Daily Caffeine Intake Induces Concentration-Dependent Medial Temporal Plasticity in Humans: A Multimodal Double-Blind Randomized Controlled Trial

Abstract

Caffeine is commonly used to combat high sleep pressure on a daily basis. However, interference with sleep–wake regulation could disturb neural homeostasis and insufficient sleep could lead to alterations in human gray matter. Hence, in this double-blind, randomized, cross-over study, we examined the impact of 10-day caffeine (3 × 150 mg/day) on human gray matter volumes (GMVs) and cerebral blood flow (CBF) by fMRI MP-RAGE and arterial spin-labeling sequences in 20 habitual caffeine consumers, compared with 10-day placebo (3 × 150 mg/day). Sleep pressure was quantified by electroencephalographic slow-wave activity (SWA) in the previous nighttime sleep. Nonparametric voxel-based analyses revealed a significant reduction in GMV in the medial temporal lobe (mTL) after 10 days of caffeine intake compared with 10 days of placebo, voxel-wisely adjusted for CBF considering the decreased perfusion after caffeine intake compared with placebo. Larger GMV reductions were associated with higher individual concentrations of caffeine and paraxanthine. Sleep SWA was, however, neither different between conditions nor associated with caffeine-induced GMV reductions. Therefore, the data do not suggest a link between sleep depth during daily caffeine intake and changes in brain morphology. In conclusion, daily caffeine intake might induce neural plasticity in the mTL depending on individual metabolic processes.

caffeine, cerebral blood flow, gray matter, plasticity, sleep

Introduction

Caffeine is the most commonly used psychostimulant worldwide and mainly consumed in forms of coffee, tea, energy drink, and soda (Barone and Roberts 1996; Frary et al. 2005; Mitchell et al. 2014; Reyes and Cornelis 2018). Although caffeine is mostly considered to be nonaddictive, the observed physical and psychological dependence (Nehlig 1999; Mills et al. 2016) consolidate its regular consumption (Fredholm et al. 1999; Ferre 2008, 2016) through the caffeine-induced reinforcing effects (Griffiths and Woodson 1988), as well as the motive to resist withdrawal symptoms (Juliano and Griffiths 2004) and to increase alertness (Mahoney et al. 2019). Higher alertness after acute caffeine intake (Einother and Giesbrecht 2013) mirrors a reduced homeostatic sleep pressure, which is also evident in a reduced depth of sleep (Clark and Landolt 2017). This is characterized by attenuated electroencephalographic slow-wave activity (EEG SWA, 0.75–4.5 Hz) in nonrapid eye movement (NREM) sleep and shortened slow-wave sleep (Landolt et al. 1995a; Drapeau et al. 2006; for review: Urry and Landolt 2015; Clark and Landolt 2017).

Disturbed sleep homeostasis can cause not only cerebral micromorphometric alterations in the mitochondria and chromatin that leads to cell death (Abushov 2010; Zhao et al. 2016, 2017) but also macrostructural changes. Lower gray matter volumes (GMVs) were observed during abnormally high sleep pressure, such as during sleep deprivation or sleep fragmentation. Liu et al. (2014) reported reduced thalamic GMV along with impaired cognitive performance in healthy adults after 72-h prolonged waking compared with baseline. Dai et al. (2018) demonstrated a GM dynamic during a 36-h course of sleep deprivation, where a decrease in right thalamus, right insula, right inferior parietal lobe, and bilateral somatosensory association cortex was observed by 32 h of sleep deprivation. In clinical studies, GMV and cortical thickness were reduced in patients with various sleep disorders (e.g. chronic insomnia [Altena et al. 2010; Joo et al. 2013], sleep apnea [Baril et al. 2017], and narcolepsy [Joo et al. 2011]) compared with healthy controls. Together, the micro- and macro-morphometric changes in GM in response to sleep deprivation might reflect a disrupted adenosine-modulated cellular homeostasis, such as cardiac microtubule dynamic (Fassett et al. 2009), astrocytic cytoskeleton arrangement (Abbracchio et al. 2001), hippocampal fiber synaptic plasticity (Kukley et al. 2005), and robustness of cortical axons and dendrites (Ribeiro et al. 2016; Ribeiro and Sebastiao 2016).

Caffeine has been shown in animals to exert neuroprotective effects through the A2A receptor antagonism on age-related (Prediger et al. 2005), disease-related (Laurent et al. 2014), stress-associated (Kaster et al. 2015), or kainate-induced (Cognato et al. 2010) cognitive decline. However, caffeine also interferes with sleep homeostasis through antagonizing A1 and A2A receptors (Elmenhorst et al. 2012; for review: Urry and Landolt 2015). Despite of its popular use, it remains unclear whether daily caffeine consumption in humans has a long-term impact on cerebral structures through the constant impact on sleep homeostasis.

Hence, we hypothesized that, through the impacts on sleep homeostasis, daily caffeine intake alters GM structures. We measured cerebral GMV by magnetic resonance imaging (MRI) after 10 days of standardized caffeine intake (vs. placebo) in young healthy habitual caffeine consumers during a strictly controlled laboratory protocol. To control for biases on the MRI signals, we adjusted the GM responses for caffeine-induced changes in cerebral blood flow (CBF) voxel-to-voxel. As an objective measure to indicate homeostatic sleep pressure, we focused on sleep SWA (0.75–4.5 Hz) derived from frontal regions in the first NREM sleep episode (i.e. from sleep stages except REM within the first sleep cycle). This frequency band is known to be most sensitive toward the variations in sleep pressure and the effect of caffeine (Landolt et al. 1995a; Dijk et al. 1997; Werth et al. 1997; Carrier et al. 2009).

Materials and Methods

The study was approved by the ethics committee of northwest/central Switzerland (EKNZ 2016-00376). The study execution followed the declaration of Helsinki, and all participants were fully informed with study details and consented in written form.

Recruitment and Participants

Applicants aged between 18 and 35 years, body mass index (BMI) ≥18 and ≤25, nonshift workers, and without a history of transmeridian travels <1 month prior to study were screened according to the following exclusion criteria:

• self-reported caffeine intake <300 or >600 mg/day (calculations were based on Bühler et al. (2013)), adapted according to a classification of (Snel and Lorist 2011) to ensure the safety of caffeine intake and to exclude extreme responses;

• bad sleep quality, that is Pittsburgh Sleep Quality Index (PSQI) >5 to control for sleep disturbances;

• extreme chronotype, as defined by Horne-Ostberg’s Morningness-Eveningness score ≤30 or ≥70 to prevent pronounced variance in circadian phase;

• self-reported regular substance use (including medication, nicotine, and other drugs) and other major medical conditions.

A habituation night in the laboratory was conducted to exclude poor sleep efficiency (SE <70%) and clinical sleep disturbances (apnea index >10, periodic leg movements >15/h). A toxicological screening right before each laboratory session served to exclude the influence of recent drug intake including cannabis, amphetamine, methamphetamine, cocaine, benzodiazepine, and morphine.

Participants, Study Protocol, and Environmental Control

Overall, 20 healthy male participants completed the study. Here, we focus on the within-subject comparison of a 10-day caffeine and a 10-day placebo condition, which took place in randomized order (10 participants in the order of caffeine–placebo and 10 in the placebo–caffeine order; the conditions were apart minimal 11 days and no longer than 2 months). The average age was 26.4 ± 4.0 years, BMI 22.7 ± 1.38 kg/m², and self-reported daily caffeine intake was 474.1 ± 107.5 mg/day.

Each protocol consisted of 9 days of an ambulatory phase, followed by the strictly controlled laboratory stay (Fig. 1). In order to examine the effects from daily caffeine intake and to avoid withdrawal effects during the laboratory phase at abstinence (Juliano and Griffiths 2004), the treatment, caffeine (3 × 150 mg/day), or placebo (mannitol) capsules had been administered for 9 days prior to the laboratory phase. The timing of intake was set to 45 min, 4 h, and 8 h daily after waking up to imitate the pattern of caffeine intake in reality (Martyn et al. 2018). During the 9 ambulatory days, participants complied to a fixed sleep–wake cycle (8 h ± 30 min in bed, no naps allowed) to control for high sleep debt. Individual bedtimes were chosen according to usual bedtimes of each participant. The compliance to the sleep schedule was monitored by actimetry and sleep diaries. Furthermore, participants were asked to abstain from caffeine-containing diets including coffee, tea, energy drink, soda, chocolate, etc. To check for compliance to the treatments, participants were instructed to collect fingertip sweat samples during the ambulatory phase with standard cleaning procedure every evening 2 h before bedtime. Sweat from the fingertips is an emerging tool for metabolomic biomonitoring in humans (Brunmair et al. 2020).

Figure 1

Overview of the study design. Every participant underwent both a placebo and a caffeine condition. Each condition consisted of 10 days—9 days of an ambulatory phase with treatment, followed by the nighttime sleep and the laboratory day. On the 10th day, the MRI scan started at 12.75 h of EEG-monitored wakefulness (roughly 13.5 h for ASL sequence, equivalent to 5.5 h after last caffeine treatment). Levels of caffeine and paraxanthine were measured in the fingertip sweat every 2 h, and visual working memory tasks (N-back) were performed every 4 h through the laboratory phase.

In the evening of the ninth day, participants started the laboratory phase, where they stayed in dim light (<8 lux), constant half-supine position (~45°), with controlled dietary and lavatory time. Water consumption was allowed ad libitum and did not differ between conditions (see Fig. S3 in supplementary materials for more information). Access to mobile or other forms of social contacts was forbidden. The participants slept at their habitual bedtime with polysomnographic recording.

On the 10th day after waking up, the treatment (caffeine vs. placebo) continued at identical times as during the ambulatory phase (Fig. 1). The time of the protocol was adapted to the individual’s habitual bedtime. A comparison of salivary dim light melatonin onset did not indicate a significant difference in circadian timing between caffeine and placebo (see Fig. S4 for the course of salivary melatonin per condition). More comprehensive results of circadian-associated variables were reported in Weibel et al. (2020b). MRI measurement was scheduled to start at 12.75 h after awakening (roughly 13.5 h for arterial spin-labeling (ASL) sequence, equivalent to 5.5 h after last caffeine treatment). Visual working memory tasks (N-back) and measurements to assess caffeine levels and caffeine metabolites in the sweat of the fingertip were operated every 4 h and every 2 h, respectively.

Data Acquisition

1. MRI

T1-weighted structural data were obtained with a 3D magnetization-prepared rapid gradient-echo sequence (1 × 1 × 1 mm³, TR = 2000 ms, TI = 1000ms, TE=3.37ms, FA=8°) on a 3T Siemens scanner (Prisma). CBF was measured by 2D-echo-planar imaging pulsed ASL sequence (4 × 4 × 4 mm³, TR = 3000 ms, TE = 12 ms, FA = 90°, FoV = 100) at the same MR scanner.

1. Sleep EEG

Polysomnography was operated via V-Amp devices (Brain Products GmbH) with a sampling rate of 500 Hz and notch filter at 50 Hz. The recording was conducted on the identical device between conditions of each subject. Electrophysiological activity was recorded above frontal (F3, F4), central (C3, C4), and occipital (O1, O2) regions against the linked mastoids (A1, A2) as the reference electrodes. In addition we measured electro-oculography, electro-myography, to determine sleep stages, as well as electro-cardiography. A thorough investigation of sleep and circadian variables has been reported elsewhere (Weibel et al. 2020a).

1. Caffeine and metabolites from fingertip sweat

Considering the inter-individual variance in caffeine metabolism (Nehlig 2018), we measured the individual levels of caffeine and caffeine metabolites in the sweat from the fingertips during the ambulatory and laboratory phases. Samples were collected once daily during the ambulatory phase, while the laboratory phase included seven times repeated sampling in 2-h intervals from the awakening until the MRI scan. Sweat was collected using 1-cm² sampling unit. Samples were then processed and analyzed according to Brunmair et al. (2020). In short, the metabolites were extracted from the sampling unit using an acidified aqueous solution and subjected to analysis by liquid chromatography–mass spectrometry using a Q Exactive HF orbitrap hyphenated with a Vanquish UHPLC chromatography system (both ThermoFisher Scientific). The collected samples were shipped from Basel (CH) to Vienna (AT) by postal service, where they were processed.

1. N-back task

Performance in visual working memory N-back tasks was exploratorily analyzed based on the results in GM in order to test whether a concomitant change in the memory domain was present. Four sessions were completed in 4-h intervals during the 13 h of wakefulness and lasted approximately 15 min per session. Every session consisted of nine trials of 3-back (30 stimuli each) and five trials of 0-back (30 stimuli) in quasi-random order. During each session, a series of letters was presented on the computer screen. The participants were asked to press key “1” when the letter presented was the same as three stimuli before (3-back condition) or was a “K” (0-back condition), while pressing “2” when it was not. For habituation, participants performed one practice session on the ninth day evening. The differences between the caffeine and placebo conditions in the average accuracy and reaction time (RT) of four sessions are reported.

Data Preprocessing and Analyses

The pipeline of the imaging analyses in the present study consisted of three steps: 1) determining the condition effects in GM by whole-brain voxel-based analysis (VBA); 2) adjusting the GM results for the interpersonal variance of CBF voxel-to-voxel; and, finally, 3) extracting the first eigenvariates of the cluster exhibiting significant GM differences for further analyses with other covariates, that is SWA, areas under the curves of the caffeine/paraxanthine (AUC-CAPX) levels, and working memory performance.

1. MRI data for GM and CBF

For the preprocessing of T1-weighted images, we used CAT12, an extension toolbox of SPM12. We adopted the built-in pipeline for repeated measures, where volumetric measures were coregistered to the mean of the two volumes from each participant. The total intracranial volume (TIV) and the volumes of each cerebral compartment (GM, white matter, and cerebrospinal fluid) of each participant were segmented based on an affine registration on a tissue probability map in SPM12. The modulated normalization was carried out by registering each participant’s image collected in the placebo condition as a baseline onto a Montreal Neurological Institute (MNI)-defined standard brain and applying the estimation onto the image collected in caffeine condition. A Gaussian kernel of FWHM = 8 × 8 × 8 mm³ was adopted for the smoothing process.

ASL images were preprocessed with FMRIB Software Library (FSL) 5.0 developed by the Oxford Center for Functional MRI of the Brain. Motion correction was estimated, followed by the generation of M0 calibration volume and tag-control pairs to calculate relative and quantify absolute CBF maps. The final absolute CBF maps were coregistered onto the individual T1-weighted images and MNI space.

For the first step of the pipeline (a), the condition effect in total GMVs, which were estimated during the segmentation process, were tested by generalized linear mixed model (Gamma distribution) in R (R Core Team). To specify regional differences of the condition effects (caffeine vs. placebo), on GM and CBF, we used a flexible factorial model in SPM12 and pair-T (equivalent to linear mixed model) on FSL, respectively. Nonparametric threshold-free cluster enhancement (number of permutations = 5000, cluster-level threshold P < 0.01) was further performed by the SPM TFCE toolbox and FSL “randomize” function on GM and CBF, respectively. A mask of GM generated from the template was applied to reduce the bias from the correction of multiple comparisons.

Since caffeine-induced reductions in CBF can bias the MRI signal distribution (Field et al. 2003; Laurienti et al. 2003; Ge et al. 2017), for the second step (b), the VoxelStats toolbox (Mathotaarachchi et al. 2016) was applied to, voxel-to-voxel, control for the covariance of CBF and GM when estimating the coefficient of condition effects in a linear mixed model (cluster-level threshold p_FDR < 0.001).

To relate (c) condition-specific regional differences in GM to sleep, caffeine levels, and working memory performance (as specified below), we extracted the first eigenvariates of GM in the cluster exhibiting a significant change from the whole-brain analysis. We examined the associations with other variables using a linear mixed model in R (R Core Team).

All the GM models were adjusted for individual TIV.

1. EEG slow wave activity

To determine NREM, two treatment-blind experimenters (Y.-S. L., and J. W., inter-rater reliability above 85%) visually scored all nights in 30-s epoch based on the criteria from AASM (Berry et al. 2012). Artifacts were detected visually and manually rejected during the scoring. The power density of SWA during NREM sleep was quantified separately for each 0.25-Hz bin in the frequency range of 0.75–4.5 Hz by a fast Fourier transform spectrum analysis of the signal recorded over frontal electrodes (4-s spectrums over 30-s epochs averaged, window function = hamming, 0% overlapped). Results per condition and bin are illustrated in supplement (Fig. S1). The condition effect (caffeine vs. placebo) in EEG SWA was analyzed using a generalized linear mixed model on R. We used a gamma distribution model to estimate the raw EEG data instead of using Gaussian on log-transformed data to maintain the originality of the data interpretation, as the data in log-transformed scale yield a slightly different meaning from the original scale (Ng and Cribbie 2017). In addition, we determined the best-fit model between two methods by the lower Akaike information criterion and Bayesian information criterion, as well as visual examination of Q-Q plots of residuals.

1. Caffeine and paraxanthine levels

We kept all values original including the ones below the limit of detection. To estimate the individual amount of caffeine exposure, we combined the AUCs of caffeine (CA) and paraxanthine (PX), abbreviated CAPX, as paraxanthine also antagonizes adenosine at A₁, A_2A, and A_2B receptors (Snyder et al. 1981), while it follows a similar metabolic pattern as caffeine. To investigate the association of acute caffeine and paraxanthine exposure with other variables (such as GM or SWA), the AUC was calculated with the trapezoidal rule over the seven samples from the awakening until the scan. The dynamic of AUC-CAPX level over the 12.75 h is illustrated in the supplement (Fig. S2).

1. N-Back

In the n-back tasks, the “correct rate” was defined as (hits + correct rejection)/all responses + no response, while the “incorrect rate” was (missed + false alarm)/all responses + no response. Accuracy was defined as the ratio of correct rate to incorrect rate. The mean RT of overall as well as of each type of responses was calculated over nine trials of 3-back and over five trials of 0-back, respectively, in each session. The condition effect (caffeine vs. placebo) in the performance of N-back tasks was analyzed using a generalized linear mixed model (Gamma distribution) on R. The orders of the conditions were adjusted in order to reduce the impact of learning effect between conditions.

Results

Caffeine Induces a Decrease in GMV

Sweat from the fingertip was used to monitor compliance during the ambulatory and laboratory phases. During the laboratory day until the MRI scan, we found significantly higher levels in the caffeine than in the placebo condition by combining the AUCs of caffeine and paraxanthine levels (AUC-CAPX, Fig. S2; t = 13.60, P < 0.001, 95% CI of coefficients = [2.473, 3.305], R²m = 0.70, R²c = 0.85).

Total GMV was lower in the caffeine condition compared with placebo (t = −3.59; P < 0.001; 95% CI of coefficients = [−0.009, −0.003]; R²m = 0.75; R² c = 0.97; Fig. 2). A VBA indicated that the reduction of GMV was most prominent in the right medial temporal lobe (mTL, including hippocampus; VBA threshold: cluster level p_FWE < 0.05; Fig. 3a), along with the following regions at trend (i.e. 0.05 ≤ p_FWE < 0.063) of a reduction: left frontal pole, right postcentral gyrus, right insula, and the cerebellum. Both total GM and mTL GM (Fig. 4) were negatively associated with AUC-CAPX, that is the larger the reduction was in GM the higher the individual AUC-CAPX was (for statistics, see Table 1). No significant increases of GM in the caffeine compared with the placebo condition were observed.

Figure 2

The associations of total GMV with total CBF and caffeine metabolites. We use center plots to display the changes in a specific variable to the treatment of caffeine and placebo in each participant. The values are the relative distance from the responses in each condition to the average response of a single participant, calculated as response_{caff or plac}—(response_caff + response_plac)/2. Each symbol represents one participant. In two dimensions, one can observe the variance of within-subject changes of two variables between conditions (by color) as well as observe the association between the changes of two variables (x and y axes). The stronger the condition effect is, the more discretely the two-colored clouds are distributed. The stronger the association between two variables is, the closer the shape to linear is. (a) The lower total GMV was associated with higher AUC-CAPX. (b) The lower total GMV was associated with lower total CBF.

Figure 3

a) Reduction in medial temporal GM. The blue area indicated the clusters that showed a significant GM reduction in medial temporal lobe (at combined voxel-cluster-level p_uncorrected  < 0.001, p_FWE = 0.032) in the caffeine condition compared with placebo, based on a voxel-based nonparametric analysis. b) Regional differences in CBF. Regions in red indicate the reduced CBF in cuneus, precuneus, and subcortical regions (p_FWE < 0.05) after caffeine intake compared with placebo.

Figure 4

The associations of medial temporal GMVs with mTL CBF and caffeine metabolites. (a) The individual response in GMV within the significant mTL clusters was positively associated with the levels of caffeine and paraxanthine. (b) No significant association between mTL GMVs and mTL CBF was found.

Table 1

The effect of treatment and pairwise associations between all physiological variables

		Outcome variables
		Total GM	mTL GM	Total CBF	mTL CBF	SWA
Predictor variables	AUC-CAPX	Negative t = −3.5; P = 0.001 [−0.00008, −0.00002] [0.76; 0.98]	Negative t = −3.8; P = 0.001 [−0.004, −0.001] [0.42; >0.99]	Negative t = −6.3; P = 0.001 [−0.0009, −0.0005] [0.26; 0.75]	Negative t = −2.8; pAUC = 0.055 [−0.021, −0.004] [0.32; 0.63]	N.S. t = −0.8; P = 0.441 [−0.002; 0.001] [0.01; 0.74]
	Total CBF	Positive ± t = 4.70; P < 0.001 [0.004, 0.009] [0.74; 0.98]	—	—	—	—
	mTL CBF	—	N.S. t = −1.7, P = 0.122 [−0.003, 0.0002] [0.51; 0.99]	—	—	—
	SWA	Positive* t = 2.9; P = 0.004 [0.00008, 0.0004] [0.78; 0.98]	N.S. t = 0.1, P = 0.931 [−0.0004, 0.0005] [0.57; >0.99]	N.S. t = 0.2; P = 0.861 [−0.002; 0.002] [0.03; 0.72]	N.S. t = −0.1; P = 0.928 [−0.003; 0.003] [0.11; 0.57]	—

Notes: The information given in each line of each cell corresponds to Line (1) the direction of the association, (2) the t and P values, (3) the 95% CI of coefficients (noted in [lower limit, upper limit]), and (4) the effect size of the coefficients (noted in [R²margianl, R²conditional]). The R² marginal includes only the variance of the fixed factors, while R² condition includes the variance of all (fixed + random factors).

N.S. indicates that the P value is below threshold (>0.05) or when 95% CI of coefficients covers 0.

^*Statistically significant interaction (t_c-p = −4.709, P < 0.001). This positive association between SWA and total GM was significantly attenuated in the caffeine condition compared with in the placebo.

± Including total CBF as a covariate fully accounted for the main effect of condition on total GM estimation (t_CBF = 2.82, p_CBF = 0.005; t_con = −0.77, p_con = 0.441).

CBF Remains Reduced at 5.5 h after the Last Caffeine Intake

Compared with the placebo condition, total CBF was significantly reduced in the caffeine condition (t = −5.2; P < 0.001; 95% CI of coefficients = [−0.119, −0.054]; R²m = 0.12; R²c = 0.71). Voxel-wise analysis suggested that the reductions occurred prominently in the midline, including cuneus, precuneus, and subcortical regions (cluster level p_FWE < 0.05) (Fig. 3b). AUC-CAPX was negatively associated with both total CBF and mTL CBF (for statistics, see Table 1).

CBF Changes Account for Caffeine-Associated Total but Not Regional GM Differences

Examining the inter-individual variances between the changes in GM and CBF, only the total CBF and total GM were positively associated, yet no association between mTL CBF and mTL GM was observed (Table 1). By extension, total CBF predominantly accounted for the variance of total GM and divested the explanatory power of the condition effect on total GM in the two-factor model, indicating a mediation effect of CBF on the observed changes in total GM (Fig. 2). On the contrary, mTL CBF did not account for the reductions in mTL GM, determined by both the multi-modal voxel-based and ROI-based linear mixed models (Fig. 4). The detailed statistics of the effects of covariates are presented in Table 1.

No Significant Association of Sleep SWA Caffeine-Induced Regional GM Changes

In NREM SWA, we did not observe significant differences between caffeine and placebo conditions (t = −0.87, P = 0.386). However, NREM SWA was positively associated with total GM and a significant interaction between condition and SWA indicated a stronger association in the placebo condition. No significant associations between SWA and mTL GM volumetric reductions were seen (Table 1).

Exploratory Analyses Indicate Poorer Working Memory Performance during Daily Caffeine Intake

As an exploratory step to observe whether a change in the memory domain might be present concomitantly with the mTL GM changes, we tentatively inspected into the only memory-related assessment (N-back) conducted in the study. The inter-individual variance of the reductions in response accuracy was not associated with the magnitude of total nor mTL GM reduction; however, a relatively poorer working memory performance was found in the caffeine condition: compared with placebo, lower accuracy was found in both 0-back (t = −2.33, P = 0.020) and 3-back (t =−2.02, P = 0.044) performance in the caffeine condition, as well as a lower net accuracy, that is response accuracy in 3-back corrected for the baseline response in 0-back performance (t_con = −1.97, p_con = 0.049; R²m = 0.41, R²c = 0.85). Furthermore, the RT of missed (t_con = 2.75, p_con = 0.006; R²m = 0.87; R²c = 0.99) and correction rejections (t_con = 1.95, p_con = 0.051; R²m = 0.90; R²c = 0.99) were longer in caffeine condition compared with placebo, albeit no difference in the overall RT and the RT of hits and false alarm responses (t_all < 1.88, p_all > 0.60). Within the caffeine condition, a higher AUC-CAPX level was associated with a better net accuracy (t_AUC = 3.58, p_AUC < 0.001; R²m = 0.87; R²c = 0.99) and a shorter net RT in all types of responses (t_all < −2.44, p_all < 0.015; R²m_all = 0.87; R²c_all = 0.99).

Discussion

Our study examined whether daily caffeine intake affects human GM through the mediation of homeostatic sleep pressure, indexed by sleep EEG SWA. We observed a GM reduction in total volume, which was, however, artificially confounded by changes in CBF, suggesting to strictly control for caffeine intake in future MRI morphometric studies. Irrespective of CBF adjustment, a caffeine-induced concentration-dependent reduction in GMV was present specifically in a cluster within the medial temporal lobe (including hippocampus, parahippocampus, fusiform gyrus). In contrast with our expectations, nighttime sleep SWA did not explain these reductions. In other words, GM plasticity, particularly in the medial temporal region, might be induced by daily caffeine intake through a parallel pathway to its acute influence on sleep. Together with our observation on the poorer working memory performance, our data represent an important piece of knowledge regarding the impact of the most common psychostimulant on human brain structure and performance. They strongly indicate to adopt a fine-grained perspective on the divergent effects of caffeine after acute intake in comparison to daily use.

Based on the evidence of sleep-homeostatic effects on brain morphology (Liu et al. 2014; Dai et al. 2018), we assumed that daily caffeine intake may consequently affect GMVs. While the data indeed support the latter, this observed alteration did not result from a proportionate disturbance in homeostatic sleep pressure during nighttime and thus may suggest another mechanism. While caffeine reinstates the adenosine-inhibited synaptic excitatory signaling in human neurons (Kerkhofs et al. 2017) and attenuates synaptic long-term potentials (Costenla et al. 2010; Lopes et al. 2019), SWA during sleep mirrors the renormalization of synaptic capacity and recovery of brain neurons from the energy consumption during prior wakefulness (Tononi and Cirelli 2003, 2014). We speculate that, between the energy usage during wakefulness and the recovery of neurons at sleep, daily caffeine intake perhaps did not slant the balance by compromising the side of the recovery process but raising the side of energy usage. In other words, an enhanced excitation in neurons induced by caffeine might increase the need for synaptic recovery, which was, however, incompletely fulfilled in a regular sleep and even further regained by the next-day consumption. The mTL GM plasticity observed in the present study might be a consequence of the lack of commensurate synaptic restoration from the cellular stress in the long run.

A few reasons might account for the cooccurrence of a caffeine-induced response in CBF and the absence of any clear-cut changes in SWA. First of all, there was a difference in timing of measurements. CBF was measured at 5.5 h after the last intake, while the start of nighttime sleep was set to 8 h after the last intake. The previous evidence with different durations from the last intake suggests that a longer duration may attenuate the effect of caffeine on sleep (Landolt et al. 1995a, 1995b; Drapeau et al. 2006). Moreover, different subtypes of adenosine receptors (A₁R and A_2AR) may mediate caffeine-induced changes in CBF and SWA. A₁R has much stronger propensity to develop a tolerance to daily caffeine treatment compared with A_2AR (Johansson et al. 1993, 1997; Halldner et al. 2000; Popoli et al. 2000; Karcz-Kubicha et al. 2003). Thus, daily caffeine intake remains its effect in CBF through the blockade of A_2AR and A_2BR (Meno et al. 2001; Ngai et al. 2001), while a tolerance in A₁R to caffeine might contribute to the absence of changes in SWA. Notably, chronic caffeine treatment has been found to strengthen A_2AR agonism, which can downregulate the affinity of A₁R to caffeine through A_2A-A₁R heteromers (Ciruela et al. 2006; Ferre et al. 2008), therefore, potentially resulting in divergent responses modulated by each type of receptors.

To investigate the effects of daily caffeine intake on human GM, we used MRI and the classical structural MRI analyses. Within these analyses, differences in the outcome, the so-called GMV, could also be associated with changes nonneuronal cells and/or in cerebral vasculature (Zatorre et al. 2012). Such contributions seem not unlikely, as caffeine has been applied to induce apoptosis of glial cells in oncological studies (Li et al. 2017). Furthermore, the effects of adenosine and its A_2A receptors on the release of growth factors have been strongly suggested (Cunha 2016), which in turn modulate the proliferation of astrocytes and can influence angiogenesis (Fredholm 2007). Direct evidence on the effect of daily caffeine intake, however, remains to be clarified.

An MRI-derived observation on GM changes can also be simply confounded by differences in CBF (Field et al. 2003; Laurienti et al. 2003; Ge et al. 2017). Accordingly, in the present study, CBF accounted for total GM changes induced by caffeine intake. These apparent total GM changes emphasize the importance of controlling for caffeine consumption in repeated MRI measures, especially in the primarily impacted regions—cuneus, precuneus, cerebellum, and subcortices. In addition to the typical caffeine-induced reductions in CBF (Field et al. 2003; Vidyasagar et al. 2013; Merola et al. 2017), our study adds that both differences in CBF and their impact on the apparent change of morphometry still remains after 5.5 h after the caffeine intake.

Interestingly, while we observed slower RT and compromised accuracy in a visual memory task during daily intake, people who had higher caffeine and paraxanthine levels (as indicated by AUC-CAPX) within the caffeine condition showed faster RT than those with lower levels. This phenomenon might implicate two possibilities: 1) since the between-subject variance of AUC-CAPX may also reflect traits in metabolism, it could be that different metabolic patterns moderate the magnitude of caffeine’s negative impact on working memory performance; or 2) an acute higher concentration of caffeine and paraxanthine could still temporarily mitigate the worsened reaction from the long-term caffeine intake to a limited degree.

Notably, we did not find a significant association between mTL volume and the performance in visual working memory. N-Back performance, as the only available measure in memory functions in the current project, might, however, not be a suitable task to detect neurocognitive changes related to mTL GM (Jeneson and Squire 2011). In fact, performing N-back task involves majorly frontal and parietal regions (Wang et al. 2019). Moreover, a change in brain structure does not necessarily come with a difference in behavior, as differential brain activity patterns can compensate or adapt to structural changes such that behavioral performance remains unchanged (Barulli and Stern 2013; Rudrauf 2014). Therefore, we strongly encourage a further investigation in functional activities during working memory as well as other cognitive domains under daily caffeine consumption.

Increasing evidence suggests that moderate caffeine intake is beneficial and reduces the risk to develop several chronic diseases (van Dam et al. 2020). Thus, it is of importance to note that our findings on cerebral structures and visual working memory were derived from young, healthy participants in a “neutral circumstance” (i.e. no manipulation of cognitive or physical states). This may set a different basis when compared with the existing evidence. Evidence from a number of randomized controlled trials (RCTs) consistently supported acute enhancement in vigilance, especially during compromised cognitive or physiological states. In respective of habitual intake, a few longitudinal cohorts revealed, in habitual caffeine consumers older than 65 years old that the years of prior caffeine intake was associated with less age-related cognitive decline (van Gelder et al. 2007; Ritchie et al. 2007; Arab et al. 2011; Vercambre et al. 2013) or a reduced risk for dementia and Alzheimer’s disease (Lindsay et al. 2002; Eskelinen et al. 2009; Tomata et al. 2016). However, other longitudinal cohort studies did not confirm these findings (van Boxtel et al. 2003; Gelber et al. 2011; Feng et al. 2012, 2018). Moreover, a recent meta-analysis (Zhou et al. 2018) on 415 530 participants collected from international databases reported no associations between cognitive functions (memory and overall capacity) and self-reported lifetime caffeine intake or genetic variants linked with caffeine intake. Finally, RCTs investigating repeated intake in healthy population under neutral circumstance did not observe benefits in various cognitive tasks (Wing 1990; Galduróz and Carlini 1996; Judelson et al. 2005; Weibel et al. 2020b). Therefore, the positive association between habitual caffeine intake and cognitive outcomes found in some of the observational studies might not necessarily refer to causality. Nevertheless, neuroprotective effects of caffeine have been cumulatively reported in animals, especially in ameliorating symptoms of Parkinson’s disease and preventing or normalizing cognitive decline in Alzheimer’s models (Arendash et al. 2006, 2009; Chen 2014; Cunha 2016). Therefore, we strongly suggest more longitudinal RCTs to investigate the neurocognitive outcomes of habitual caffeine intake in stratified population (e.g. by genetic trait, age, sex, in healthy participants, or in patients) to characterize the adequate applications that can bring most benefits and least harm.

Our study also bears some limitations that require a careful interpretation. First, we had a relatively small sample size. However, we minimized the individual variances by the within-subject design and strictly controlled for the influence of timing and the environment. We also implemented an ambulatory phase in each condition to ensure the participant entering the laboratory phase with a standardized condition. We also coped with this potential limitation in the analyses, by employing nonparametric permutation tests in all the voxel-based analyses. Secondly, one might argue that the absence of a significant difference in SWA might be due to a genetically predisposed insensitivity to caffeine of the selected population (Retey et al. 2007). However, as the withdrawal from daily intake of 450 mg of caffeine has been shown to induce clear-cut responses in the sleep-homeostatic regulation of these participants (Weibel et al. 2020b), it seems unlikely that the selected population was insensitive to the stimulant. Notably, the habitual amount was calculated from all types of caffeinated diets, not only coffee intake but also tea, chocolate drinks, energy drinks, soda, and so on.

Overall, our findings derived from this double-blind randomized cross-over laboratory study yield an insight on the mTL GM plasticity induced by repeated intake of caffeine in a long-term course. As caffeine effects might differ with respect to the pattern of intake (sporadically vs. daily) and the brain’s state (healthy vs. diseased), we call for more well-controlled clinical trials to investigate the potential of caffeine to affect human brains and cognition. This might clarify the conditions for both detrimental and beneficial impacts of caffeine, a freely available psychostimulant all over the world.

Notes

We sincerely appreciate the contribution of our M.Sc. student Rowena Waldis, interns Andrea Schumacher and Laura Tincknell, M.Sc. student Sven Leach, and all the study helpers assisting in the experiment. We also thank Dr Corrado Garbazza and Dr Helen Slawik for the health check during screening process. We are grateful to the Core Facility of Mass Spectrometry at the Faculty of Chemistry and the Joint Metabolome Facility, members of the Vienna Life-Science Instruments. We especially appreciate all our participants for their volunteering and cooperation. Conflicts of interest: No conflicts of interest to declare.

Funding

Swiss National Science Foundation (grant 320030-163058).

Author Contributions

Y-SL: acquisition and analysis of data, draft of the manuscript and figures; JW: acquisition and preprocessing of data; H-PL: conception of the study; FS: design of the study; MM, JB, SM-M, and CG: acquisition of data; SB: design of the study and data analysis; CC: conception and design of the study; CR: conception and design of the study, draft of the manuscript.