(C) PLOS One
This story was originally published by PLOS One and is unaltered.
. . . . . . . . . .



Longitudinal single-subject neuroimaging study reveals effects of daily environmental, physiological, and lifestyle factors on functional brain connectivity [1]

['Ana María Triana', 'Department Of Computer Science', 'School Of Science', 'Aalto University', 'Espoo', 'Department Of Neuroscience', 'Biomedical Engineering', 'Juha Salmi', 'Aalto Behavioral Laboratory', 'Aalto Neuroimaging']

Date: 2024-10

Our behavior and mental states are constantly shaped by our environment and experiences. However, little is known about the response of brain functional connectivity to environmental, physiological, and behavioral changes on different timescales, from days to months. This gives rise to an urgent need for longitudinal studies that collect high-frequency data. To this end, for a single subject, we collected 133 days of behavioral data with smartphones and wearables and performed 30 functional magnetic resonance imaging (fMRI) scans measuring attention, memory, resting state, and the effects of naturalistic stimuli. We find traces of past behavior and physiology in brain connectivity that extend up as far as 15 days. While sleep and physical activity relate to brain connectivity during cognitively demanding tasks, heart rate variability and respiration rate are more relevant for resting-state connectivity and movie-watching. This unique data set is openly accessible, offering an exceptional opportunity for further discoveries. Our results demonstrate that we should not study brain connectivity in isolation, but rather acknowledge its interdependence with the dynamics of the environment, changes in lifestyle, and short-term fluctuations such as transient illnesses or restless sleep. These results reflect a prolonged and sustained relationship between external factors and neural processes. Overall, precision mapping designs such as the one employed here can help to better understand intraindividual variability, which may explain some of the observed heterogeneity in fMRI findings. The integration of brain connectivity, physiology data and environmental cues will propel future environmental neuroscience research and support precision healthcare.

Funding: AMT acknowledges the support from the Ella and Georg Ehrnrooth foundation. https://www.ellageorg.fi/en . The Aalto Brain Centre, Aalto University provided funding for MRI data collection. https://www.aalto.fi/en/school-of-science/aalto-brain-centre . JPS acknowledges support from the Research Council of Finland (project NetResilience, grant numbers 345188 and 345183). https://www.aka.fi/en/strategic-research/ . JS acknowledges support from the Research Council of Finland (project Bringing real-life to attention, grant numbers 325981, 328954, and 353518). https://www.aka.fi/en/strategic-research/ . The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.

Data Availability: Data collected from the experiment are available for the scientific community at https://zenodo.org/records/10571956 . Since the data for this study are personal data, they are subject to the General Data Protection Regulation (GDPR). For this reason, access to the data can be requested from the institution data stewards ( [email protected] ) and a data transfer agreement contract must be stipulated between the organization who is the controller of the personal data (Aalto University) and the recipient institution. If the recipient country is not within the EU/EEA, the recipient must also provide a level of data security that is compatible with what is required by the GDPR. The data will be provided with a data processing agreement that the recipient shall also accept before initiating any data transfer. Data collected from the experiment are shared in a similar fashion across the scientific community for research purposes. All individual quantitative observations underlying the data summarized in the figures and results are available in two public repositories. Pilot data is available in the Zenodo release at https://doi.org/10.5281/zenodo.13208496 . Unprocessed study data can be found in the Zenodo dataset release at https://zenodo.org/records/10571956 . Processed results derived from the study data are accessible in the GIT repository at https://zenodo.org/doi/10.5281/zenodo.13208811 .

Copyright: © 2024 Triana et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Received: April 14, 2022; Accepted: August 8, 2024; Published: October 8, 2024 Note: As this is a Preregistered Research Article, the study design and methods were peer-reviewed before data collection. The time to acceptance includes the experimental time taken to perform the study. Learn more about Preregistered Research Articles .

All data collected were pseudonymized and released with this registered report. This experiment presents a reductionist approach as reliability and robustness are embedded in the design. By using data from one subject, we removed many sources of confounds that are common in cross-sectional designs; for example, controlling for anatomical and functional factors, weight, movement, and many more. We hope this data set will be useful to other researchers for investigating within-subject variability in brain function during tasks, alongside natural physiological variations, and for developing new methodologies for the analysis of multivariate brain and sensor data. This data set allowed us to understand how task-evoked brain activity is influenced by daily physiological, behavioral, and lifestyle factors, advancing the understanding of brain–behavior dynamics. It also allowed us to use fewer artificial stimuli while taking into account more natural situations over different timescales, which are a current challenge to current human behavior research [ 112 ]. This setting is not only important to understand how population-level findings are related to results at the individual level, but also to understand the conditions for the phenomena to be reproducible. We hope that the results from this experiment support the design of future precision-mapping studies with clinical populations for personalized medicine.

(A) One subject provided frequently sampled data over 19 weeks from 3 data sources. (i) The subject underwent MRI sessions where structural, functional, and diffusion MRI data were collected. (ii) During the 19 weeks, we also employed smartphones and wearables to collect objective data outside the scanner about the subject’s behavior. The subject also performed a short PVT and n-back test to assess cognition daily. (iii) The subject also answered questionnaires about her mental state on a regular basis. These multiple data sources allowed us to map data collected in the scanner, under laboratory conditions, with data collected outside the scanner, under quotidian situations. (B) MRI data was collected on Mondays and Fridays, depending on the availability of the scanner. During those days (in green), the participant performed a 10-min PVT in the scanner, while on the other days (pink), the participant performed a 5-min PVT at her office or home. (C) The participant was requested to answer 5 mood questionnaires every day. Two surveys were scheduled on a fixed time (yellow), while the other 3 surveys were scheduled at random between the fixed schedule (brown). (D) Timeline of data collection. All data was collected in 19 weeks. From day 1, behavioral data (green) was continuously collected using smartphones and wearables. Answers to PHQ9 [ 113 ] and PSS [ 114 ] questionnaires were also saved at the beginning of the experiment and were continuously collected each week (orange). Brain activity data collection began 2 weeks after the start of the experiment and was gathered twice per week (purple). PSS, Perceived Stress Scale; PVT, Psychomotor Vigilance Test.

Here, we intensively collected fMRI, daily behavioral, and daily physiological data from a 33-year-old female subject over 133 days (19 weeks). The subject underwent 30 fMRI sessions in which we measured functional brain connectivity while performing 4 different tasks. Such tasks include PVT [ 42 ] to assess sustained attention, resting-state, movie-watching as a naturalistic paradigm (Part 2, The Grand Budapest Hotel [ 44 ]), and dual 1-back and 2-back tests to assess working memory [ 43 ]. To take into account possible confounds in the scanning sessions, concurrent eye-tracking, heart rate, and respiration rate signals were measured in the scanner. During these 19 weeks, we also collected daily objective physiological, behavioral, and lifestyle data employing a smartphone app and 2 wearable sensors that track the subject’s signals continuously and passively, that is without the need of the subject’s input ( S1 Table provides a detailed description of these signals). Wearables and smartphone data were also collected 2 weeks before the MRI sessions began and 2 weeks after the MRI sessions ended. Fig 1 shows a description of the experimental design.

With multiple scans of a single subject, we could control for phenotypical, anatomical, and functional brain differences, so that the classification accuracy for each scanning session could then be directly related to daily fluctuations in behavioral, physiological, and lifestyle factors. As this type of analysis has not been conducted in prior studies, we kept this as an exploratory analysis. We conducted similar time-segment classification with the leave-one-out approach similar than Visconti di Oleggio Castello and colleagues [ 111 ], using fMRI movie data from N-1 days to train the classifier and testing it on the day that is left out for each of the days (runs) in the data set. The classification accuracy for each daily scanning session was then compared to other measured external factors using regression models, similarly to the other presented hypotheses H1–H4.

In machine learning approaches in neuroimaging, a classifier is often trained on data from N-1 individuals and its accuracy is then tested with the subject that is left out and the average classification accuracy is reported. However, any observed interindividual differences in classification accuracy in cross-sectional studies might depend on multiple confounding factors, which are difficult to disentangle. For example, one cannot tell if the low classification accuracy for a subject is due to the brain anatomy and function per se or whether it is confounded with a subject’s state at the time of the measurement related to behavior. These include whether the subject slept well the night before the scanning session and the subject’s phenotype (height, weight, etc.). Moreover, phenotypic factors have also been linked to in-scanner confounding variables such as spatiotemporal patterns of head motion [ 109 ], making the challenge even more difficult. Previous studies have tested the accuracy of a classifier which identified consistent spatiotemporal patterns for movie segments of 15 s using hyperalignment [ 110 ] to minimize interindividual differences and maximize classification accuracy [ 111 ]. Findings show highest classification accuracy in primary sensory areas as well as in areas related to the theory of mind, which are important during movie watching with social cues.

Following the rationale of H5 and H6, we further investigate the effects of sleep, ANS activity, and mood on resting-state functional connectivity at different timescales. We leveraged previous longitudinal findings that have demonstrated that for rs-fMRI, increased sleep regularity is associated with a more efficient network structure, better sleep quality is associated with an increase in FC, and heart rate variability is associated with the functional connectivity trajectory over time when recovering from a head injury [ 101 , 108 ]. Similarly, using data from MyConnectome, Mirchi and colleagues [ 8 ] demonstrated that resting-state functional connectivity patterns can be used to track daily fluctuations in mood, with positive mood being marked by an integrated architecture of functional connectivity.

Similar to H5, we leveraged the previous hypothesis (H2) for investigating the influence of sleep and physical activity on functional connectivity across different timescales, taking into account 3 factors. Firstly, the evidence of longitudinal effects of sleep and physical activity on improving working memory performance [ 102 , 103 ], and the evidence of a close relationship between motor and working memory development [ 104 ]. Secondly, better n-back performance has been shown to be positively correlated with increased sleep and physical activity [ 105 , 106 ]. Finally, research studying the longitudinal effects of sleep and physical activity on functional connectivity suggest that the connectivity of a set of brain regions increases when performing a working memory task following low sleep episodes [ 107 ].

We investigated how sleep affects functional connectivity on different timescales in terms of days and weeks based on 4 reasons. Firstly, previous precision functional mapping research has shown that sleep duration during the previous week is related to brain cortical thickness, and that these effects are stronger from the second to third night before measurement [ 13 ]. Secondly, it has been demonstrated that sustained attention networks are sensitive to within-subject variability at different timescales, including days and weeks [ 5 ]. Thirdly, the effects of sleep on the PVT task have been studied longitudinally, revealing a significant influence of the duration of sleep on alertness, a long-timescale modulating influence of sleep on vigilance performance, and significant co-deterioration in the sleep patterns and cognitive performance across long periods of sleep instability [ 98 – 100 ]. Finally, research on the longitudinal effects of sleep on functional connectivity has highlighted the association between poor sleep quality and a decrease in functional connectivity [ 101 ]. These findings lead us to hypothesize that variation in sleep patterns will correlate with vigilance performance, which modulates functional connectivity during attention tasks over days or even weeks. Therefore, building on H1, we chose to investigate further time-lagged cross-correlations of sleep and functional connectivity.

While the prespecification of the hypotheses H1–H4 and their analyses makes this registered report ideally suited for confirming and extending previous research results, the richness of the data collected in this study should allow for further exploratory analyses [ 97 ], in particular for studying timescales (Q2). We approach the more exploratory question Q2 through 4 specific hypotheses.

There are 3 reasons for formulating this hypothesis. Firstly, naturalistic tasks such as movie-watching offer a good trade-off for improving ecological validity and reducing vulnerability to confounds [ 88 , 89 ]. Secondly, although the effect of external factors has been less studied in these naturalistic tasks [ 88 ], there is evidence that sleep deprivation is correlated with decreased activation in the fronto-parietal and visual regions during visual selective attention tasks [ 60 , 90 ], and that physiological changes are correlated with different stimuli within the movies [ 91 , 92 ]. Regarding mood, Lyndon-Staley and colleagues [ 93 ] found that increased sadness is correlated with connectivity in the FPN and DMN. In a precision functional mapping study, Mirchi and colleagues [ 8 ] found that positive mood is related to the integration of the brain, while negative mood relates to segregation measurements. Furthermore, Nummenmaa and colleagues [ 94 ] have found that negative valence is associated with increased inter-subject correlation (ISC) in the thalamus, ventral striatum, insula, and in the DMN. Thirdly, similar to the other hypotheses, the effects of sleep and ANS were tested using wearables [ 63 , 87 , 95 ]. For mood assessment, we used ecological momentary assessments (EMAs) on smartphones [ 96 ].

We chose to study the link between sleep and ANS activity during the resting-state task in these regions because of the following reasons. Firstly, rs-fMRI paradigms are widely used to study the brain due to the low effort required from the subjects and simple signal acquisition [ 75 ]. Secondly, evidence suggests that brain activity in the absence of tasks (i.e., resting state) is affected by sleep [ 76 – 78 ], mood [ 79 ], and ANS functions (e.g., heart rate variability [ 50 ], breathing, and heart rate [ 80 – 82 ]. In fact, changes in the ANS can influence the blood oxygen-level dependent (BOLD) fMRI signal and elicit large fluctuations in the fMRI time series [ 79 , 82 ]. Thirdly, previous studies have shown several brain regions to be associated with sleep and ANS activity during the resting state. For example, functional connectivity within the DMN, insula, and intraparietal sulcus are known to be affected by sleep deprivation [ 76 , 77 ]. Moreover, functional connectivity between ACC, basal ganglia, thalamus, amygdala, midbrain, dorsolateral prefrontal cortex (dlPFC), and medial prefrontal cortex (mPFC) is correlated with heart rate variability [ 83 – 85 ]. Further, activity in the pons, thalamus, striatum, periaqueductal gray matter, hypothalamus, hippocampus, SMA, motor areas, and parietal cortices is related to breathing rate [ 86 ]. Note that most of these studies collect ANS data simultaneously with fMRI or for a very short period of time outside the scanner. This clearly differs from our study, in which we measured physiological markers of ANS activity inside and outside the scanner for a prolonged period of time. Finally, similarly to H1 and H2, sleep and ANS activity can be recorded using wearables [ 63 , 87 ].

Similarly to H1, our choices are motivated by 4 reasons. Firstly, sleep and physical activity are known to affect working memory [ 47 ]. While sleep deprivation causes deficits in working memory performance [ 64 , 65 ], physical activity improves working memory function [ 66 , 67 ]. Secondly, such working memory function can be measured using the n-back task [ 68 ] by measuring the accuracy of responses and reaction times [ 65 ]. Thirdly, previous studies have shown that during working memory tasks, connectivity in the FPN, DMN, and supplementary motor area (SMA) is affected by sleep [ 69 , 70 ] and physical activity [ 71 ]. Moreover, the type of physical activity influences activations in the anterior cingulate cortex (ACC) and SMA [ 72 ]. Finally, both external factors can be reliably measured using existing wearables [ 63 , 73 , 74 ].

We chose to investigate the relationship between sleep and attentional tasks in these regions because of 4 reasons. Firstly, results from sleep studies suggest that the amount of accumulated awake time is associated with the performance in attentional tasks [ 54 , 55 ], thus making attentional maintenance more variable and inconsistent [ 56 ]. Secondly, such variability is manifested in errors of omission (i.e., failure to respond in a timely manner or attention lapses) and errors of commission (i.e., response to stimuli that are not present). These errors can be detected by specialized, reliable, and valid tests such as the PVT [ 46 , 57 ] which is simple enough to avoid learning effects and is sensitive to sleep loss [ 58 ]. Thirdly, brain studies have shown that sleep deprivation (SD) is positively correlated with decreased activity in the prefrontal cortex (PFC), visual, parietal, and premotor areas during attention tasks [ 59 – 62 ]. Sleep loss also affects the thalamus although not uniformly [ 47 ]. In the sleep-rested state, sustained ascending arousal input from the thalamus supports a reciprocal inhibition between the fronto-parietal network (FPN) and default mode network (DMN), which becomes erratic at sleep-deprived states [ 47 ]. Finally, sleep duration, onset and offset can be reliably recorded using wearables [ 63 ].

The relationship between the aforementioned factors (i.e., sleep, physical activity, ANS activity, and mood) and functional connectivity has been investigated cross-sectionally using a wide variety of paradigms. Although the setting differs from precision functional mapping, these studies still provide valuable hints about which brain areas and external factors are generally associated at the population level, and are thus worth investigating at the individual level as well. For example, several authors have reviewed in depth the relationship between brain activity and factors, such as sleep [ 45 – 47 ], physical activity [ 48 , 49 ], ANS activity [ 50 , 51 ] (comprising respiration rate, heart rate, and heart rate variability), emotions, and mood [ 52 , 53 ]. Based on these reviews and the studies cited therein, we approached Q1 and Q2 through 8 specific hypotheses (see Table 1 ).

To address the 2 challenges discussed above, we collected a precision functional mapping data set from a single individual. This data set contains both brain activity data under a set of different fMRI tasks and objective data from external factors collected through automatic sensors. As an index of brain function, we selected functional connectivity [ 41 ] (a pattern of statistical dependencies between brain areas), as this method can be applied across all selected experimental conditions. A carefully selected set of experimental conditions, including the Psychomotor Vigilance Test (PVT) [ 42 ], an adaptive n-back [ 43 ], resting-state, and movie-watching [ 44 ] tasks, allowed us to understand the daily within-subject variability in attention, working memory, resting, and naturalistic stimuli tasks. At the same time, state-of-the-art smartphones and wearable devices made it feasible to measure (with low effort) external factors like sleep, physical activity, autonomic nervous system (ANS) activity, and mood [ 34 ]. These data offers a mean in real-life settings to answer the following research questions posed in this study:

While these studies have provided strong proof-of-concept for the benefits of repeatedly sampling the brain activity of an individual, 2 challenges remain. Firstly, although studies report dynamic changes in individuals’ brain activity at rest [ 8 , 27 , 28 , 30 ], less is known about these fluctuations during tasks. This is noteworthy, as difficult cognitive tasks can provide a means for minimizing common confounds such as head motion, for obtaining more robust brain networks, and for exploring how the current findings generalize to ecologically valid contexts [ 31 – 33 ]. Secondly, even though precision functional mapping studies have shown that external factors may also modulate brain activity [ 8 , 27 , 28 ], automatic sensors have rarely been employed to collect this type of data. Nevertheless, collecting such data is entirely feasible, owing to the development of smartphones and wearables that can be more suitable than traditional log methods for obtaining objective, quantitative data in real-life settings with minimal subject burden (see Sheikh and colleagues [ 34 ] for a review). The use of automatic sensors could not only avoid key limitations of the gold-standard behavioral methods due to subjective biases [ 35 ], but could also improve the reliability of atypical conducts [ 36 , 37 ]. In fact, dynamic analyses of these data have yielded promising results as markers for mental health disorders [ 38 – 40 ].

In recent years, this need for more trials within individuals has translated into an increasing number of studies that intensively collect brain activity data via functional magnetic resonance imaging (fMRI) from a few subjects (see Gratton and colleague’s work [ 12 ] for a full review). This approach is known as precision functional mapping [ 20 , 21 ]. Precision functional mapping studies have demonstrated that individual-specific functional connectivity differs from group averaged functional connectivity [ 22 – 25 ], and that groups of individuals who differ in behavior share common network variants (i.e., brain regions that differ from group network organization) [ 26 ]. Moreover, precision functional mapping studies have also illustrated that the picture emerging from intensive sampling of one individual’s functional brain activity can be very different from that obtained from a single session recording, after taking into account temporal variation in factors such as behavior [ 27 ], hormonal changes [ 28 ], or even migraine symptoms [ 29 ].

In light of the above, there is a clear need for longitudinal studies with frequent measurement points to study brain–behavior relationships. Moreover, analysis of longitudinal data should consider the dependence between time points, as external factors from previous days are known to be correlated with brain measurements on the following days [ 13 ]. While large sample sizes are needed in cross-sectional designs [ 14 ], small N repeated-measures designs are also desirable due to their high power and inferential validity [ 15 ]. Designs that carefully sample a small number of individuals have already been used since the 1960s [ 16 , 17 ]. For example, in psychophysics, researchers traditionally conduct numerous trials on a small number of participants [ 18 ]. However, this sampling method is still overlooked in cognitive neuroscience, where researchers strive to optimize the numbers of trials and participants in order to gain sufficient statistical power for significant group averages. Since it is commonly assumed that an individual’s mental states and cognitive abilities are somewhat invariant, just a few trials are considered sufficient for correctly sampling an individual’s brain activity and behavior. However, this assumption falls short, prompting researchers to use larger numbers of trials per participant instead [ 19 ].

Traditionally, the relationship between behavior and the state of the brain is studied with cross-sectional designs which sample many individuals at one specific point in time. Although cross-sectional designs have contributed greatly to the understanding of the brain–behavior relationship, there are concerns that studies sampling large populations over a short period of time may not translate into similar findings within single individuals [ 9 , 10 ]. More specifically, it has been shown that the variance within individual measurements can be up to 4 times greater than the variance in groups [ 11 ]. This lack of group-to-individual generalizability means that it is challenging to translate cognitive neuroscience findings into practice, as almost any treatment would benefit from personalized planning [ 12 ].

Every day, we wake up as a slightly different person, as our mental states are influenced by many external factors. The quality of sleep, the level of physical activity, and the nature of our social interactions all affect the state of our brains at different timescales. These timescales range from milliseconds (rapid detection of sounds [ 1 ]) to seconds (preparation for motor action [ 2 , 3 ]), minutes (mood changes [ 4 ]), and days (fluctuations in attentional state) [ 5 ]. Thus, different timescales reveal different aspects of brain dynamics. For example, brain areas and networks are engaged differently over time when performing a specific task [ 6 ], major psychiatric disorders show large fluctuations in brain function over different timescales [ 7 ], and functional brain connectivity patterns accurately track daily fluctuations in mood [ 8 ]. Hence, the timescales of both brain activity and external factors are important. However, few studies have considered brain activity to be not only a function of the cognitive and psychological characteristics of the sampled individual, but also a function of the specific moment in time when sampling the individual.

Results from the third pilot study suggest successful adaptation of the tasks in the MRI scanner, good tolerance of the subject to the protocol (as evidenced by FD < 0.2 for 99.2% of the time for all tasks), and reliable datastreams from wearables and smartphones. Thanks to the pilot data, we could select the preprocessing strategies stated in the analysis plan. The pilot data were also used to compute the power analysis (see Sampling Plan, Sampling size).

Results from the first 2 pilot studies demonstrated reliable datastreams from wearables and smartphone sensors. For the PVT task, data suggest there are no learning effects. For the n-back task, data suggest there are some learning effects that span no longer than 10 days.

To demonstrate the feasibility of our design, we ran 3 separate pilot studies. The first 2 pilot studies aimed to check the data quality of the sensors we were to use, and to check the possible learning effects for the PVT and n-back tasks. The third pilot study aimed to test the tasks in the scanner, the tolerance of the subject to the protocol, the MRI preprocessing strategies, and to check the sensor data quality again.

We chose an adaptive n-back test design where the perceptual load varies at each session. By balancing the perceptual load across sessions, we wanted to make sure that the task remains challenging over time and that the participant’s effort remains maximal and at a stable level across both, between and within sessions. To assure our results reflected the working memory load, we examined the influence of perceptual load and attention in the n-back task. To this end, we conducted a separate analysis, where the threshold value obtained with the adaptive n-back task was included as a confound (see Salmela and colleagues [ 43 ]). Like the PVT, we expected these results to reflect the level of attention, instead of the working memory.

Although connectivity analysis can be applied to all tasks, more traditional methods such as standard mass univariate statistics can be used to analyze the PVT and n-back tasks. Consequently, we run a supplementary analysis for H1 and H2. In this analysis, we computed the activation maps for the PVT and n-back tasks using nilearn [ 170 ]. Task-related activation was determined by identifying voxels showing a significant difference in BOLD signal, taking into account the average hemodynamic response. Event-related design matrices were generated for the PVT tasks, where we took into account the RT. Block-related design matrices were built for the n-back task. Results from the first level analysis were employed to regress the effects of behavioral and lifestyle factors according to each hypothesis. Mean FD was included as a confound. Statistical validity was assessed via FSL randomise over 10,000 permutations.

The richness of the collected data allowed several questions to be explored in different ways. Despite our intention to mainly focus on connectivity, we acknowledge some interesting analyses worth exploring. Many of these additional analyses are checkpoints to further what could be driving our results or serve as data checks. Because of their nature, these analyses are therefore reported in the Additional analysis section.

The proposed classification analysis in this manuscript differs from the original one in the following: Firstly, in the original study the TR was 1 s, i.e., the temporal distance between consecutive time segments is 1 s which is almost twice as big than the distance in this study (TR = 0.594). Because of this, we explored a few temporal segmentation approaches by shifting the sliding window with 1TR, 2TRs, 4TRs. Secondly, contrary to the original study, we did not perform hyperalignment as the subject is the same across all sessions. Thirdly, in the original study, only the cortical surface was considered after transforming the volumetric data with freesurfer. Here, we kept the data in volumetric format as it has 2 advantages: (i) we could also explore subcortical areas; and (ii) we could better control for false positives using the threshold-free cluster enhancement correction as implemented in FSL randomize since the classification accuracy for each day was correlated with other behavioral and physiological time series obtained outside the scanner. Finally, although still widely used in the neuroimaging literature, leave-one-out approaches have limitations due to potential overfitting and outliers effect [ 169 ]. For this reason, we also explored other forms of cross-validation with repeated random splits by leaving out 20% of the sessions.

Following the work of Visconti di Oleggio Catello and colleagues [ 111 ], we used the same leave-one-out classifier on our naturalistic movie-watching data. The approach is described in detail in the zenodo release [ 168 ]. Briefly, the movie fMRI time series were divided into overlapping segments of 14.85 s (25 TRs, in the original paper segments were 15 s long). The segments were obtained with a sliding window of 25 TRs, with sliding time equal to 1 TR, i.e., each segment has 24 time points overlapping with the previous segment for a total of 946 segments. Then, the classifier is implemented by calculating the pairwise similarity (correlation-based distance) between the n-th segment from the left-out movie fMRI session and the average signal from the n-th segment from the other sessions. The pairwise similarity was computed over a searchlight sphere of 10 mm radius and the segment with the highest similarity was chosen as the predicted one (chance level equals 1/946). The accuracy of the classifier was then estimated as the percentage of correctly classified segments across the whole run that is left out. The procedure was repeated for all searchlight spheres covering the whole gray matter brain areas. For each day, a brain map with values of classification accuracy was obtained. These daily maps were then correlated with external factors as described in the previous sections using linear regression and correction for multiple comparisons with FSL randomise over 10,000 permutations.

After conducting permutation models with synthetic data for each subnetwork in H5, H6, and H7, we adjusted for the number of subnetworks, assuming independence for each variable-lag pair. This means we treated variables on consecutive days as independent; for example, variable v at lag l1 is independent from a variable v at lag l2 and variables are independent. However, it is important to note that this assumption may not hold true for all variables, as autocorrelation levels and covariance can vary. For example, sleep patterns may exhibit higher autocorrelation compared to mood patterns or sleep patterns may be correlated to activity patterns. Therefore, while corrections were made for the number of networks in H5 (4), H6 (3), and H7 (3), there are still 15 lags to consider. We have provided all statistical values in the GIT (see Code availability) for transparency. Finally, multiple comparisons corrections for H4 and H8 were performed using FSL’s Randomise tool.

After conducting the permutation tests, we applied corrections for multiple comparisons. For the network estimates (global efficiency and participation coefficient) in H1, H2, and H3, we corrected for the number of networks for each variable in the model, namely 4 for H1 and 3 for H2 and H3. For the links in H1, H2, and H3, we corrected for the number of links of the chosen subnetworks, for each variable in the model.

Statistical significance was evaluated by estimating a null model with 10,000 nonparametric permutations between functional connectivity estimates and surrogate behavioral data. For each permutation, surrogate behavioral data were synthesized using spectral synthesis, by shuffling the phase of the Fourier transform of each behavioral predictor and applying the inverse Fourier transform. The surrogate behavioral data were then correlated with the subnetwork’s measure (average participation coefficient or average global efficiency) for the estimation of the null model. P-values were then obtained from the null-model probability density function and corrected for multiple comparisons with the same approach used in other analysis (false discovery rate (FDR)).

To understand what timescales behavioral, physiological, and lifestyle factors have effects on functional brain connectivity, we run a series of time-lagged cross-correlation analyses. This analysis helped us identify the relationship between functional brain connectivity estimates (participation coefficient and global efficiency) and past lags of the external factors. For each ROI, we computed the lagged cross-correlation between pairs of functional connectivity estimates and external factors, according to the hypothesis being tested. This means that, e.g., for H5, we computed the cross-correlation between total sleep duration and global efficiency for the DMN ROIs. Lagged-cross correlation coefficients were organized in carpet plots where the x-axis represented lags (in days) and y-axis represented behavioral factors. This carpet plot was organized by network, so following the example, this visualization yielded a global efficiency carpet plot for all behavioral factors and the FPN for H5 across 15 days.

Regarding H3, we included the percentage of prolonged eyes closure as a covariate, i.e., the percentage of time when the eyes have been closed for longer than 10 s (median brief microsleep episodes) [ 166 ].

We did not include performance as a confound regressor in the PVT and n-back tasks models because our main focus is the variation of task performance as reflected in functional brain connectivity. Nevertheless, to verify the relationship between task performance and behavioral factors, we run 2 standardized regression analysis between: (1) mean 1/RT [ 42 ] and total sleep time, awake time, and restless sleep for H1; and (2) accuracy and total sleep time, awake time, and restless sleep, steps, and inactive time for H2. Results from these analyses are reported in the Additional analysis section.

Similarly, we also performed a standardized regression analysis between the subnetwork measurements and the external factors selected for each hypothesis (see Table 1 ). For each model, we performed a nonparametric permutation test using 10,000 iterations under the null hypothesis of no temporal association between subnetwork measurements and external factors. We used the lmPerm R package.

To assess time-synchronous variation in functional connectivity associated with external factors, we performed a standardized regression analysis between: (1) links from the FPN, DMN, somatomotor, and cingulo-opercular networks (CONs) and total sleep time, awake time, and restless sleep for H1; (2) links from the DMN, FPN, and somatomotor areas and total sleep time, awake time, and restless sleep, steps, and inactive time for H2; and (3) links from the DMN, FPN, and CONs and total sleep time, awake time, restless sleep, heart rate variability, respiration rate, and mood computed from the selected PANAS questionnaire for H3 (see Table 1 ). For each model, we computed empirical null distributions of test statistics via 10,000 iterations of nonparametric permutation testing under the null hypothesis of no temporal association between connectivity and external factors. We used the lmPerm R package.

Here, we adapted the intersubject representational similarity analysis (IS-RSA) framework [ 88 ] to a single subject. In this case, the intuition is that days in which the neural responses are more similar should also be more similar in behavior. To compute the inter-daily representational similarity analysis (ID-RSA), we constructed 2 pairwise (i.e., day-by-day) similarity matrices for the brain and the behavioral data. Then, we assessed the significance of the comparison between these matrices via the Mantel test. The brain similarity matrix was computed using the ISC framework between ROIs defined by Seitzman and colleagues [ 158 ]. Similarly to ID-RSA, we adapted the ISC framework to a single subject by treating daily data as one subject in the ISC framework. The behavior similarity matrix was computed using 2 models, the nearest neighbors (NN), and the Anna Karenina (AK) structure [ 88 ] based on the behavioral value extracted from the daily answers to the selected PANAS questionnaire, daily total sleep time, awake time, restless sleep, breathing rate, and heart rate variability.

Both the participation coefficient and global efficiency estimates are computed on thresholded, binarized networks. Given the lack of consensus over a desirable value, we run all our models using 3 widely adopted values of proportional thresholding (10%, 20%, and 30%). These values are within the range known to produce the most consistent results [ 164 ]. To ensure a fully connected network, we first compute the maximum spanning tree of the correlation matrix, rank the links from strongest to weakest, and then add the strongest positive links to the network until reaching the chosen proportional threshold, similar to Kujala and colleagues [ 165 ]. Here, we reported results using the threshold 10%; results for other thresholds are reported in the S1 Text .

We computed 2 network measures, participation coefficient and global efficiency using the Brain Connectivity Toolbox [ 41 ]. We opted for these 2 graph metrics as they have been previously used to analyze a precision functional mapping data set [ 8 ], and they include both between and within-network measurements. The participation coefficient quantifies the relation between the number of links connecting a node outside its community and the total number of links for that particular node (i.e., between-network measurement). To estimate the participation coefficient, we used the 300 × 300 connectivity matrices and a vector of network IDs defined by Seitzman and colleagues [ 158 ]. Summary participation coefficients were obtained by computing the mean of the participation coefficients of the network nodes. Global efficiency quantifies the ease of information transfer within a network and it is defined as the average inverse shortest path length in a network [ 163 ]. To calculate the global efficiency, the 300 × 300 adjacency matrices were subdivided into smaller network-specific matrices; this means that the global efficiency was only calculated among within-network nodes. Summary global efficiency coefficients were obtained by computing the mean of the global efficiency coefficients of the network nodes. Only the measurements for the networks of interest were taken into account, according to each hypothesis (see Table 1 ).

Individual adjacency matrices were Fisher-transformed to stabilize the variance for all correlation values. Average head motion across sessions (as measured by mean FD of a session) were regressed out for each link in the network [ 28 , 162 ]. Finally, we applied the Fisher inverse transformation to the regressed adjacency matrices.

For task-fMRI (i.e., PVT and n-back), individual adjacency matrices were computed using the beta series analysis [ 161 ]. This multivariate method employs trial-to-trial variability to characterize dynamic inter-regional interactions. In short, we constructed a GLM in which every stage of every trial is modeled with a separate covariate and obtained trial-to-trial parameter estimates (beta series) for each voxel. Then, we computed the averaged beta series of the voxels belonging to an ROI and correlated these averaged beta series between ROIs. This should yield a weighted adjacency matrix of size 300 × 300.

For rs-fMRI, functional links between nodes were defined as the Pearson correlation coefficients between the averaged time series of the voxels belonging to an ROI. This should yield a weighted adjacency matrix of size 300 × 300.

Functional network nodes were defined based on the sets of brain regions defined by Seitzman and colleagues [ 158 ]. In their work, Seitzman and colleagues [ 158 ] generated new subcortex and cerebellum regions, which they later combined with previous parcellations from Power and colleagues [ 159 ] (set 1) and Gordon and colleagues [ 160 ] (set2). Set 1 comprises 300 regions of interest (ROI) and set 2 comprises 394 ROIs. We will employ set 1 to derive the main results in the manuscript. Results using set 2 are reported in S1 Text .

After the fMRIPrep preprocessing, we applied a 240-s-long Savitzky–Golay filter to remove scanner drift (similar to Çukur and colleagues [ 151 ]). To control for motion and physiological artifacts, we regressed from the BOLD time series 24 motion-related regressors, 16 signals from the WM and CSF (the signals, their derivatives, and their powers), heart and respiration rate as preprocessed by the Drifter software package [ 131 ], and motion outliers as detected by fmriprep. All confounds were also filtered using a 240-s-long Savitzky–Golay filter before regressing out their effect to avoid re-introducing artifacts [ 152 ]. Finally, the cleaned BOLD signal was filtered with a high-pass filter at 0.01 Hz cut-off frequency. No spatial smoothing was applied for the connectivity analysis, as previous research has shown that spatial smoothing affects connectivity measurements in non-uniform and systematic ways [ 153 , 154 ]. However, for task-related activation analysis with voxelwise univariate general linear model (see Additional analysis), we applied a 6 mm full width at half maximum (FWHM) kernel to smooth the data spatially using the FSL software [ 155 – 157 ].

For each of the BOLD runs found per subject (across all tasks and sessions), the following preprocessing was performed. First, a reference volume and its skull-stripped version were generated using a custom methodology of fMRIPrep. Susceptibility distortion correction (SDC) was omitted. The BOLD reference was then co-registered to the T1w reference using flirt (FSL 5.0.9 [ 143 ]) with the boundary-based registration [ 144 ] cost-function. Co-registration was configured with 9 degrees of freedom to account for distortions remaining in the BOLD reference. Head-motion parameters with respect to the BOLD reference (transformation matrices and 6 corresponding rotation and translation parameters) were estimated before any spatiotemporal filtering using mcflirt [ 145 ] (FSL 5.0.9). BOLD runs were slice-time corrected using 3dTshift from AFNI 20160207 [ 146 ] (RRID:SCR_005927). The BOLD time series (including slice-timing correction when applied) was resampled onto their original, native space by applying the transforms to correct for head-motion. These resampled BOLD time series were referred to as preprocessed BOLD in original space or just preprocessed BOLD. The BOLD time series was resampled into standard space, generating a preprocessed BOLD run in MNI152NLin6Asym space. First, a reference volume and its skull-stripped version were generated using a custom methodology of fMRIPrep. Several confounding time series were calculated based on the preprocessed BOLD: FD, DVARS, and 3 region-wise global signals. FD was computed using 2 formulations following Power (absolute sum of relative motions [ 147 ]) and Jenkinson (relative root mean square displacement between affines [ 145 ]). FD and DVARS were calculated for each functional run, both using their implementations in Nipype (following the definitions by Power and colleagues [ 147 ]). The 3 global signals were extracted within the CSF, the WM, and the whole-brain masks. Additionally, a set of physiological regressors were extracted to allow for component-based noise correction (CompCor [ 148 ]). Principal components were estimated after high-pass filtering the preprocessed BOLD time series (using a discrete cosine filter with 128 s cut-off) for the 2 CompCor variants: temporal (tCompCor) and anatomical (aCompCor). tCompCor components were then calculated from the top 2% variable voxels within the brain mask. For aCompCor, 3 probabilistic masks (CSF, WM, and combined CSF+WM) were generated in anatomical space. The implementation differs from that of Behzadi and colleagues [ 148 ] in that instead of eroding the masks by 2 pixels on BOLD space, the aCompCor masks were subtracted from a mask of pixels that likely contain a volume fraction of GM. This mask was obtained by thresholding the corresponding partial volume map at 0.05, and it ensures components are not extracted from voxels containing a minimal fraction of GM. Finally, these masks were resampled into BOLD space and binarized by thresholding at 0.99 (as in the original implementation). The head-motion estimates calculated in the correction step were also placed within the corresponding confounds file. The confound time series derived from head motion estimates and global signals were expanded with the inclusion of temporal derivatives and quadratic terms for each [ 149 ]. Frames that exceeded a threshold of 0.5 mm FD or 1.5 standardized DVARS were annotated as motion outliers. All resamplings can be performed with a single interpolation step by composing all the pertinent transformations (i.e., head-motion transform matrices, susceptibility distortion correction when available, and co-registrations to anatomical and output spaces). Gridded (volumetric) resamplings were performed using antsApplyTransforms (ANTs), configured with Lanczos interpolation to minimize the smoothing effects of other kernels [ 150 ]. Non-gridded (surface) resamplings were performed using mri_vol2surf (FreeSurfer).

The T1-weighted (T1w) images were corrected for intensity non-uniformity (INU) with N4BiasFieldCorrection [ 138 ], distributed with ANTs 2.3.3 [ 139 ] (RRID:SCR_004757), and used as T1w-reference throughout the workflow. The T1w-reference was then skull-stripped with a Nipype implementation of the antsBrainExtraction.sh workflow (from ANTs), using OASIS30ANTs as target template. Brain tissue segmentation of cerebrospinal fluid (CSF), white matter (WM), and gray matter (GM) was performed on the brain-extracted T1w using fast FSL [ 140 ] (FSL 5.0.9, RRID:SCR_002823). Volume-based spatial normalization to 2 standard spaces (MNI152NLin6Asym, MNI152NLin2009cAsym) was performed through nonlinear registration with antsRegistration (ANTs 2.3.3), using brain-extracted versions of both T1w reference and the T1w template. The following templates were selected for spatial normalization: FSL’s MNI ICBM 152 nonlinear sixth Generation Asymmetric Average Brain Stereotaxic Registration Model [ 141 ] [RRID:SCR_002823; TemplateFlow ID: MNI152NLin6Asym], ICBM 152 Nonlinear Asymmetrical template version 2009c [ 142 ] [RRID:SCR_008796; TemplateFlow ID: MNI152NLin2009cAsym].

Data from smartphones measuring battery life and screen were aggregated by calculating the mean per hour. Calls and SMS data were aggregated by day. GPS location data were pseudonymized by computing daily traveled distance and daily travel time.

Data from mood and everyday experiences were preprocessed in 3 different ways. Firstly, negative affect and positive affect scores were computed from the selected PANAS questions. Negative affect (NAF) is the sum of the answers from the afraid, nervous, upset, hostile, and ashamed questions. Positive affect (PAF) is the sum of the answers from the active, determined, attentive, inspired, and alert questions. Then, daily median NAF and PAF scores were computed. Secondly, menstruation days were recorded via free-text in the smartphone app. Finally, free-text data were not preprocessed, but used as a possible guideline into the subject’s life events.

Data from the smart ring and wristband were preprocessed automatically by the web services. For the smart ring, the data were automatically aggregated by day and did not require any further preprocessing. For the wristband, heart rate variability, and respiration rate are automatically computed by the platform and only needs to be aggregated daily. However, due to the change of device, we no longer need to compute the heart rate variability from the interbeat interval, as stated in the Stage 1 manuscript. Change accepted by the Editor on 18 January 2023.

Most behavioral data required little preprocessing. Data from the initial assessment were preprocessed by calculating the questionnaire scores according to each questionnaire’s guidelines. Scores from the cognitive tasks were computed using in-house scripts (see Code availability). The scripts calculate PVT scores according to Basner and Dinges [ 42 ] based on the reaction times provided by the Presentation software PVT script. The scripts also calculate the median reaction time, and the number of correct, wrong, and missing answers for the n-back test based on the response and reaction time provided by the Presentation software n-back script.

Due to the quality controls imposed on the data (see Methods , Quality control), we expected little to no missing data points. However, in case of missingness in the measurement of external factors (see Methods , Behavioral data), we replaced missing values with the mean of each variable where missingness is found. The imputation method was selected after trials with the pilot data (see S1 Text , Pilot data, pilot III). In case of missingness in the MRI data, we rescheduled the session, schedule permitting; the length of the overall data collection was extended accordingly. If more than 20% of the sessions needed to be rescheduled to a different hour or day, time and day of the scanning session was taken into account in a regressor.

Data were only excluded because of low quality. Behavioral data were excluded if there are more than 40% missing values over the whole time series (i.e., 19 weeks). For data measured with wearables and smartphones, we completed missing values with the mean of the data (see Analysis plan, missing data). Structural MRI data were excluded in case part of the brain in the image was missing or if extreme artifacts are found; e.g., movement or ghosting artifacts that are notorious and were not corrected by the preprocessing. rs-fMRI volumes were excluded because of high movement (i.e., volumes whose FD > 0.2). Scrubbed volumes could be different for each session of the rs-fMRI data. For movie-fMRI, we detected the volumes with high movements (FD > 0.2) in each session and censored them across the 30 sessions. Discarding volumes did not affect the sampling size (i.e., the number of scans to be taken).

As assessing power calculations for model-free stimuli (i.e., resting-state and ISC) is more difficult, we based our calculations on proposed sample sizes that have been previously employed. For resting-state fMRI, 30 scans have been used in similar precision mapping studies where fluctuations of functional connectivity across time were investigated [ 5 , 28 ]. For naturalistic stimuli, previous work [ 133 ] has established that 30 scans are within the optimal number of samples for the ISC statistics to converge.

Power analysis for (A) PVT and (B) n-back tasks computed by neuropower [ 132 ]. Power curves are shown for uncorrected values (green), random field theory (orange), and Bonferroni (blue) corrected values. Fourteen scans are necessary to achieve 95% power using Random Field Theory for both tasks. The curves also show that samples bigger than 20 scans assure enough power for all types of corrections in both tasks. MRI Pilot data available on request (see Data availability). PVT, Psychomotor Vigilance Test.

Power analysis of pilot data for PVT and n-back tasks revealed that a minimum of 14 samples are required to detect voxel activations given the presented stimuli. These 14 samples yielded 95% power, once values are corrected with random field theory. Power analysis for PVT and n-back tasks were done using second-level general linear models (GLMs) from the pilot data (see S1 Text , pilot III) and computed by neuropower [ 132 ]. Results from these analyses are shown in Fig 2 and are openly available (see Code availability).

We collected daily behavioral data for 19 weeks from 1 individual (N = 1). Simultaneously, we also collected 30 MRI scans over 15 weeks from the same individual. We opted for 30 scans as a compromise between the experiment load to the subject and sufficient number of scans for all tasks according to the power analysis.

We also checked the integrity of the eye-tracking, heart rate, and respiration rate data collected in the scanner to look for possible artifacts. In case more than 20% of the data per session were corrupted (e.g., outliers, NaNs), we investigated the causes and took corrective actions for the next session to avoid data loss.

The quality control also included checking that MRI data were transferred and stored correctly. After the correct storage, the MRI data were preprocessed using fmriprep (see Analysis plan, MRI preprocessing) and framewise displacement (FD) and temporal signal-to-noise ratio (tSNR) were investigated. Similarly to behavioral data, these checkups were only performed to take corrective actions (e.g., remind the participant to stay still in the scanner) and avoid data loss.

Due to the blinding, only basic quality control was performed during the data collection. This includes daily checks to ensure that behavioral data were transferred and stored. We only checked the data timestamps for possible missingness, i.e., no data were analyzed. In case some data were missing, we investigated the possible causes (e.g., malfunction of the devices, low battery) and took corrective actions and avoided further data loss.

To prevent any conscious or unconscious bias in the experiment, no data analysis was performed during data collection. Instead, only quality control checks were run to ensure that the data is being correctly saved (see Design, Quality control). In addition, to adhere to best practices, initial hypothesis and analysis plans were recorded in the present registered report.

Eye-tracking, breathing rate, and heart pulse signals were collected during fMRI. Eye movements were recorded during all fMRI acquisitions with an EyeLink 1000 eye-tracker (SR Research, Mississauga, Ontario, Canada). The eye-tracker was calibrated once with a five-point calibration prior to each session. Heart pulse and respiration were monitored with the Biopac system (Biopac Systems, Isla Vista, California, USA). Instantaneous values of heart rate and breathing rate were estimated with the Drifter software package [ 131 ].

Movie: The participant watched part 2 from “The Grand Budapest Hotel” by Wes Anderson [ 44 ]. Similarly to the PVT stimuli, we introduced a 30-s washout before and after the 9.5-min movie stimuli. Likewise, the washout is not included in the analysis.

Rest: a 10-min resting-state fMRI was included at each session. The first 30 s of the scanner were discarded before preprocessing to exclude drifting effects. The participant was instructed to let her mind wander while keeping the gaze fixated at a cross in the center of the screen (black background, white cross).

PVT: The participant performed a 10-min version of the behavioral PVT (see Behavioral data, Cognitive tasks, PVT) with 1 modification to adapt to the scanner conditions. To avoid noise from the scanning start, we introduced a 30-s washout before and after the PVT task. During this washout, the subject saw a centered white cross on a black background. This washout signal was not taken into account in the analysis.

All stimuli were back-projected on a semitransparent screen using a data projector (PROPixx MRI/MEG, VPixx Technologies, Saint-Bruno, Canada) and the Presentation software. Answers to the stimuli from the PVT and n-back were recorded using a 4-button diamond keypad (RESPONSEPixx/MRI, Pixx Technologies, Saint-Bruno, Canada).

Full-brain BOLD images were acquired in an interleaved fashion using gradient-echo-planar imaging with fat suppression, multiband acceleration factor 4, TR 594 ms, TE 16 ms, flip angle 50, 64 × 64 matrix, 44 axial slices, slice thickness 3 mm, 3 × 3 mm in plane resolution, anterior-posterior phase encoding. This sequence was used to record brain data from the subject while performing the PVT, resting-state, movie-watching, and n-back tasks, in that order. We opted to fix the order of scans to make sure that the same condition (PVT, resting state, movie, and n-back) would be comparable between different measurement sessions. Another option would have been to counterbalance the order of scans to account for potential effects changing over the whole session. However, a possible drawback of using counterbalancing could be that the amount of variability brought by the same task sometimes being conducted first and sometimes last (reflecting, for example, the level of arousal during the scanning session) could be greater than the variability of the external factors of interest. In addition, the fixed-order design has been previously used in massive data collection projects such as the Human Connectome Project [ 129 , 130 ]. While counterbalancing could be used with a larger number of scanning sessions by treating it as an external confound, in single-participant studies, fixing the order of scans could increase the likelihood in detecting brain behavior interactions while keeping the number of scanning sessions reasonably low.

MRI was performed in a fixed schedule, Mondays and Fridays at 1,130 hours, subject to the scanner availability. Imaging was performed on a Siemens MAGNETOM Skyra 3T MRI scanner (Siemens Healthcare, Erlangen, Germany) at the Advanced Magnetic Imaging Center, Aalto University, using a 30-channel head coil. At each session, the following data was collected in the same order: localizer, PVT, resting-state, movie watching, N-back, and structural T1-weighted imaging. At the first and last session, structural T2-weighted and diffusion MRI data were also gathered after the structural T1-weighted imaging (see Fig 1 ), although we did not analyze these in this study.

After each fMRI session, the subject completed a questionnaire about the experience in the scanner. Using a 5-point Likert scale, the subject rated the engagement on each task. Additionally, the subject also wrote short freeform accounts on the thoughts while performing rs-fMRI and movie-watching tasks.

Additional passive data were collected using the AWARE app. In short, data from battery levels, screen use, calls, messages (SMS), and global positioning system (GPS) were collected. However, these data are not used in the current experiment.

Mood and everyday experiences were recorded using the Aware Framework application ( https://awareframework.com/ ) and the koota service [ 127 ]. The application is freely available for Android and iPhone. It is a tool to collect mobile context information by sensor instrumentation capable of gathering data passively (i.e., no user-input required) and actively (i.e., requires input from the user). We used the app to collect the participant’s answers to 3 questionnaires, morning, evening, and mood. Morning questionnaires were asked at 10:00 AM and inquiry about the sleep on the previous night. Evening questionnaires were asked at 11:00 PM and included questions about social interactions, exercise, and alcohol, caffeine, and theine consumption. In addition, mood was monitored using the short version of the international PANAS [ 128 ] that was asked with the morning and evening questionnaires, plus at 3 random times between 10:00 AM and 11:00 PM. In addition, a free-text box is available with the selected PANAS questionnaire in case the participant would like to provide further contextual information. Finally, the participant answered weekly PHQ9 and PSS questionnaires. The complete set of questions is available in the S2 Table .

ANS activity was measured using a wrist monitor (EmbracePlus, Empatica, Empatica, Boston, Massachusetts, USA)**. The device is a medical-grade wearable that measures physiological data via photoplethysmography (PPG), accelerometry, and electrodermal activity. EmbracePlus is the successor of the E4 wristband, which has been validated for measurement of heart rate and heart rate variability [ 87 , 125 , 126 ] under different conditions. The E4 wristband was originally registered in the protocol. However, Empatica no longer sold the device when the data collection was to start. Therefore, to ensure the reproducibility of this protocol, we decided to employ the EmbracePlus device. We received the Editorial approval for this change on 18 January 2023. We used the monitor to track respiration rate and heart rate variability.

Sleep and physical activity were measured using the Oura ring (Oura Health Ltd, Oulu, Finland). The smartring is a commercially-off-the-shelf device that has been validated against actigraphy for sleep measurements under free-living conditions [ 63 ]. The ring has also been used in physical activity studies under normal life conditions [ 73 , 124 ]. It tracks 55 metrics that are daily aggregated. However, many of the metrics are highly correlated with each other and therefore, we only used a subset of them in our study (see S1 Text , experiment data, selection of features from wearables). We used the ring to track daily total sleep time, awake time, restless sleep, steps, and inactive time. Originally, we chose to use daily total sleep time, sleep efficiency, sleep latency, steps, and inactive time measurements based on pilot data (see S1 Text , pilot III, selection of features from smartring). However, during data quality check-ups (i.e., after finishing the data collection and before the initial analysis), we noticed a high correlation between other sleep measurements and sleep latency and sleep efficiency. This led us to reconsider the variables to employ. For a full description, please see S1 Text , experiment data, selection of features from wearables.

Two wearables and 1 smartphone app were used to obtain objective quantitative data from external factors with minimal subject burden. External factors such as sleep, physical activity, autonomic nervous system activity, and mood were measured. The participant wore the devices at all times, except when charging the devices was required or when it was not allowed (e.g., going to the pool for some devices).

The grayscale gratings orientation ranged between 0° and 360° in steps depending on the task performance. Each grating had a spatial frequency of 2 c/deg and was displayed in a Gaussian envelope (diameter 3°). The grating orientation varied according to the participant’s performance in a staircase procedure. The lower part of the grating was kept at the center and the grating rotated. The maximum change is constrained to 90°. Complex colored textures were used as visual distractors (size 16 × 24°). For consistency across trials, we cut off a circular area of 6° from the center of the textures. The root mean square contrast of the visual distractors is 0.3.

The stimulus pairs were presented in onset-to-onset intervals of 1,800 ms and their duration was 300 ms. The frequency of the auditory stimuli ranged between 600 and 1,800 Hz in steps according to the participant’s task performance in a staircase procedure. The maximum change in pitch was limited to 0.5 octaves. The auditory distractors were band-pass filtered at 200 and 7,000 Hz and notch-filtered at 1,000 Hz with a two-octave wide filter. The sound maximum intensity was 80 dB SPL.

On each trial, auditory tones and visual gratings were presented, and either tone pitch or grating orientation changed compared to the previous trial. On 1/3 of the trial, either visual or auditory distractors concurred with the task-relevant stimuli. Visual distractors were spectrally complex textures and auditory distractors were spectrally complex sounds (e.g., car honks). The participant’s task was to indicate with a button press, in which modality and to which direction the stimulus changed with respect to the previous trial (1-back) or to the 2 trials before (2-back). When there was a change in the auditory stimulus, the participant should have pressed “up” if the current pitch was higher, or “down” if the current pitch was lower, compared to the n trial. When the change was in the visual stimuli, the participant should have pressed “right” if the current grating orientation rotated clockwise, or “left” if the current grating orientation rotated counterclockwise, compared to the n-th trial. Only one modality changed per trial. The difficulty of the task was adapted to the participant’s responses, i.e., with more accurate answers, the changes became more subtle. This adaptation aimed to keep the rate of correct responses at 70.7% with an adaptive staircase method based on trials with no distractors. A schematic of this process is shown in the S1B Fig .

Dual n-back: The participant performed an adaptive dual n-back task as to the one used by Salo and colleagues [ 43 ] for 1 and 2 back tasks in 8 alternating blocks of 20 trials each. At the beginning of each block, instructions showed which n-back task the participant should perform (1-back or 2-back). Then, the participant was presented with synchronous sinewave tones and sinewave gratings with occasional auditory or visual distractors. At the end of each block, general feedback was displayed. All instructions and feedback were one line of text in English. Stimuli and procedures have been previously described in other papers [ 43 , 123 ].

Based on previous results [ 42 ], this task has shown good performance in the partial sleep deprivation condition, with large effect sizes (d > 0.8) for the median 1/RT, number of lapses, and performance score for PVTs longer than 5 min. In fact, Basner and colleagues [ 121 ] have demonstrated that 3 to 5 min is usually sufficient to measure the effects of sleep on vigilance. Two versions of the PVT were employed, a 10-min version which was used in the scanner, and a 5-min version which was employed every day except for those days where the 10-min PVT was used (see Fig 1B ). We used a longer PVT task in the scanner because usually in fMRI, longer versions are preferred to collect more volumes. Nevertheless, using a 5-min and a 10-min PVT task should not pose any problems given the 5-min PVT is a reliable substitute for the 10-min PVT [ 122 ].

PVT: The participant performed a PVT task [ 42 ]. In short, the participant was instructed to observe a red rectangular box on the computer screen and to press a button as soon as a yellow counter appears on it. When the button was pressed, the counter stopped and it was subsequently reset. The period between the last response and the new stimulus varied between 2 and 10 s. The participant was instructed to press the button as soon as the yellow counter appeared, so that the reaction time (RT) was as low as possible, but avoiding lapses (i.e., press the button when no stimulus is yet displayed). A schematic of the PVT stimuli is shown in the S1A Fig .

Two cognitive tasks were employed in this study, a PVT and a dual n-back. Cognitive performance from these tasks was measured daily during the 19 weeks. Both tasks were shown using the Presentation software (Neurobehavioral Systems, Albany, California, United States of America) and their code is available (see Methods , Code availability).

On the first day of the experiment, the participant answered a battery of questionnaires, which includes the Big Five Inventory [ 115 – 117 ], Perceived Stress Scale (PSS) [ 114 ], Patient Health Questionnaire (PHQ9) [ 113 ], Generalized Anxiety Disorder 7 Item Scale (GAD-7) [ 118 , 119 ], and Pittsburgh Sleep Quality Index (PSQI) [ 120 ]. These allowed us to establish baseline metrics.

The participant (author AMT) is a right-handed female, aged 33 years for the majority of the study. The participant has no history of severe psychiatric disorders or any neurological disorders. To date, she is generally healthy with no history of chronic medical conditions. She is not prescribed special medication, except oral contraceptives (Levonorgestrel and Ethinylestradiol), which she has taken for more than a year. She is active, cycles regularly, and exercises 5 times a week (mean total energy expenditure 1.467, mean daily calories burned 2416.14, based on the pilot data, see S1 Text , pilot data, pilot III). She does not have any dietary restrictions, does not smoke, has never used any recreational drugs, and rarely consumes alcohol. Her native language is Spanish, but she also uses English at a professional level.

In this single-subject correlational study, the participant underwent 133 days (19 weeks) of data collection in total. On the first day, the participant signed the consent forms, answered a battery of questionnaires (see Design, Behavioral data) and the behavioral data collection began. To establish a baseline, only behavioral data was collected during the first 2 weeks (for details, please see Design, Behavioral data). The MRI data collection began in the third week. There were 2 MRI sessions per week over 15 weeks, where we collected structural and functional MRI (please see Design, Brain imaging data acquisition for details). The order of the acquisitions and tasks did not change in the MRI sessions. However, during the first and last session of the experiment, diffusion MRI and T2-weighted images were collected. After the 15 weeks of MRI data collection, the subject collected 2 more weeks of behavioral data. An overview of the experiment is shown in Fig 1 . The Stage 1 protocol is available in the Open Science Framework https://osf.io/5hu9c/ .

This study was approved by the Aalto University Research Ethics Committee (28th of June, 2019, Helsinki, Finland) and was carried out in accordance with the principles expressed in the Declaration of Helsinki. Two updates were required to include new collaborators, adjust the timeline and to include the privacy notice for the study. Both updates were accepted by the Aalto University Research Ethics Committee. Written informed consent was obtained from the participant. As approved by the Committee, the participant did not receive any compensation for this study.

Standardized regression analysis between the PVT’s mean 1/RT and the total sleep time, sleep efficiency, and sleep latency did not yield any significant results. Similarly, we found no significant results for the regression analysis of the n-back’s accuracy and the total sleep time, sleep efficiency, sleep latency, steps, and inactive time.

Following the proposed methods, we computed 3 additional analyses on the task-fMRI data. These computed mass univariate statistics for: (i) the sustained attention task; (ii) the working memory task; and (iii) the working memory task including the perceptual load. We found a significant association between the previous day’s sleep duration and the BOLD signal during the PVT task (t = 7.39, p < 0.05, MNI coordinates x = 55.04, y = −58.54, z = 19.89) ( S42 Fig ). No other significant results were found.

All the mentioned analyses employed a LOO cross-validation for the decoding maps. While each LOO analysis took up to 8 hours, a leave-20% cross-validation would require an impractical 33 days. Therefore, supplementary analyses using alternative cross-validation methods were not conducted. Nevertheless, we analyzed some centroids of ROIs and compared these findings with our main classification accuracy maps. The high correlation between these results suggest that different cross-validation methods would likely yield similar results ( S41 Fig ).

Supplementary analyses, including longer sliding windows and global signal regression, demonstrated that the maximum respiration rate from the prior 8 and 11 days is correlated with the classification accuracy ( S38 – S40 Figs and for full statistics and peak MNI coordinates, see S29 – S31 Tables). In fact, the maximum respiration rate experienced 8 days prior exhibited greater activation in the right postcentral gyrus ( Fig 9C ) across all approaches. Conversely, the maximum respiration rate experienced 11 days before showed consistent brain activation in the right precentral gyrus only when the global signal was regressed from the data ( S40D Fig and S31 Table ). Nevertheless, for this particular variable-lag combination, consistent activation in the right postcentral gyrus was found in longer sliding windows ( S38D and S39E Figs), but not in the main analysis. Supplementary analyses also revealed a consistent association in the right superior temporal gyrus between the classification maps and the standard deviation of the positive affect experienced 2 days prior, across different sliding windows.

The classification accuracy is related to: (A) standard deviation of the positive affect at lag 2, (B) maximum respiratory rate at lag 5, (C) maximum respiratory rate at lag 8, (D) maximum respiratory rate at lag 11, and (E) total sleep duration lag 15. Plots were generated with nilearn [ 170 ]. Unprocessed study data can be found in the Zenodo data set release [ 175 ]. Processed results derived from the study data are accessible in the GIT repository [ 176 ], under the results folder.

To understand the relationship between brain activity and past external factors such as sleep, mood, and ANS activity, during movie-watching stimuli, we run a series of regression models on classification accuracy brain maps. These maps represent the accuracy of a classifier to correctly identify specific brain patterns based on the highest similarity observed between segments of different sessions. Regression results reveal that the classification accuracy is significantly correlated with past positive affect (t = 6.55, p < 0.05, MNI coordinates x = 66, y = −30, z = 8), respiration rate, and sleep duration (t = 5.66, p < 0.05, MNI coordinates x = 24, y = −76, z = 42) ( Fig 9 ). In particular, the maximum respiration rate experienced 5 (t = 5.47, p < 0.05, MNI coordinates x = −40, y = −66, z = −14), 8 (t = 6.62, p < 0.05, MNI coordinates x = 54, y = −12, z = 40), and 11 days (t = 5.76, p < 0.05, MNI coordinates x = −48, y = −74, z = 28) prior correlates with the classification accuracy in the left fusiform, the right postcentral gyrus, the right precentral gyrus, and the middle occipital gyrus.

Similar to H5 and H6, we re-analyzed the resting-state data using other thresholds ( S34 and S35 Figs and S25 and S26 Tables), parcellation ( S36 Fig and S27 Table ), and including the global signal as a regressor in the fMRI data preprocessing stage ( S37 Fig and S28 Table ). Fig 8 shows the variable-lag pairs that we found statistically significant using at least 2 different analysis approaches. Results from re-analyses indicate high stability in time-lagged correlations of ANS factors. Specifically, correlations between the DMN’s global efficiency and the respiration rate 7 and 13 days prior seem consistently significant across all analysis variants. A similar pattern is observed for the FPN’s participation coefficient and the previous day’s HRV, showing significant correlations in every analysis. Overall, these findings suggest that for rs-connectivity, there is an important relationship between HRV on both within- and between-network integration, occurring within a timeframe of 7 to 13 days prior.

Notably, most of the dynamic correlations between ANS factors and the DMN’s global efficiency are negative (see S26 Table ). This suggests that lower respiratory rates and HRV may be linked to more efficient information transfer within the DMN ( Fig 8A ). Conversely, the majority of time-lagged correlations between the ANS activity and the networks’ participation coefficients are positive ( Fig 8D, 8E and 8F ). Likewise, we noted positive short-lag correlations between mood factors and the CON’s connectivity. In particular, more negative feelings (lag 5, ρ = 0.61, p < 0.05 and lag 6, ρ = 0.48, p < 0.05) and stress levels (lag 7, ρ = 0.51, p < 0.05) are linked to higher CON’s within and between network integration ( Fig 8C and 8F ).

In contrast, correlations between the DMN’s participation coefficients and past ANS activity seem less dominant. Instead, factors like awake time in bed from 15 days prior (ρ = −0.44, p < 0.05) and recent mood changes, up to 2 days before (negative affect, lag 2, ρ = −0.4, p < 0.05 and positive affect, lag 3, ρ = 0.53, p < 0.01), appear more influential in the DMN ( Fig 8D ). Similarly, the FPN’s participation coefficient shows a strong, short-lag correlation with mood effects and HRV ( Fig 8E ). A similar trend is seen in the CON’s participation coefficient, especially with short-lag correlations. For example, sleep patterns, specifically sleep duration and restlessness from the past 3 days, suggest a correlation with the CON’s between-network estimate ( Fig 8F ). All correlation and p-values are reported in the S26 Table . Therefore, we observed that the between-network integration seems more closely linked to shorter cycles of sleep, mood, and ANS activity, compared to the within-network estimate.

Patterns from 2 to 14 days in the past are correlated with the global efficiency in the DMN (A), fronto-parietal (B), and cingulo-opercular (C) networks. Similarly, sleep, activity and ANS activity patterns from the previous day up to 15 days in the past are correlated with the participation coefficient in the DMN (D), fronto-parietal (E), and cingulo-opercular (F) networks. Significant correlations are shown by an asterisk (*). The data was also analyzed using different thresholds, parcellations, and regressing the global signal. Squares show correlations that we found significant in at least 2 analyses. External factor measurements are correlated with whole-brain connectivity at different time lags, which suggests that different aspects of the individual’s sleep, mood, and ANS history have specific, time-sensitive relationships with distinct brain networks. Unprocessed study data can be found in the Zenodo data set release [ 175 ]. Processed results derived from the study data are accessible in the GIT repository [ 176 ], under the results folder. ANS, autonomic nervous system; DMN, default mode network.

Our results suggest that the DMN’s global efficiency predominantly correlates with ANS activity (respiratory rate and HRV) experienced in the prior days, but not with prior sleep factors ( Fig 8A ). All correlation and p-values are reported in the S25 Table . Similarly, the FPN’s global efficiency is mostly associated with prior respiratory rate (lag 2, ρ = 0.5, p < 0.01 and lag 3, ρ = −0.46, p < 0.05), although sleep hours from 10 days earlier (ρ = 0.42, p < 0.05) also bear a significant correlation ( Fig 8B ). Instead, positive affect lags are significant factors influencing the CON’s global efficiency (lag 2, ρ = 0.4, p < 0.05 and lag 8, ρ = 0.4, p < 0.05) ( Fig 8C ). These patterns indicate that while the DMN and FPN’s global efficiency may be dynamically linked with ANS activity, mood factors may play a more substantial role in influencing the CON within-network integration.

To investigate the time-lagged relationships between external factors and resting-state brain network estimates, we ran a series of cross-correlation analyses. We assessed significance by comparing our results with 10,000 correlations based on surrogate data. We also corrected for multiple comparisons, assuming each tuple of variable-lag is independent, i.e., 3 networks for each variable-lag tuple.

Upon re-analyzing the data with other thresholds ( S30 and S31 Figs and S21 and S22 Tables), another parcellation ( S32 Fig and S23 Table ), and including the global signal as a regressor ( S33 Fig and S24 Table ), we observed several significant time-lagged cross-correlations not found in the principal analysis, especially for the FPN’s participation coefficient. Fig 7 shows the stability of the correlations upon re-analyses of the data. Nevertheless, we found that several time-lagged correlations remained stable across different analyses. For example, the DMN’s global efficiency showed repeated time-lag correlations with awake time and inactivity. Similarly, the network’s participation coefficient consistently related to dynamic sleep and activity patterns. These results suggest that in working-memory tasks, an individual’s previous sleep and, more importantly, physical activity steadily correlates both the within- and between-network integration of the DMN and somatomotor network. The correlation appears to follow short cycles of approximately 3 days and longer cycles of about 11 days.

For time lags exceeding 7 days, our results show correlations between awake time, number of steps, and inactive time with the global efficiency in the DMN ( Fig 7A ) and somatomotor network ( Fig 7C ). For full statistics (correlation and p-values), see S21 Table . We also observed negative long-lag correlations between sleep patterns and the participation coefficient of the DMN (lag 15, ρ = 0.43, p < 0.05) ( Fig 7D ) and somatomotor network (lag 15, ρ = 0.37, p < 0.05) ( Fig 7F ). Moreover, the number of steps and inactive time experienced 11 to 12 days prior are associated with the participation coefficient in the FPN (ρ = 0.49, p < 0.01) ( Fig 7E ) and somatomotor network (ρ = 0.5, p < 0.01) ( Fig 7F ). These results suggest that reduced physical activity and sleep hours from more than a week ago correlates with decreased number of links between the FPN, somatomotor areas, and other networks compared to the links inside each network. Notably, many of these significant correlations occur around the eighth-day or fourteenth-day period, indicating a consistent pattern aligned with 1 and 2-week intervals.

Similarly, we noted significant short-lag correlations between the number of steps taken 5 days earlier and the participation coefficient of the DMN (ρ = −0.37, p < 0.05) ( Fig 7D ), FPN (ρ = −0.48, p < 0.01) ( Fig 7E ), and somatomotor network (ρ = −0.54, p < 0.01) ( Fig 5F ). These negative correlations suggest that reduced step count 5 days prior is associated with increased between-network estimate of the 3 aforementioned networks. In contrast, sleep duration 7 days prior showed a significant positive correlation with the participation coefficient in the DMN (ρ = 0.4, p < 0.05) and somatomotor (ρ = 0.4, p < 0.05) networks.

Sleep and activity patterns from 2 to 14 days in the past are correlated with the global efficiency in the DMN (A), and somatomotor networks (C). Similarly, sleep and activity patterns from 5 to 15 days in the past are correlated with the participation coefficient in the DMN (D), fronto-parietal (E), and somatomotor (F) networks. Significant correlations are shown by an asterisk (*). The data was also analyzed using different thresholds, parcellations, and regressing the global signal. Squares show correlations that we found significant in at least 2 analyses. Sleep and activity measurements are correlated with whole-brain connectivity at varying time lags, which suggests that different aspects of sleep and activity history have specific, time-sensitive relationships with distinct brain network functionalities. Unprocessed study data can be found in the Zenodo data set release [ 175 ]. Processed results derived from the study data are accessible in the GIT repository [ 176 ], under the results folder. DMN, default mode network.

Results from these analyses revealed correlations for short lags (i.e., less than 7 days) in awake time, number of steps, and inactive time with global efficiency in the DMN ( Fig 7A ) and somatomotor network ( Fig 7C ). For full statistics (correlation and p-values), see S21 Table . In particular, the inactive time (ρ = 0.43, p < 0.05) and number of steps (ρ = −0.42, p < 0.05) taken on the third day prior were found to be related to the DMN’s global efficiency.

Using the same approach as H5, we investigated the relationship between external factors, such as past sleep and activity patterns, and brain connectivity estimates. These estimates are computed for the DMN, FPN, and somatomotor network when undergoing working memory tasks. We also corrected for multiple comparisons, assuming each tuple of variable-lag is independent, i.e., 3 networks for each variable-lag tuple.

We found different significant time-lagged cross-correlations when analyzing the data with other proportional thresholds ( S26 and S27 Figs and S17 and S18 Tables), another parcellation ( S28 Fig and S19 Table ), and when including the global signal as a regressor ( S29 Fig and S20 Table ). Fig 6 shows the variable-lag tuples that were found to be significant in at least 2 separate analyses of the data. Few lags were consistent across all analyses. In particular, we observed that the DMN’s and FPN’s global efficiency consistently correlates with short lags of awake time and long lags of sleep duration, while their participation coefficient remained unaffected. Conversely, the CON’s and somatomotor network’s participation coefficient is associated with the awake time (ρ = 0.4 for CON, ρ = 0.36 for somatomotor, p < 0.05) and restlessness (ρ = 0.53 for somatomotor, p < 0.01) 6 days prior. This result suggests that in attention tasks, the influence of awake time on both within- and between-network integration occurs within a 1-week timeframe.

The majority of the findings showed a positive correlation, indicating that more hours of sleep, awake time in bed, and interruptions during sleep are associated with an increase in global efficiency and participation coefficient. This pattern suggests that multiple factors involved in the sleep quality—and not only the sleep duration—are important for the efficiency of node communication within and between networks.

The analyses also showed that sleep patterns experienced beyond the previous day are related to how easily the DMN, FPN, CON, and somatomotor networks communicate with other networks. For example, sleep duration (ρ = 0.4, p < 0.05) experienced on the 15th day prior and restlessness (ρ = 0.39, p < 0.05) experienced on the seventh day prior are significantly correlated with the participation coefficient of the DMN ( Fig 6E ) and FPN ( Fig 6F ). Further, awake time in bed 6 days earlier is associated with the between-network connectivity for the CON (ρ = 0.40, p < 0.05) ( Fig 6G ) and somatomotor network (ρ = 0.36, p < 0.05) ( Fig 6H ). Likewise, prior sleep duration and restlessness correlate with the somatomotor network’s participation coefficient with lags spanning from 6 to 14 days (for full statistics, see S18 Table ).

Sleep patterns from 3 and 14 days in the past are correlated with the global efficiency in the DMN (A), and FPNs (B). Similarly, sleep patterns from 6 to 15 days in the past are correlated with the participation coefficient in the DMN (E), fronto-parietal (F), cingulo-opercular (G), and somatomotor (H) networks. Significant correlations are shown by an asterisk (*). The data was also analyzed using different thresholds, parcellations, and regressing the global signal. Squares show correlations that we found significant in at least 2 analyses. Sleep measurements are correlated with whole-brain connectivity at different time lags, which suggests that different aspects of sleep history have specific, time-sensitive relationships with distinct brain network functionalities. Unprocessed study data can be found in the Zenodo data set release [ 175 ]. Processed results derived from the study data are accessible in the GIT repository [ 176 ], under the results folder. DMN, default mode network; FPN, fronto-parietal network.

Cross-correlation analyses revealed that the global efficiency within nodes of the DMN and FPN is correlated with the sleep habits of 2 weeks in the past. In particular, sleep duration (ρ = 0.51, p < 0.05), awake time in bed (ρ = −0.48, p < 0.05), and restless sleep (ρ = 0.46, p < 0.05) experienced on the 14th day prior correlate with the global efficiency in the DMN ( Fig 6A ), and the FPN ( Fig 6B ). In addition, the awake time spent in bed 3 days prior is associated with the FPN within-network integration estimate (ρ = 0.54, p < 0.01) ( Fig 6B ). For other correlation statistics, see S17 Table .

We employed time-lagged cross-correlation analyses to identify the relationships between past sleep behaviors and brain network estimates (global efficiency and participation coefficient) for the DMN, FPN, CON, and somatomotor network. Null distributions were created based on 10,000 correlations between the brain estimates and surrogate data. We also corrected for multiple comparisons, assuming each tuple of variable-lag is independent, i.e., 4 networks for each variable-lag tuple.

Other external factors were found to be related to the daily-ISC when we analyzed the data with the additional scrubbing percentages and models ( S16 Table ). From these results, we highlight the negative relationship between the previous day’s restless sleep and ISC in the left superior parietal sulcus using the NN model (ρ = −0.26, p < 0.05) for percentage-based scrubbing at 5% when the global signal is regressed ( S16 Table ). Likewise, the positive relationship between the mean HRV and the ISC activity in the left medial superior frontal gyrus using the AK model (ρ = 0.19, p < 0.05), which is consistent with for 2 scrubbing strategies when the data is analyzed with a second parcellation.

In an effort to preserve the high standards of data quality while also keeping more fMRI volumes, we also adopted a percentage-based scrubbing approach for the movie watching task. This method involves discarding a particular volume from all sessions if it is flagged (FD > 0.2) in at least a percentage of sessions. For example, for a 30 session sample, a 10% threshold excludes all volumes where FD > 0.2 in at least 3 sessions. We applied this percentage-based scrubbing with 2 values 10% and 5% and we were able to improve the percentage of discarded volumes (5% discarded and 32% discarded, respectively, for the 2 percentage-based scrubbing strategies). The number of volumes scrubbed using these percentages are reported in the Supporting information ( S24 and S25 Figs).

Using this strict criterion, we found no significant relationship between behavioral factors and brain ISC for any of the parcellations. Nevertheless, when regressing the global signal, we identified a significant relationship between the brain ISC and the previous day’s mean HRV (ρ = −0.19, p < 0.05) in the left medial frontal gyrus (MNI coordinates x = −5.5, y = 29.3, z = 44) using the AK model. We also observed a significant relationship between the previous day’s mean negative affect (ρ = −0.19, p < 0.05) in the left middle temporal gyrus (MNI coordinates x = −53.1, y = −11.4, z = −16) using the NN model. Both results were obtained using the alternative parcellation.

To assess the relationship between brain connectivity and behavior (i.e., sleep, ANS, or mood) during movie-watching, we computed Mantel tests between the pairwise ISC for each ROI and the similarity matrix for each behavioral variable. It is important to keep in mind that in order to compute the pairwise ISC for each ROI, all sessions must have comparable time series. Therefore, any volumes scrubbed from one session due to high head movement must inevitably be scrubbed from the other sessions as well. To verify the quality of the data, we computed the total number of scrubbed volumes by session, following the exclusion criteria FD < 0.2. Upon inspection, we noticed that following such rigorous scrubbing discarded 61% of the volumes for all sessions, leaving only a few untouched segments in the first third part of the film ( S23 Fig ).

Finally, regression results from the within- and between-networks estimates showed that the previous day’s minimum HRV was the only significant predictor in the model, showing a notable impact on the FPN participation coefficient (β = 0.86, p < 0.01) ( Fig 5D and 5E ). In other words, lower minimum HRV in the previous day is associated with lower between-network integration for the FPN. This pattern persisted when the global signal was regressed from the rs-fMRI data ( S22 Fig and S15 Table ), but not when the data was analyzed with a second parcellation. Instead, when employing a second parcellation, we found that mean HRV was associated with CON participation coefficient (β = 0.83, p < 0.05). We also found that total sleep duration (β = −0.62, p < 0.05), mean negative affect (β = −0.52, p < 0.05), mean HRV (β = 0.81, p < 0.05), and maximum HRV (β = −0.52, p < 0.05) were found to drive changes in the DMN participation coefficient ( S21 Fig ).

Supplementary analyses, including a second parcellation and global signal regression, revealed that the prior’s day maximum HRV predicted link-weights in the DMN, FPN, and CON, albeit being different from the ones yielded in the main analysis ( S21 and S22 Figs and S13 and S14 Tables). Awake time in bed also emerged as a significant predictor of brain connectivity in the 3 networks when we analyzed the data with a second parcellation ( S21 Fig and S13 Table ), but not when the global signal was regressed. Despite the links found in each analysis being different, the emergent pattern remains: links associated with awake time have positive slopes (β ranges between 0.88 and 1.33, p < 0.01) and links related to HRV have mixed effects (β absolute values ranges between 0.73 and 1, p < 0.01). In other words, spending more time in bed without sleeping increased the connectivity in selected links. In addition, lower maximum HRV enhanced the connectivity in some links, while decreasing it on others. We also found links related to the percentage of eyes closure in the scanner for all parcellations and global signal regression analyses (β absolute values ranges between 0.68 and 1, p < 0.01).

(A) Linear regression models on individual links showed significant associations between the previous day’s awake time and the connectivity between the prefrontal cortex in the DMN (purple) and a subcortical area in the CON (cyan). (B) Analyses also revealed significant relationships between the microsleep time in the scanner and the resting-state connectivity in the DMN, cingulo-opercular, and FPNs. (C) Similarly, regression analysis demonstrated a direct proportional relationship between prior night’s maximum heart rate variability and the connectivity between links in the DMN and CON. Red colors indicate positive correlations and blue colors indicate negative correlations. (D) Nodes from the FPN employed to compute the network’s participation coefficient. (E) Partial regression plot showing that the FPN’s participation coefficient is strongly predicted by the previous day’s minimum heart rate variability. Results are empirically thresholded via 10,000 iterations of nonparametric permutation testing and further corrected for multiple comparisons (corrected p < 0.05). Brain plots were generated with netplotbrain [ 171 ]. Unprocessed study data can be found in the Zenodo data set release [ 175 ]. Processed results derived from the study data are accessible in the GIT repository [ 176 ], under the results folder. CON, cingulo-opercular network; DMN, default mode network; FPN, fronto-parietal network.

To understand the effect of the previous day’s sleep, mood, and ANS activity patterns on resting-state functional connectivity, we ran a series of regression models on the unthresholded link-weights within the DMN, FPN, and CON. Results from these models revealed that the previous day’s awake time in bed was negatively related to the connectivity between the left dorsolateral superior frontal gyrus and the right thalamus (β = 1, p < 0.01) ( Fig 5A ). Results also indicated that the previous day’s maximum HRV was associated with the connectivity between the left insula and the left middle temporal gyrus (β = 0.95, p < 0.01), and the connectivity between the right posterior cingulate gyrus and the right cerebellum 8 (β

[END]
---
[1] Url: https://journals.plos.org/plosbiology/article?id=10.1371/journal.pbio.3002797

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/

You are viewing proxied material from magical.fish. The copyright of proxied material belongs to its original authors. Any comments or complaints in relation to proxied material should be directed to the original authors of the content concerned. Please see the disclaimer for more details.