Control vs. salience: a new axis of circadian brain-body organization

Abstract

Circadian robustness is usually cast on a single weak–strong continuum, but multi-system data suggest a different picture. We followed 52 healthy young adults for ~30 days with wearable locomotor (accelerometry; ACC) and autonomic (heart rate; BPM) signals and paired these with structural and resting-state fMRI. From person-level circadian feature vectors (stability, amplitude, acrophase, and ACC–BPM alignment/lag), we uncovered a Control–Salience axis of brain–body organization. A control-anchored archetype showed ACC-dominant rhythms—higher activity stability and amplitude, later BPM acrophase, and a longer ACC → BPM phase lead—together with stronger connectivity in cognitive control networks. A complementary salience-anchored archetype exhibited BPM-dominant rhythms—earlier BPM acrophase, higher BPM relative amplitude, tighter ACC–BPM coupling—and stronger connectivity in salience and attention networks. Across individuals, cross-system alignment (ACC–BPM lag) tracked control-network coherence, whereas rhythm timing and amplitude related selectively to cortical geometry and network strength. These findings recast circadian health as axis-based and system-specific: individuals organize along a spectrum from stability-anchored, locomotor-led profiles to coupling-anchored, autonomic-dominated profiles with distinct neural correlates. The Control–Salience axis refines mechanistic models of circadian risk and points to alignment-aware, network-targeted strategies for monitoring and intervention.

Introduction

Circadian rhythms regulate physiology and behavior across nearly every domain of human biology, from sleep–wake timing and cognition to cardiovascular, metabolic, and immune function. These rhythms emerge from a multiscale timing system in which a central pacemaker coordinates peripheral clocks, aligning brain and body processes to the 24-h day¹. Robust rhythms provide a temporal scaffold for neural processing, energy balance, and autonomic control, while weak or irregular rhythms increase risk for mood disturbance, cognitive decline, and cardiometabolic disease^2,3. Two widely used markers of circadian health are stability—day-to-day consistency—and amplitude—the contrast between day and night^4,5. Across species, stronger, more stable rhythms are linked with healthier aging and longer healthspan^6,7.

Individual differences are shaped by chronotype (morningness–eveningness) and quantitative metrics such as acrophase (peak timing), relative amplitude (RA), and interdaily stability (IS)⁵. Lower RA predicts greater odds of depression, worse mood stability, and slower reaction times². Cultural and environmental forces also matter: global smartphone data revealed shorter sleep in Japan/Singapore versus longer sleep in the Netherlands⁸, while U.S. surveys showed foreign-born students kept earlier, shorter weekend schedules than domestic peers⁹.

Most studies focus on amplitude and phase, yet two aspects remain underexplored: (i) the coordination between multiple rhythmic markers, and (ii) divergence between behavioral (activity) and autonomic (heart rate) rhythms. Irregular sleep–wake timing predicts delayed circadian phase and worse grades¹⁰, while fragmented rhythms in older adults forecast faster cognitive decline^6,11. Brain network studies further link chronotype and activity timing to connectivity between the default-mode and ventral attention networks¹².

Disrupted circadian rhythms have measurable consequences for brain structure and health outcomes. Fragmented or dampened activity rhythms predict reduced cortical thickness and white-matter integrity, while stable rhythms align with stronger neural architecture^13,14. Blunted amplitude and increased fragmentation elevate dementia and cardiovascular risk^3,15,16, and behavioral–autonomic misalignment impairs sustained attention and reaction speed in shift-work paradigms¹⁷. These converging findings underscore that circadian health is not only a matter of rhythm strength but also of cross-system coordination, with consequences for mood, cognition, and long-term brain integrity.

Despite this, most studies still frame circadian robustness along a single “weak–strong” axis, overlooking that behavioral and autonomic systems may serve as distinct anchors of a 24-h profile. Recent machine-learning approaches suggest multi-dimensional circadian phenotypes with differentiated health risks¹⁸, yet no principled taxonomy of whole-person circadian organization exists. To address this gap, we leverage continuous measures of locomotor activity (accelerometry; ACC) and heart rate (BPM)¹⁹ to build person-level circadian feature vectors that capture stability, amplitude, and alignment, and test how these features map onto both functional and anatomical brain metrics

Results

Figure 1: Subjects and Circadian Metrics

Fig. 1. Group differences in circadian metrics.

A Demographics: participants were evenly split by age (≤30 vs > 30) and sex (male vs female); just over half were domestic, one-third international; 7 subjects with acrophase > 24 h were excluded. B 24 h accelerometer (ACC) curves: younger participants showed higher mesor; international students had later acrophase (–2.5 h) and lower mesor. C 24 h heart rate (BPM): females had higher values overall, but no significant group differences in acrophase/mesor. D Intra-day ACC metrics: international students showed reduced relative amplitude and a trend toward lower stability; BPM amplitude was also lower. E Heatmaps and interday correlation matrices illustrate individual differences in stability. F Group ACC curves highlight exclusion of “bad subjects”. G Data validity was higher for ACC than for BPM. H Correlation matrix shows stability positively linked to amplitude and negatively to variability; cross-modal correlations were weaker.

Figure 1A summarizes the demographic breakdown of the study population across three categories. Age was evenly distributed, with 44.2% of participants aged 30 or younger (n = 23) and 42.3% older than 30 (n = 22). The sample also showed a fairly balanced sex distribution, with 40.4% male (n = 21) and 46.2% female (n = 24). Finally, just over half of participants were domestic students (53.8%, n = 28), while nearly one-third were international students (32.7%, n = 17). In total, 7 subjects were removed from the analysis due to having an acrophase greater than 24 h.

Figure 1B shows average 24-hour accelerometer activity (ACC) curves by age, sex, and origin groups, with shaded error bars indicating SEM. For age (left), younger participants (≤30) displayed overall higher activity levels than older participants (>30), with a significant mesor difference (t = 2.23, p = 0.0319). Interestingly, the test for acrophase yielded a negative t-value (t = –1.22, p = 0.231), suggesting that younger participants tended toward an earlier peak when tested at the individual level. For sex (middle), males and females showed very similar diurnal profiles, with no significant differences in acrophase or mesor, but a trend towards females peaking slightly earlier (p = 0.26). For origin (right), international students showed both a significantly later acrophase (t = 3.11, p = 0.00399; Δφ = –2.50 h) and a lower mesor (t = –2.59, p = 0.0131) compared to domestic students, indicating earlier peak timing and reduced overall activity in the international group.

Figure 1C shows average 24-hour heart rate (BPM) curves separated by age, sex, and origin groups. For age groups (left), younger (≤30) and older (>30) participants exhibited nearly identical profiles, with no significant differences in acrophase (t = 0.43, p = 0.67; Δφ = +0.23 h) or mesor (t = 0.68, p = 0.5). For sex groups (middle), females had consistently higher BPM across the day compared to males, along with a modestly earlier acrophase (Δφ = –0.95 h). However, these differences were not statistically significant (φ: t = 0.99, p = 0.331; mesor: t = –1.63, p = 0.111). For origin groups (right), international and domestic students showed very similar daily BPM curves, with only a small, nonsignificant phase difference (Δφ = –0.74 h, t = 1.11, p = 0.275) and no difference in mesor (t = 0.12, p = 0.906). Overall, unlike the accelerometer data, no significant group-level differences were observed in heart rate rhythms, aside from elevated BPM in females vs males.

Figure 1D shows subject-level intra-day metrics, which complement the averaged 24-hour rhythm parameters (acrophase, mesor) by quantifying stability, variability, and amplitude of daily rhythms on an individual basis. Interdaily Stability (ACC): No significant differences were observed across age (t = 0.46, p = 0.644) or sex (t = 0.40, p = 0.689). International students trended toward lower stability compared to domestic students (t = –1.95, p = 0.057), suggesting somewhat weaker synchronization to the 24-hour cycle. Intradaily Variability (ACC): No significant group differences emerged (all p > 0.3), indicating comparable levels of within-day fragmentation in activity across age, sex, and origin groups. Relative Amplitude (ACC): International students showed significantly lower amplitude than domestic students (t = –3.54, p = 0.001), pointing to dampened daily activity rhythms. No differences were found by age (t = 1.60, p = 0.119) or sex (t = –0.15, p = 0.880). Relative Amplitude (BPM): A modest but significant difference was observed, with international students again showing reduced amplitude compared to domestic peers (t = –2.21, p = 0.032). No significant effects were detected for age (t = 1.01, p = 0.320) or sex (t = 0.86, p = 0.393). Nocturnal BPM Stability: No group showed significant differences (all p > 0.5), suggesting broadly similar night-time heart rate stability across age, sex, and origin.

Overall, while most intra-day metrics did not differ by age or sex, international students consistently showed weaker and less well-defined rhythms, particularly in amplitude measures.

Figure 1E illustrates three example subjects with differing levels of interday stability, highlighting how daily patterns translate into stability metrics. For subject s52 (left), the day×hour heatmap shows a clear and consistent diurnal structure with high activity during the day and low activity at night across nearly all days. This regularity is reflected in the inter-day correlation matrix, which shows strong and uniform correlations across days, yielding a high IS value (0.65). For subject s23 (middle), the day×hour map displays a moderately structured rhythm, but with variability in the timing and amplitude of daily activity peaks. The inter-day correlation matrix is patchier, with weaker correlations across days, consistent with a low-to-moderate IS value (0.18). For subject s12 (right), the day×hour map shows little discernible daily rhythm, with noisy and irregular activity patterns. The inter-day correlation matrix is highly fragmented, with minimal correlation between days, resulting in a very low IS value (0.02). Together, these examples demonstrate how the heatmaps capture within-subject daily organization while the correlation matrices quantify day-to-day reproducibility, with higher IS values linked to more stable and consistent 24-hour activity patterns.

Figure 1F: Shown are all-subject 24-hour accelerometer (ACC) curves. Most subjects displayed a clear daily rhythm with daytime peaks and nighttime troughs (blue lines). A subset of participants (red lines) showed poorly defined rhythms with fitted acrophases exceeding 24 h, indicating shift work or irregular/poor data coverage; these were labeled “bad subjects” and excluded from group analyses. The averaged curves emphasize how excluding these subjects sharpens the group-level rhythmic signal.

Figure 1G: This panel shows the percentage of valid data across the 24-hour cycle (mean ± SEM across subjects) (number of days with at least one valid recording in that hour/total number of days). ACC validity remained consistently high throughout the day (~75–85%), while BPM had lower coverage overall (~35–65%), especially during midday and afternoon hours. This highlights stronger data density for activity than for heart rate, due to the rigorous denoising protocol to eliminate motion-corrupted PPG epochs during high-activity hours.

Figure 1H: The inter-subject correlation matrix displays pairwise relationships among rhythm metrics (φ, mesor, interdaily stability, intradaily variability, relative amplitude, and nocturnal BPM stability). Strong positive correlations emerged between ACC-derived stability measures: interdaily stability (IS) was positively related to mean interday correlation (r = 0.82) and relative amplitude (r = 0.66), indicating that subjects with more stable rhythms also had stronger daily amplitude. Conversely, IS was negatively correlated with intradaily variability (r = –0.36), consistent with the interpretation that more fragmented activity patterns reflect weaker day-to-day stability. Cross-modal correlations (ACC vs BPM) were generally weaker, reflecting that activity-based and heart-rate-based rhythmic features capture related but not identical aspects of daily physiology.

Figure 2: Brain-Circadian associations

Fig. 2. Brain–circadian associations.

A Partial correlations (adjusted for sex, age, origin) between cortical structure/function and circadian timing metrics. Later BPM acrophase correlated positively with curvature in attention/default networks, while lower ACC mesor was linked to a larger cortical area in control/default/visual networks. Longer ACC–BPM lag correlated negatively with control network FC. B Stability/amplitude metrics: higher ACC stability related to stronger visual network structure/FC; greater ACC RA to stronger control within-FC; higher BPM RA to stronger salience within-FC. C RSN cortical parcellation. D Covariates: males showed larger cortical area; older age was associated with cortical thinning (notably salience, control, default); international students had greater sensorimotor curvature. E Scatterplots highlight the strongest effects, including control FC vs ACC–BPM lag and default curvature vs BPM acrophase.

Figure 2A presents partial correlations (adjusted for sex, age, and origin) between cortical structure/function and circadian timing metrics derived from accelerometry (ACC) and heart rate (BPM). Significant associations (p < 0.05, highlighted by black squares) were observed in several specific cells. For BPM φ (unwrapped), positive correlations emerged with curvature in the attention and default-mode network, suggesting that a later heart rate acrophase is linked to increased cortical curvature in these regions. For ACC mesor, negative associations with cortical area were found in the control, default, and visual networks, implying that lower mean activity was related to larger cortical surface area in these regions. Finally, for the ACC–BPM lag, robust negative correlations were seen with control network FC and within-FC, indicating that a longer delay between activity and heart rate rhythms corresponds to stronger control network coherence.

Figure 2B shows partial correlations (adjusted for sex, age, and origin) between cortical structure/function and circadian stability/amplitude measures, including interdaily stability (IS), intradaily variability (IV), relative amplitude (RA), and nocturnal BPM stability (NBS).

Significant associations (p < 0.05, black squares) emerged in a few targeted cases. For interdaily stability (ACC), negative correlations were observed with cortical area, curvature, and FC in the visual network, suggesting that more consistent day-to-day rhythms in activity align with stronger cortical structure and connectivity in visual regions. Stronger behavioral amplitude (RA ACC) was positively correlated to control network within-FC (p = 0.019), and stronger BPM RA was correlated to increased salience network within-FC (p = 0.02,) although neither was FDR-significant.

Figure 2C displays the cortical surface parcellation of the left hemisphere, with regions colored according to their resting-state network (RSN) membership. Visual, sensorimotor, attention, salience, limbic, control, and default networks are delineated, providing the anatomical framework for the brain–circadian correlation analyses.

Figure 2D shows the covariate effects of sex, age, and origin on cortical structure and connectivity after controlling for all circadian features. As expected, males exhibited greater surface area across nearly all RSNs, with several significant associations (e.g., visual: p = 0.017, sensorimotor: p = 0.013, attention: p = 0.0041, salience: p = 0.0034, limbic: p = 0.0036, default: p = 0.0018).

For the age effects, partial correlations revealed a trend toward reduced cortical thickness with older age across most networks, reaching significance in salience (p = 0.01), control (p = 0.022), and default (p = 0.012) networks. This reinforces the well-documented thinning of cortex with advancing age, particularly in higher-order association networks.

In terms of origin, only weak effects were observed; domestic students showed thicker cortex compared to international peers. The one significant effect was observed for sensorimotor curvature (p = 0.049), indicating increased curvature in the international student cohort.

Figure 2E: These scatter plots highlight the strongest non-adjusted brain-circadian relationships.

Top left (Within_FC in Control vs. ACC–BPM lag): A robust negative association (r = –0.47, p = 0.0012) shows that longer (more negative) ACC-BPM lag is linked to stronger within-network connectivity in the control network. Individuals whose ACC curve peaked earlier relative to their BPM curve had stronger within-FC in the control network.

Top right (Curvature in Default vs. BPM φ): A positive correlation (r = 0.45, p = 0.0021) indicates that later heart rate acrophase is associated with increased cortical curvature in the default network. This suggests a structural correlate of circadian timing, where delayed BPM rhythms may align with specific geometric features in the association cortex.

Middle left (Within_FC in Control vs. BPM φ): Connectivity in the control network was positively associated with later BPM acrophase (r = 0.40, p = 0.0066). Individuals with later peaks in heart rate rhythms tended to show stronger within-network coherence, suggesting a link between circadian phase delay and enhanced control network integrity.

Middle right (Visual Area vs. ACC mesor): A significant negative association (r = –0.37, p = 0.0133) was found between mean activity (ACC mesor) and surface area in the visual network. Higher overall activity levels corresponded to a smaller visual cortical area, suggesting allometric scaling, with larger individuals moving less overall.

Bottom left (Default Area vs. ACC mesor): A similar negative correlation (r = –0.38, p = 0.0098) was observed for surface area in the default network, again indicating that higher average activity levels are linked to smaller cortical surface area in association cortex. This is parallel to the visual findings point to a consistent relationship between mesor and cortical area.

Bottom right (Default Thickness vs. BPM mesor): Finally, a negative association (r = –0.32, p = 0.03) suggests that individuals with higher average heart rate (BPM mesor) had thinner cortex in the default network. This relationship implies that elevated baseline physiological arousal may be linked to reduced cortical thickness in the association cortex.

Figure 3: Cluster-based associations

Fig. 3. Circadian archetypes and brain associations.

A Residualized features revealed two clusters: Cluster 0 (“BPM-dominant/tight-coupling”) showed earlier BPM acrophase, lower ACC mesor, stronger ACC–BPM synchrony, lower ACC stability and amplitude, and higher BPM amplitude. Cluster 1 was more “ACC-dominant/loose-coupling”. B Brain comparisons showed Cluster 0 with greater cortical metrics overall, with significant effects in visual area, salience curvature, and FC within attention/salience networks. D ROI-wise analyses highlighted stronger connectivity in Cluster 0 across somatomotor, dorsal attention, salience/ventral attention, limbic, and default hubs, whereas Cluster 1 showed sparse prefrontal/ control network increases. Together, results suggest tight autonomic–behavioral coupling aligns with enhanced sensory-attentional-salience integration, while looser coupling emphasizes frontal-control circuits.

Figure 3A shows a two-cluster solution derived from residualized circadian features, highlighting distinct “circadian archetypes.” Several features clearly differentiate the clusters. Cluster 0 (blue) is characterized by lower BPM acrophase (earlier heart rate peak; p = 0.012), lower ACC mesor (reduced mean activity; p = 1.9e–05), tighter ACC-BPM coupling (p = 0.04), lower interdaily stability in ACC (p = 2.8e–05), lower ACC relative amplitude (p = 0.0002), and higher BPM relative amplitude (p = 1.83e−-05). Thus, cluster 0 can be viewed as a ‘BPM-dominant’ group with higher ACC-BPM synchrony (tighter coupling), while cluster 0 is an ‘ACC-dominant’ group with looser ACC-BPM coupling. This dichotomy emphasizes that circadian organization can stratify individuals into distinct profiles.

Figure 3B shows the results of t-tests comparing brain metrics between the two circadian clusters (Cluster 0 – the BPM-dominant, tight-coupling group – minus Cluster 1). Warmer/redder values indicate relatively greater structural or functional measures in Cluster 0.

Overall, the most consistent trend is that Cluster 0 shows higher values across multiple metrics, especially in curvature, area, and connectivity measures. Significant differences (p < 0.05, boxed) appear in: visual area (p = 0.022), salience curvature (p = 0.016), attention FC (p = 0.046), salience FC (p = 0.047), and salience within-network FC (p = 0.025). These results suggest that the “BPM-dominant” or “synchronized” circadian archetype (Cluster 0) is associated with a larger cortical area in visual regions, distinct geometric features in salience cortex, and stronger functional connectivity within attention and salience networks.

The broader trend of red shading across the map further indicates that Cluster 0 generally has greater cortical structural and functional measures compared to Cluster 1, though many effects fall short of significance. Taken together, this pattern suggests that stronger coupling between motor and autonomic signals, and more profound fluctuations in autonomic rhythms, may align with preserved or enhanced cortical integrity, particularly in networks important for attention and salience processing.

Figure 3D: The ROI-wise (200 Schaefer parcels) within-FC analyses reveal a clear functional signature of the strong/stable circadian archetype (Cluster 0 > Cluster 1). The most consistent positive effects appeared in bilateral somatomotor regions, posterior dorsal attention nodes, and salience/ventral attention areas, along with robust contributions from limbic OFC and default mode hubs (parietal and PCC). This profile suggests that individuals with tighter autonomic-behavioral alignment show enhanced intrinsic connectivity in sensorimotor systems, attentional control networks, and salience-related frontal/insula regions, as well as midline hubs of the salience network. Together, these effects highlight a signature of preserved integration across sensory, attentional, and self-referential domains in strong circadian responders.

In contrast, negative effects (Cluster 1 > Cluster 0) were sparse but clustered in prefrontal regions of the default network and control cingulate. These regions overlap with higher-order association areas involved in executive control and flexible reconfiguration, suggesting that subjects whose behavioral rhythms vary independently of autonomic rhythms may shift functional emphasis away from integrative sensorimotor-attentional circuits toward more frontal and visual territories. Thus, the archetype of tight behavioral-autonomic coupling can be summarized as enhanced integration in core sensory-attentional-salience-midline networks, whereas subjects whose behavioral fluctuations supersede or are out of tune with their autonomic systems exhibit relatively greater prefrontal and control network coupling.

Discussion

This is the first study to compare structural and functional brain metrics with long term continuous naturalistic physiological and behavioral monitoring in a healthy sample. Our primary finding suggests that rather than ‘weak’ or ‘strong’ circadian rhythms, the variance between circadian archetypes is better along a salience-control axis:

1. Control-types: higher functional connectivity in the control network. Accelerometer-dominated circadian cycles with higher stability and relative amplitude in ACC, and a longer lag between activity and autonomic (BPM) rhythm. Their behavioral and autonomic rhythms are often misaligned, with an ACC-BPM phase lag of 6+ hours in some subjects.

2. Salience-types: higher functional connectivity in salience and attention networks. Autonomic (BPM) driven circadian cycle (higher RA, earlier acrophase) and tighter ACC-BPM coupling. Their activity levels are closely aligned with autonomic rhythms, with ACC-BPM lags of <2 h.

This dissociation highlights that amplitude is system-specific: control-types achieved circadian robustness through higher activity (ACC) amplitude and consistency, whereas salience-types were defined by stronger autonomic (BPM) rhythms and behavioral-autonomic synchrony. Thus, circadian organization stratifies individuals into qualitatively different brain–body coupling modes—one anchored in behavioral consistency but looser cross-system synchrony, and another anchored in autonomic amplitude and cross-system alignment. We speculate that these clusters share both biological and environmental origins, as we also observed correlations between DMN curvature and BPM acrophase, suggesting structural differences may accompany the differences in FC^20–24.

Circadian theory posits that the SCN coordinates both behavioral and autonomic rhythms¹, but the strength of behavioral-autonomic coupling varies across individuals¹⁹. Within this framework, the emergence of stability-anchored (high IS, ACC-RA) versus salience-anchored (BPM-RA, early BPM phase, ACC–BPM coupling) archetypes is coherent: one emphasizes consistent activity entrainment, the other synchronized autonomic output^4,5. These patterns provide a framework for population-level findings that low RA predicts poorer mental health and slower reaction times², and that cultural/environmental factors systematically shape rest–activity cycles⁸. In our sample, international students exhibited later activity acrophase, lower mesor, and reduced amplitude (ACC and BPM RA; Fig. 1B–D), echoing reports of distinct schedules in foreign-born students⁹.

A key extension of prior work is our demonstration that cross-system alignment matters. Increased ACC–BPM lag (larger delay between locomotor and autonomic rhythms) correlated with increased within-network connectivity in control systems (Fig. 2A, E), suggesting that ‘control’ types are able to desynchronize their daily rhythms due to lifestyle requirements (alarm clock use, etc.). This converges with evidence that connectivity between default-mode and ventral attention networks mediates links between chronotype and daily activity¹². In line with earlier work, more stable rhythms predicted stronger cortical integrity: our salience archetype showed a larger visual area, increased salience curvature, and stronger functional connectivity across attention and salience networks (Fig. 3B, D). These patterns build on prior associations between stable rhythms, white-matter integrity¹⁴, and cortical thickness^15,16.

Mechanistically, tight behavioral–autonomic coupling may support predictable neural resource allocation, while misaligned ACC and BPM rhythms may induce state instability that recruits executive networks. However, the causal direction remains unclear: individuals with stronger control FC may be more prone to externally structuring their routines (e.g., using alarms), whereas those with stronger salience network features may instead follow their intrinsic autonomic rhythms more naturally. This interpretation aligns with experimental misalignment studies showing decrements in sustained attention and visuo-motor performance^17,25. Group contrasts further underscored that activity-based rhythms diverged more by origin than autonomic rhythms (Fig. 1B, C), emphasizing system specificity. Sex and age covariates were consistent with known physiology: females had higher BPM, and older age predicted thinner cortex in association networks^6,11,26.

Viewed against the broader literature, the two archetypes align with known risk gradients. Low autonomic amplitude and misaligned rhythms (‘control’ type in our cohort) are tied to mood and cognitive risk^2,3, while more profound autonomic cycles and tighter behavioral alignment predict healthier trajectories^6,14. Our data support a dual-axis model of circadian health: (i) behavioral stability (day-to-day reproducibility) at the cost of inter-system synchrony and (ii) salience/coupling (synchrony between behavioral and autonomic signals). Both axes appear relevant for brain structure and network function.

Methodological factors warrant caution. ACC coverage exceeded BPM by over twofold during daytime hours, raising possible noise asymmetries (Fig. 1G). Scan time and timing varied widely across participants (from 9:00 AM to 10:00 PM) (Supplementary 1), and although diurnal effects on FC are relatively modest²⁷, they may nevertheless have influenced our single-subject FC measures, as only one scan of approximately 7.8 min was performed per participant. Brain–circadian associations were exploratory, and none of the correlations in our partial-r matrices (Fig. 2) survived correction; however, the observed brain–wearable correlations were comparable in magnitude—or in some cases stronger—than the well-established effect of a 10–15-year age gap on cortical thickness (older adults showing thinner cortex) in our sample, though still weaker than the robust sex differences we observed in cortical surface area. Additional measures, such as light exposure, DLMO, and melatonin assays, would further refine archetype taxonomy⁸.

From a translational perspective, our results argue for viewing circadian health as a multivariate phenotype encompassing both reproducibility and cross-system alignment. This perspective is consistent with clinical observations linking circadian disruption to mood, cognition, and dementia risk^{2,3,11,28–31}. Low-burden wearables can capture IS, IV, RA, acrophase, and ACC–BPM lag, and integration with cognitive and network biomarkers could yield actionable risk scores and personalized interventions to regularize schedules and tighten brain–body synchrony. Moving beyond a one-dimensional “strong vs. weak” model, we identify two archetypes of brain–body organization: a locomotor stability-anchored control-type and a salience-anchored type with stronger behavioral-autonomic synchrony. These archetypes showed distinct structural and functional brain correlates (Figs. 2, 3), align with known links to cognitive health, and suggest system-specific and alignment-aware definitions of circadian health.

Methods

Participants

Fifty-two healthy adults (27 male, 25 female) between the ages of 18 and 38 years (mean = 31.1, SD = 7.9) were recruited from the Bishop’s University and Université de Sherbrooke communities. Participants included both domestic (n = 31) and international students, with the latter group primarily from Iran (n = 21), having arrived in Canada in the past 2–3 years (MSc. and Ph.D. students). All procedures were approved by the Research Ethics Board of the Centre intégré universitaire de santé et de services sociaux de l’Estrie – Centre hospitalier universitaire de Sherbrooke (CIUSSS de l’Estrie – CHUS; approval no. 2021–2626 and were conducted in accordance with the Declaration of Helsinki. All participants provided written informed consent prior to participation.

Screening procedures excluded individuals with self-reported neurological or psychiatric disorders, sleep disorders, or use of medications affecting the central nervous system. However, participants were not pre-screened for alarm clock use, shift work, or other lifestyle factors that could influence circadian rhythms, and no standardized questionnaires were administered beyond the initial consent form.

Smartwatch data acquisition

Each participant wore a Samsung Galaxy Watch Active 2 continuously over a 30-day monitoring period. This device features an 8-channel green LED PPG sensor array with a 3-axis accelerometer and gyroscope, running Tizen OS with direct sensor access, powered by a 340 mAh battery lasting ~36 h per charge. A custom Tizen app written in C was deployed on the devices, sampling PPG and accelerometry in burst mode (1 min every 10 min at 10 Hz; i.e., not continuous) to balance fidelity and battery life. Participants were instructed to wear the device continuously except during charging. Charging windows were self-selected “low-activity” periods (<30 min/day) rather than standardized times, and wear adherence was quantified as the percentage of burst epochs with finite data per clock hour (reported as cohort mean ± SEM in Fig. 1G; descriptive only). Visual inspection was used only to flag gross anomalies after automated QC; the adherence metric, not visual review, governed inclusion.

MRI data acquisition

MRI data were collected using a 3T Philips Ingenia scanner at the CIUSSS de l’Estrie – CHUS imaging facility. Due to scheduling constraints, scan times were widely distributed across subjects, ranging from 9:00 AM to 10:30 PM, with most subjects clustering between noon and 2:00 PM.

Anatomical Scan

A high-resolution T1-weighted structural image was acquired using a 3D gradient echo sequence (TR = 7.9 ms, TE = 3.5 ms, flip angle = 8°, FOV = 240 × 240 × 150 mm, matrix = 240 × 240 × 150) with 1 mm isotropic voxels. All anatomical images were visually inspected for motion artifacts.

Functional MRI

Gradient-echo EPI; voxel size = 2.53 × 2.53 × 3.00 mm; matrix = 96 × 96 × 42; TR = 1350 ms; total = 350 volumes (~7.8 min); eyes open fixation. Echo time (TE) was 30 ms, and flip angle was 90°. Philips “dummy” stabilization volumes are not reconstructed and thus do not enter the time series. We acknowledge that ~7.8 min is relatively short for resting-state network estimation and note this as a limitation regarding network stability. Susceptibility distortion correction (fieldmap/TOPUP) was not applied, which may reduce spatial fidelity in orbitofrontal and temporal regions; however, ANTs nonlinear warp can partially mitigate these effects, though this remains a limitation.

fMRI preprocessing

AFNI 25.0.07. Volumes were realigned to the first time point (3dvolreg); six motion parameters entered nuisance regression. WM/CSF time series (3dmaskave) from subject-specific masks aligned to EPI space were included. Temporal preprocessing used 3dTproject with linear detrending (polort = 2), bandpass 0.01–0.1 Hz, volume censoring at framewise displacement > 0.3 mm, and 4 mm FWHM spatial smoothing. All nuisance regressors and censor vectors were applied in a single projection step to avoid re-introducing artifacts. We used the 6-parameter motion model with aggressive censoring; although expanded models (derivatives/squares) are common, we chose this compact model for transparency and note this as a limitation. Registration/parcellation accuracy was quality-controlled by visual inspection of atlas overlays in both T1 and native EPI space for every subject; no gross misregistrations (≈≤3 mm) were observed.

Anatomical processing and Schaefer Atlas fMRI alignment

T1 images were processed with FreeSurfer v6.0 (intensity normalization, skull-strip, surfaces, cortical parcellation). Atlas alignment used ANTs: the EPI mean (vol0) was rigidly registered to T1; MNI ICBM152 1 mm was nonlinearly (SyN) registered to T1; tissue priors and Schaefer-200 (7-network) labels were warped to T1 and then to native EPI via the inverse rigid transform (nearest-neighbor for labels). This ensured consistent masks for nuisance regression and ROI-based analyses. Connectivity analyses were restricted to the cortex (Schaefer-200); subcortical regions were not included. Surface-derived structural metrics were mapped to the same ROI framework via FreeSurfer surface-to-volume correspondence to reconcile surface (structure) and volume (fMRI) representations.

Extraction of structural and functional metrics in Schaefer ROIs

From FreeSurfer outputs, we computed cortical thickness, mean curvature, and surface area per Schaefer ROI. For fMRI, mean time series per ROI were extracted in native EPI space to form 200 × 200 correlation matrices. All correlation coefficients (FC, Within-FC) were Fisher r-to-z transformed prior to statistical testing. Within-network FC for an ROI was defined as correlation with the average time series of the other ROIs in the same Yeo-7 network (self-excluded); this yields a single interpretable coherence measure and closely tracks within-network pairwise-average FC at the level of inference presented here.

FC and within-FC

Functional connectivity (FC) metrics were derived from resting-state fMRI by extracting mean time series from each of 200 Schaefer atlas ROIs and computing pairwise Pearson correlations, yielding a 200 × 200 correlation matrix per subject. For each ROI, within-network FC was defined as the correlation between that ROI’s time series and the average time series of all other ROIs in the same functional network (excluding self-connections), based on the Yeo 7-network partition. This produced a per-ROI within-FC value reflecting how strongly each region cohered with its broader network.

PPG signal preprocessing and derivation of autonomic metrics

PPG was sampled at 10 Hz (1-min/10-min bursts). Signals were band-pass filtered (0.35–4.0 Hz; 4th-order Butterworth) and resampled to 30 Hz. Ten-second epochs were retained if the normalized autocorrelation showed a positive peak at 0.85–1.35 s lag with a negative trough at 0.35–0.7 s. Peaks were detected with a minimum peak distance = 0.6 s; IBIs and RMSSD were computed within runs and aggregated to hourly values. The conservative autocorrelation gate prioritizes low-motion, high-fidelity IBI and may under-sample very high heart rates; this potential selection bias is acknowledged as a limitation. QC combined automated gates with targeted visual review of flagged/representative segments rather than manual inspection of all epochs.

Accelerometry preprocessing and derivation of 24-hour activity profiles

Tri-axial accelerometry (10 Hz bursts) was summarized as the standard deviation of the X-axis acceleration (AccX) per burst and aggregated by recording hour across days to form per-subject 24-h curves. We used AccX as a simple gross-activity proxy; in pilot checks, day–hour patterns closely matched vector-magnitude summaries. Day-to-day variability is captured by interdaily stability (IS), intradaily variability (IV), and relative amplitude (RA) computed from the full day×hour matrices.

Circadian acrophase and autonomic-locomotor lag

For each subject, activity (AccX) and IBI were averaged within clock hour across days to form 24-h curves. We fit a 24-h cosinor model $y\left(t\right)=\mathit{MESOR}+A\cos \left(\frac{2\pi t}{24}-\varphi \right)t$ to each series (IBI converted to BPM so higher values reflect higher cardiac activation), converting $\varphi$ to clock hours. Fits required ≥6 finite hourly bins; the single-component 24-h cosinor provides a first-order summary and may not capture multimodal patterns, which we acknowledge as a limitation. Subjects whose unwrapped activity acrophase exceeded 24 h (cutoff 06:00 rule) were pre-specified as “atypical” and excluded from all inferential analyses (n = 7), remaining only in descriptive pies/overlays; all figures report n after this exclusion.

Inter-day stability and relative amplitude metrics

From the raw smartwatch data, we constructed matrices of hours by days for both activity (accelerometry) and cardiac measures (beats per minute derived from inter-beat intervals). From the hour-by-day matrices of valid smartwatch data, we derived the following inter-day metrics, using NaN-aware operations to handle missing values:

• Interdaily Stability (ACC): Quantifies how similar a subject’s daily activity profile is from one day to the next. Higher values indicate more regular repetition of the 24-hour pattern.

$$IS=\frac{D\cdot {\sum }_h{\left(\overline{\overline{X}}\left(h\right)-\overline{X}\right)}^2}{{\sum }_{h,d}{\left(X\left(h,d\right)-\overline{X}\right)}^2}$$

where $\stackrel{\circledR }{X}$ is the mean across days at hour $h$, $\stackrel{\circledR }{X}$ is the overall mean, and $D$ is the number of days

• Intradaily Variability (ACC): Captures the degree of fragmentation in daily activity. It increases when there are frequent, rapid transitions between active and inactive states.

$$\mathit{IV}=\frac{\frac{1}{N-1}{\sum }_{t=1}^{N-1}{\left(x_{t+1}-x_t\right)}^2}{\mathit{Var}\left(x\right)}$$

where $x_t$ is the vectorized hourly series across all days in column-major order (hours nested within days), and $N$ is the number of valid time points.

• Relative Amplitude (ACC): Measures the strength of diurnal amplitude in accelerometry by contrasting the most active period of the day with the least active period.

$$R{A}_{\mathit{ACC}}=\frac{M10-L5}{M10+L5}$$

where $M10$ is the maximum mean value across all contiguous 10-hour windows of the daily profile, and $L5$ is the minimum mean value across all contiguous 5-hour windows.

• Relative Amplitude (BPM): Same calculation applied to beats-per-minute time series, reflecting how strongly cardiac activity distinguishes between active and rest periods.

• Nocturnal BPM Stability (BPM): Assesses the night-to-night consistency of cardiac state by examining variability in average heart rate during a fixed nocturnal window (02:00–05:00). Lower values indicate more stable nightly rhythms.

$$NBS=StdDe{v}_d\left(\frac{1}{H}\sum _{h\in \left\{2,3,4\right\}}X\left(h,d\right)\right)$$

where the nightly mean is taken across hours 02:00–05:00 (3 h), then the standard deviation is computed across days. Lower NBS indicates more stable nightly rhythms.

Subjects flagged as “bad” by the predefined accelerometry acrophase criterion were excluded from these calculations; otherwise, missing data were tolerated through NaN-aware averaging and variance estimates.

Statistical methods

Figure 1 (wearable circadian features and group contrasts)

For age (≤30 vs >30), sex (male vs female), and origin (international vs domestic), we compared 24-h activity (ACC, z-scored) and heart rate (BPM) curves. Subjects whose ACC acrophase (φ) from a 24-h cosinor fit exceeded 24 h after unwrapping were flagged “bad” and excluded from all inferential analyses; they appear only in the demographic pie charts and as red traces in the “all-subjects” panel. Group curves show mean ± SEM across hours, with cosinor fits overlaid for visualization only. For inference, we computed per-subject φ (from individual 24-h cosinor) and per-subject mesor (the subject’s mean level across 24 h), then used two-sided Welch’s t-tests (unequal variances) to compare groups; panel titles report t and p for φ and mesor with the comparison direction explicit. Hour-wise differences between groups were assessed at each hour by separate Welch tests; hours with p < 0.05 are shaded red (uncorrected, descriptive). The “% valid” curves plot mean ± SEM and are not hypothesis-tested. Representative day×hour maps and day×day correlations (IS, IV, RA, NBS summaries) are descriptive only. Across-metric associations (10 × 10 subject-level matrix: φ/Mesor for ACC/BPM; IS, IV, RA_ACC, RA_BPM, NBS, mean inter-day ACC correlation) use Pearson’s r with two-sided p-values computed by listwise deletion; multiple testing is controlled within the lower triangle by Benjamini–Hochberg false discovery rate (BH-FDR, α = 0.05), with gold boxes indicating q < 0.05 and black boxes indicating nominal p < 0.05 but q ≥ 0.05. Violin plots (same five metrics) show group distributions; each pairwise contrast is annotated with Welch’s t and p (two-sided, uncorrected). Missing values are handled by finite-value filtering, and day-level correlations require ≥12 valid hours/day.

Figure 2 (brain–circadian associations, partial correlations)

We related 10 circadian features (five “timing”: ACC φ, BPM φ, ACC mesor, BPM mesor, ACC–BPM lag; five “stability/amplitude”: IS, IV, RA_ACC, RA_BPM, NBS) to network-averaged brain measures (Thickness, Curvature, Area, FC, Within-FC) aggregated from Schaefer-200 ROIs to Yeo-7 resting-state networks. Subjects were aligned across datasets; those with ACC φ > 24 h (unwrapped) were excluded, and analyses used only finite observations (listwise deletion). FC/Within-FC values were Fisher-z transformed prior to modeling. For each (brain metric × network) cell and each circadian predictor, we fit an OLS model with the circadian feature as the target coefficient while adjusting for sex, age, and origin; partial r was obtained from the t-statistic, and two-sided p-values are from the OLS coefficient tests. Color encodes partial r; cells show compact raw p-values; multiplicity is controlled separately within each 5 × 7 matrix by BH-FDR (α = 0.05), with gold boxes for q < 0.05 and black boxes for nominal p < 0.05. A companion “covariate-effects” row estimates partial r for sex, age, and origin while controlling for all 10 circadian features simultaneously (BH-FDR within each covariate matrix). The 3 × 2 scatter figure displays the six strongest bivariate associations (smallest unadjusted p); these panels report simple Pearson r and p (illustrative, not covariate-adjusted or FDR-corrected).

Figure 3 (covariate-independent circadian clusters and brain differences)

To derive clusters independent of demographics, each of the 10 circadian features was residualized against sex, age, and origin via OLS; residuals were z-scored and partitioned by k-means (k = 2; n_init = 20). For interpretability, cluster labels were oriented so that Cluster 0 had the earlier mean ACC φ. Histograms show the original (non-residualized) metrics by cluster; differences are tested with two-sided Welch’s t-tests (unequal variances), with t and p printed per panel (uncorrected). Brain analyses reused the RSN aggregation above; FC/Within-FC were Fisher-z transformed. For each (brain metric × network) cell, we tested Cluster 0 vs Cluster 1 with Welch’s t; the t-map is shown with raw p-values overlaid. Multiple testing across the full matrix is controlled by BH-FDR at α = 0.05 (gold boxes for q < 0.05; black boxes for nominal p < 0.05). Additionally, per-ROI tests for WITHIN_FC (200 ROIs) used Welch’s t with BH-FDR across ROIs (unthresholded). All tests are two-tailed with α = 0.05; missing data are handled by finite-value filtering, and day-level reliability metrics require ≥12 valid hours/day.

Author contributions

S.G. recruited participants and led data acquisition, including MRI and smartwatch recordings. O.D. developed the smartwatch firmware in C for raw data capture and extraction. C.L. performed statistical analyses and signal processing. R.B. conceived and supervised the study, provided funding, and oversaw all aspects of study design and interpretation. All authors contributed to manuscript preparation and approved the final version.

Data availability

All raw and preprocessed data generated and analyzed during this study are available from the corresponding author (rbutler@ubishops.ca) upon reasonable request.

Code availability

All Python and Bash scripts used for data processing, analysis, and visualization are available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.