants being 1 and 0.53, respectively. This pattern mirrors the declining trend of slow wave activity (SWA) across sleep cycles, likely due to the inherent relationship between SWA and SO. Additionally, the SO timing model for N3 showed an increasing trend in the model parameters (μ, λ) across cycles. The increase rate per cycle followed a power law relationship with a rate of 0.83 and an exponential relationship with a rate of 4.59, respectively. The variances of the increase rates were 0.02 for μ and 0.44 for λ across participants. Conclusion: This study establishes a predictive model for SO occurrence during NREM sleep, providing insights into its organization in successive cycles and at different EEG channels, which is relevant to development of personalized stimulation paradigms. These findings imply that personalized model parameters can be estimated by incorporating SO information in the first sleep cycle, and hence SO timing can be predicted before its occurrence with a probability distribution, enabling more precise targeting of SOs.