The role of age-related sleep EEG changes in memory decline: experiments and computational modeling

Abstract

The slow oscillation (SO) observed during deep sleep is known to facilitate memory consolidation. However, the impact of age-related changes in sleep electroencephalography (EEG) oscillations and memory remains unknown. In this study, we aimed to investigate the contribution of age-related changes in sleep SO and its role in memory decline by combining EEG recordings and computational modeling. Based on the detected SO events, we found that older adults exhibit lower SO density, lower SO frequency, and longer Up and Down state durations during N3 sleep compared to young and middle-aged groups. Using a biophysically detailed thalamocortical network model, we simulated the “aged” brain as a partial loss of synaptic connections between neurons in the cortex. Our simulations showed that the changes in sleep SO properties in the “aged” brain, similar to those observed in older adults, resulting in impaired memory consolidation. Overall, this study provides mechanistic insights into how age-related changes modulate sleep SOs and memory decline.

Clinical Relevance—

This study contributes towards finding feasible biomarkers and target mechanism for designing therapy in older adults with memory deficits, such as Alzheimer’s disease patients.

I. Introduction

Humans spend approximately one-third of their lives sleeping, yet its function and the impact of its changes over a lifetime remain unknown. While sleep plays many critical roles in biological organisms, sleep-dependent long-term memory consolidation has recently become a prevalent topic of investigation. Sleep has been identified as a state that enables the consolidation of newly acquired memories [1], which is believed to depend on the replay or reactivation of the neuron ensembles engaged during learning. Replay events were most frequently observed during Slow-Wave Sleep (SWS, often referred to as deep sleep or N3 stage) in many species, including rodents [2] and humans [3].

Slow oscillations (SOs, ~1 Hz), the hallmark of SWS in humans and animals, are characterized by the coordinated activity of large neural populations consisting of alternating between depolarized (Up) and hyperpolarized (Down) states [4]. Precise temporal interaction between cortical SOs, thalamic spindles (12–16 Hz) and hippocampal ripples (~ 80–100 Hz) forms a hierarchy that plays a critical role in memory consolidation [5]. Precise coordination between SOs and spindles was reported to be impaired in older adults [6], which may negatively impact overnight memory consolidation and lead to memory deficiencies. In addition, recent studies revealed that memory function impaired with aging could be improved by acoustic stimulation applied near the Up states of sleep SOs [7] or by transcranial alternating current stimulation (tACS) [8] in older adults. Our previous modeling work has revealed intrinsic and synaptic mechanisms of memory consolidation [9] and identified a critical SO phase for acoustic stimulation to improve consolidation [10]. Despite these observations, the mechanisms underlying the changes in sleep SOs with age and their impact on the process of memory consolidation remain unknown.

Computational models make it possible to bridge the ionic current dynamics to the network modulation of sleep activities. In this study, we investigated the impact of aging on sleep SOs and memory function by combing electroencephalography (EEG) recordings and computational modeling. First, we compared SO features from EEG signals for three groups of adults: young, middle-aged, and old. Then, we built a biophysically detailed thalamocortical network with “young” and “aged” brains and discussed the effect of age-related SO change on memory consolidation. Notably, the model closely captures the distinguishing properties of “young” and “aged” brains. This study provides mechanistic insight into the changes of the sleep SOs with aging and suggests a possible pathway explaining age-related memory decline.

II. Materials and methods

A. Subjects and signals

The sleep data were obtained from the Sleep-EDF database [11], containing 197 whole-night PolySomnoGraphic sleep recordings, with EEG, EOG, chin EMG, and event markers. Corresponding hypnograms (sleep stages), manually scored by well-trained technicians according to the Rechtschaffen and Kales manual, are also available. Participants were categorized by age into young (age 25–45 years, n = 20, 11 females and 9 males), middle-age (age 46–69 years, n = 32, 20 females and 12 males), and old groups (age >= 70 years, n = 24, 11 females and 13 males). In this study, we focused on the analysis during N3 stage. Therefore, four middle-aged and five older participants without the N3 stage were excluded. We selected the first night’s sleep EEG data to analyze the sleep SOs in three different age groups.

B. Computational modeling

To investigate the changes of sleep slow oscillations in aging and effects of these changes on memory consolidation, we built a thalamocortical network model to simulate sleep SOs in “young” and “aged” brains.

Network geometry.

The thalamocortical network model incorporated 60 thalamic relays and 60 reticular neurons in the thalamus, 300 pyramidal (PY) neurons and 40 inhibitory interneurons in the cortex organized with local synaptic connectivity, as in the network geometry described in our previous studies [9][10][12]. The local field potential (LFP), which is in the same time scale of EEG data, was calculated by the summation of the neural activity of all pyramidal neurons in the cortex and filtered in the range of 0.2–20 Hz.

Neuromodulators and sleep stages.

The model implemented the change of neuromodulators, such as acetylcholine (ACh), histamine (HA), and GABA, in the intrinsic and synaptic currents to model transitions between sleep stages. Compared to the awake state, in stage N3 sleep, the levels of ACh and HA were reduced while the level of the inhibitory neurotransmitter GABA was increased [13].

Training and Test.

During the training session, we stimulated the network sequentially A->B->C->D->E, where each letter represents a group of five neurons[10][12]. For example, if the sequence started at neuron #50, these 5 groups were: A (#50–54), B (#55–59), C (#60–64), D (#65–69), E (#70–74). Each group was stimulated by a step current that led to a suprathreshold response with a duration of 10 ms and a delay of 5 ms between groups. During the test sessions (sequence recall), the model was only presented with the first input at group “A” to recall the trained sequence “ABCDE” within a 350 ms response window. During both training and test sessions, each trial was repeated every 1 s. The performance was defined by the probability of the recalled sequence to have 80% similarity to the ideal sequence ABCDE as measured during each recalled test session.

Age-related changes in the model.

In the “young” (control) brain, the radius of PY cells in the cortex is 5 and the probability of AMPA and NMDA synaptic connection between PY neurons is 1 (p=1). Some degree of brain atrophy and subsequent brain shrinkage due to synaptic damage and cell loss is common with aging. It is characterized by a progressive loss of brain volume at an estimated rate of 5% per decade after age 40 [14]. In the model, the “aged” brain is simulated by a reduced probability (p<1) of connections to model the loss in the number of AMPA and NMDA in the cortex. For example, p=0.9 represent a reduction of 10% brain mass.

C. Data analysis and Statistical analysis

Preprocessing.

Sleep EEG with electrodes based on Fpz-Cz/Pz-Oz, Fpz-Cz were selected for our study. We used MNE-Python [15] for Fpz-Cz channel EEG pre-processing. First, EEG data were resampled to 250 Hz. Second, raw EEG data were filtered with a bandpass of 0.1–30 Hz. Third, the first 30 minutes and the last 30 minutes of filtered EEG signal of wake stages were removed.

SO detection.

Event detection was performed for each channel separately based on previously established algorithms [16]. First, the raw signals were filtered between 0.5–1.25 Hz with a two-pass finite impulse response (FIR) bandpass filter; Second, after zero-crossings were detected, events were selected based on duration (0.8~2 s) and amplitude (positive peak: 75 percentile, positive peak to negative peak: 75 percentile) criteria. We calculated the density, frequency, Up and Down state durations for each detected SO event during the corresponding N3 stage.

Kernel density estimation (KDE).

We used the KDE method to estimate the distribution of SO properties. The Gaussian function was chosen as the kernel function $K$. Suppose we observe samples $x=\left\{x_1,x_2,\dots ,x_n\right\}$ from a univariate distribution. The density estimation at a point $y$ is given by:

$${\rho }_K(y)=\sum _{i=1}^{n}K\left(y-x_i;h\right)$$ (1)

In Eq. (1), $h$ was bandwidth, acts as a smoothing parameter, estimated by Scott’s rules [17].

Statistical analysis.

When data had a normal distribution, the parametric test was used; otherwise, the equivalent nonparametric test was applied. If two groups of data were compared, the two-sample t-test (parametric) or the Mann–Whitney U test (nonparametric) was used. When more than two groups of data were compared, one-way ANOVA (parametric) or Kruskal–Wallis test (nonparametric) with Bonferroni’s post hoc test was applied.

III. Results

A. The effect of aging on the sleep SOs

In this study, we analyzed single SO events, to investigate the effect of aging. We detected the SO events (0.5~1.25 Hz) based on the previously established algorithms (see Methods for details). The characteristic sleep profiles over a full night of sleep from young, middle-aged, and old adults, are shown in Fig. 1a–c. As expected, from the power spectra (Fig. 1, middle) and SO density (Fig.1, bottom) analysis, sleep SOs appeared more prominent during the N3 stage.

Figure 1.

The characteristic examples of sleep profile in the young (a), middle-aged (b), and old (c) adults. Top: raw EEG data; Middle top: color-coded power spectra of consecutive 30-s epochs; Middle bottom: sleep stages; Bottom: SO density.

For each detected single SO event in all subjects, both histogram and KDE showed the gradual elongation of Up (Fig. 2a) and Down state duration (Fig. 2b) as well as SO duration (Fig. 2c) from young to old adults. We found that young adults have significantly higher SO density compared with middle-aged and old adults, which suggests that SO density is decreased with age (Fig. 3a). For each subject, the features of SO events during N3 stage, such as SO frequency, Up and Down state duration were calculated. Three groups exhibit significant differences (one-way ANOVA with Bonferroni correction for multiple tests) in terms of SO frequency (Fig. 3b), Up (Fig. 3c) and Down (Fig. 3d) state durations. The SO frequency (the inverse of SO duration) was slower, Up and Down state duration was longer in the old adult groups compared to middle-aged and young adults, indicating that a prominent elongation of SO duration was observed with age.

Figure 2.

The histogram (left) and kernel density estimation (right, KDE, see Methods for details) of SO Up state duration (a), Down state duration (b), and SO duration (c) for young (blue), middle-aged (green) and old adults (red).

Figure 3.

The comparisons and boxplots of SO density (a), SO frequency (b), Up (c) and Down state duration (d), among three groups. P-values are calculated using the one-way ANOVA corrected by the Bonferroni’s post hoc method for multiple tests. *p<0.05, **p<0.01, ***p<0.001.

In short, during human natural sleep, we observed that the number of SOs decreased, and the duration of SOs (both Up and Down state durations) increased with age.

B. The effect of age-related brain atrophy on the sleep SOs and memory consolidation in the model

To investigate the role of age-related brain atrophy on the modulation of sleep SOs and memory consolidation, we implemented a biophysically detailed computational model of the thalamocortical network. The model consists of conductance-based neurons, distributed in one dimension (Fig. 4a, the example of PY connections in the cortex) and interconnected through synaptic dynamics updated with spike timing dependent plasticity (STDP).

Figure 4.

The simulation of “young” and “aged” brain in the model. a) The example of connectivity between PY cells with a radius of 5 in the cortex. b) Learning procedure includes test and training period in the awake state, and N3 sleep. c-d) Neural activities during N3 sleep in the “young” (p=1, c) and “aged” brain (p=0.9, d). The characteristic example of SO spatiotemporal patterns (top), LFP (middle), and single-cell activity (bottom) of neuron #50 during N3 sleep. e-f) The bar plot of performance during each recalled test session in the “young” (e) and “aged” (f) groups. Error bars indicate SEM. *p<0.05. **p<0.01. ***p<0.001, N.S. represents no significant difference.

We simulated the learning procedure (Fig. 4b) with training and test session during the awake stage similar to our previous work [10]. We compared two models representing the “young” (control) and “aged” brains. The changes in the probability of connections in the model were implemented to simulate the effect of aging (see Methods for details). In agreement with EEG data from human subjects, during N3 sleep, the characteristic examples of sleep SO activity revealed that the Up and Down state duration in LFP was elongated and the firing rate was reduced in the “aged” brain (Fig. 4c), compared with the “young” brain (Fig. 4d). The SO density per minute was significantly reduced in the “aged” brain (31.2±0.29) compared to that in “young” brain (33.5±0.41, two-sample t-test, p<0.001), consistent with human nature sleep. These observations indicate that the reduced SO activity could be explained by the brain atrophy with aging.

We observed a significant difference in the performance of a sequence recall among all three test sessions as determined by one-way ANOVA (n=10, Fig. 4e–f). In the control “young” brain (Fig. 4e), the performance was significantly higher after sleep (46%±3.5%) compared to that before sleep (31.6%±2.06%, p<0.001, after Bonferroni corrections), indicating the memory consolidation during N3 sleep. In the “aged” brain (Fig. 4f), the baseline performance before learning (9.2%±0.95%) was reduced compared to that in the control group (17.4%±1.19%). Importantly, there was no significant difference after sleep (21.4%±1.4%) compared to that before sleep (18.2%±2.16%, p=0.494, after Bonferroni corrections) in the “aged” brain (Fig. 4f), indicating the impairment of memory consolidation.

IV. Discussion

In this study, we investigated the role of age-related sleep EEG changes in memory decline by combining data from human subjects and computational modeling. Our results demonstrate that the density of SO decreases and the duration of Up and Down states increases with aging in the human sleep data. By simulating the “young” and “aged” brains, we confirmed that the density of sleep SOs is reduced because of the loss of connection between neurons in the cortex, resulting in the impairment of memory consolidation during sleep. Thus, a decrease in synaptic connectivity could be one of the critical factors contributing to memory reduction with aging. Although the experimental sleep data does not involve any memory-related tasks, previous studies have shown that age-related reduction in SWS was correlated with impaired sleep-associated memory consolidation [18]. Future studies will conduct sleep experiments related to memory across the entire age range. In addition, this study provides relevant evidence for memory decline resulting from age-related changes in sleep. Improving sleep quality may be a feasible approach to reduce memory decline caused by aging and lower the risk of developing Alzheimer’s disease in the future.