Substitutions of Sleep, Physical Activity, and Sedentary Behaviour and the Risk of Dementia and Stroke

Authors: Sanne J. W. Hoepel, M. Kamran Ikram, M. Arfan Ikram, Trudy Voortman, Annemarie I. Luik

Abstract

Sleep, sedentary behaviour, and physical activity (PA) are important for brain health. Spending more time in one behaviour always substitutes time in another, which may affect associations and should be considered in prevention strategies. We assessed how substitutions of sleep, sedentary behaviour, and PA are associated with incident dementia and stroke, using compositional analysis. Participants (mean 71.3 ± 9.26 years, 51.6% female) without prevalent dementia (N = 1899) or stroke (N = 1854) from the Rotterdam Study wore an accelerometer for ≥ 4 days to estimate the duration of sleep, sedentary behaviour, light PA, and moderate‐to‐vigorous PA. Participants were continuously followed up for dementia (median: 4.5 years) and stroke (median: 5.1 years). Compositional Cox regression with isotemporal substitution analysis was used to assess associations of 30‐min pair‐wise substitutions with dementia and stroke. In total, 50 (2.6%) participants were diagnosed with dementia and 75 (4.0%) with a first‐time stroke. Spending more time in moderate‐to‐vigorous PA and less in other behaviours was associated with a lower risk of dementia (Hazard Ratio [HR] for 30 min less sedentary behaviour 0.36; 95% CI: 0.24–0.55) and so was more sleep and less sedentary behaviour (HR: 0.87; 0.79–0.97) or light PA (HR: 0.43; 0.27–0.68). Those with more light PA and less sedentary behaviour had a higher risk of dementia (HR: 1.78; 1.19–2.66). Only having more sleep and less sedentary behaviour was associated with having a stroke (HR: 1.14; 1.03–1.27). More time in sleep and moderate‐to‐vigorous PA, substituting particularly sedentary behaviour, may be a modifiable risk factor for dementia. No consistent effects for stroke were found, warranting future research.

Introduction

The importance of various individual lifestyle behaviours for healthy brain aging is increasingly acknowledged. For example, getting sufficient sleep and increasing levels of physical activity (PA) might mitigate the risk of dementia and stroke (Livingston et al. 2020; Lloyd‐Jones et al. 2022). Yet, it is often overlooked that sleep, sedentary behaviour, and PA are interdependent; all behaviours occur within the constraints of the 24‐h day and are thus mutually exclusive (Dumuid et al. 2020, 2019). Therefore, it is important to consider the compositional nature of 24‐h activity increasing time in one behaviour will always lead to a decrease in one or more of the other behaviors (Dumuid et al. 2020, 2019; McGregor et al. 2020).

The isometric log ratio methodology provides a framework to integrate compositional data as an exposure. Recently, this framework was applied in several studies to investigate the cross‐sectional association of 24‐h activity and cognition, which showed that more moderate‐to‐vigorous PA, replacing any other behaviours, was associated with better cognitive performance (Feter, de Paula, et al. 2023; Dumuid et al. 2022; Hyodo et al. 2022; Wu et al. 2023; Mitchell et al. 2023). Compositional data‐analysis has also been extended to time‐to‐event data in the context of mortality and cardiovascular disease (McGregor et al. 2020; Chastin et al. 2021; Yerramalla et al. 2021; Niemelä et al. 2023; Walmsley et al. 2021), again showing a beneficial effect of spending more time in moderate‐to‐vigorous PA, but it has not yet been applied to dementia and stroke risk. By modelling the effect on an outcome that is associated with substitutions between behaviours, this framework gives insight into the relative importance of different 24‐h activity behaviours (Dumuid et al. 2020). For example, we can hypothesise that more physical activity is associated differently with the risk of dementia when replacing sleep than when replacing sedentary behaviour.

Sleep is an important part of 24‐h activity and has been linked to the risk of dementia and stroke (Leng et al. 2015; Shi et al. 2018). However, sleep is often not considered in studies using compositional data analysis (e.g., in (Yerramalla et al. 2021)) or self‐reports of sleep duration or time spent in bed as a proxy for sleep (e.g., in (Feter, de Paula, et al. 2023; Dumuid et al. 2022; Hyodo et al. 2022; Wu et al. 2023; Niemelä et al. 2023)). Estimating 24‐h activity by combining objective activity measures with self‐reported measures might bias findings. For example, time in bed overestimates sleep duration by not accounting for time spent awake in bed. Also, multiple demographic and health‐related factors, including poor cognition, are linked to discrepancies between self‐report and accelerometer‐based sleep duration (Van Den Berg et al. 2008). Integrating objective measures of sleep in studies of 24‐h activity composition and brain health will improve our understanding of the role of sleep within 24‐h activity.

In the current study, we aimed to assess the relative importance of sleep, sedentary behaviour, and PA for healthy brain aging by assessing the association of the composition of objectively measured 24‐h activity with the risk of dementia and stroke within the prospective population‐based Rotterdam Study.

Methods

Study Setting and Population

The current study was embedded within the population‐based Rotterdam Study, an ongoing prospective cohort of middle‐aged and elderly persons in Rotterdam, the Netherlands. This cohort aims to investigate determinants and consequences of aging and age‐related disease (Ikram et al. 2024). The Rotterdam Study has been approved by the Medical Ethics committee of Erasmus MC (registration number MEC 02.1015) and by the Dutch Ministry of Health, Welfare and Sport (Population Screening Act WBO, licence number 1071272‐159521‐PG) and has been entered into the Netherlands National Trial Register (NTR; www.trialregister.nl) and the WHO International Clinical Trials Registry Platform (ICTRP; www.who.int/ictrp/network/primary/en/) under shared catalogue number NTR6831. All participants provided written informed consent to participate in the study.

Between 2011 and 2016, 2778 participants were invited to wear an accelerometer for 7 consecutive days and complete a sleep diary, of whom 2338 accepted. We excluded 392 participants who had invalid accelerometer data for one or more of the following (i) did not have 4 days of measurement with at least 1200 min of measurement per day; (ii) did not have a complete sleep diary; (iii) device malfunction. In addition, 21 participants did not consent to follow‐up. Our final sample consisted of 1925 participants, who were followed up until loss to follow‐up, death, or January 1st, 2020, whichever occurred first. For dementia, 85.2% of the potential person‐years were observed and 98.8% for stroke. Participants with prevalent dementia or missing data on prevalent dementia (n = 26) were excluded to create a dementia‐free sample for incident dementia analysis (n = 1899). Similarly, those with prevalent stroke (n = 71) were excluded to create a stroke‐free sample for the incident stroke analysis (n = 1854). A flowchart of the study sample is provided in Figure 1.

24‐h Activity

Participants were instructed to wear a triaxial accelerometer (GENEActiv; Activinsights Ltd., Kimbolton, UK) on the non‐dominant wrist and fill out a sleep diary for 7 days. Details of data processing can be found elsewhere (Koolhaas et al. 2017). In short, non‐wear time was excluded, and accelerometer data were processed with PAMPRO software in Python (2.6.6.). The accelerometer sampled at 50 Hz, and activity was expressed relative to gravity (g units; 1 g = 9.81 m/s^2^). Activity was categorised into inactive time (< 48 mg), light PA (48–154 mg), and moderate‐to‐vigorous PA (> 154 mg) (White et al. 2016). Participants noted their bedtime and getting‐up time in the sleep diary, which was used to calculate time in bed and to set the window for sleep detection. Sleep duration was estimated with the validated GGIR algorithm, version 1.6‐7 (van Hees et al. 2015). Sedentary behaviour was estimated as the difference between sleep duration and inactive time. Additionally, we divided sedentary behaviour into time spent awake in bed (e.g., wake after sleep onset, sleep onset latency, etc.) and sedentary time out of bed. Time spent awake in bed was calculated as the difference between time in bed—based on the sleep diary—and sleep duration and was subtracted from total sedentary behaviour to calculate sedentary time out of bed. Time spent in each behaviour was > 0 for each individual, so we did not need to account for zeros in the compositional data. For those who had less than 1440 min of data available per day, we rescaled data to represent daily activity as a composition summing up to 1440 min (24 h) per day.

Dementia and Stroke Ascertainment

Participants who consented to follow‐up were screened at baseline and continuously monitored for stroke events and dementia through linkage of the study database with medical records from general practitioners, the regional institute for outpatient mental health care, and nursing homes. Additional medical information (e.g., clinical notes and neuro‐imaging reports) was obtained from hospital records, if available. Whenever a potential incident event occurred, records were reviewed by research physicians and subsequently verified by a consultant neurologist.

Dementia was defined according to standard criteria for dementia (using Diagnostic and Statistical Manual of Mental Disorders III—revised) (de Bruijn et al. 2015). Participants were screened for dementia at baseline and subsequent centre visits with the Mini‐Mental State Examination and the Geriatric Mental Schedule organic level. Those with a Mini‐Mental State Examination score of less than 26 or Geriatric Mental Schedule score of more than 0 underwent further investigation and informant interview, including the Cambridge Examination for Mental Disorders of the Elderly. Available information on clinical neuroimaging was used when required for the diagnosis of dementia subtype (de Bruijn et al. 2015).

Stroke was defined according to the World Health Organisation criteria, describing a syndrome of rapidly developing symptoms of focal or global cerebral dysfunction lasting 24 h or longer, or leading to death, with an apparent vascular cause (Berghout et al. 2023). Subarachnoid haemorrhages were not considered stroke events. Prevalent stroke was assessed at study entry during an interview, and these data were verified with medical records.

Covariates

Age, sex, education, smoking, paid employment, body mass index, and coronary heart disease were included in all models as possible confounders. All variables were assessed during a home interview at the time of the accelerometer measurement, except for body mass index which was measured at the research centre. Education was categorised as primary, lower, intermediate, and higher education (Unesco Institute for Statistics 2012). Smoking was divided into current, past, or never. Paid employment was categorised as yes/no. Body mass index was calculated using height and weight (kg/m^2^), measured on calibrated scales without shoes and heavy clothing. Existing coronary heart disease at the time of actigraphy was based on continuous follow‐up of medical records and was defined as the presence of a previous myocardial infarction, coronary artery bypass grafting (CABG) and/or percutaneous coronary intervention (PCI). Occurrence of missing data in covariates was low (max. 2.4%). Missing data were imputed using the missForest package (version 1.5), an iterative imputation method based on a random forest (Stekhoven and Bühlmann 2012).

Compositional Data Analysis

We used a compositional data analysis approach to analyse 24‐h activity behaviours as exposure. All analyses were performed in R (version 4.3.0). Mathematical background and details of this analysis have been described in detail elsewhere (Dumuid et al. 2020; McGregor et al. 2020; Walmsley et al. 2021). The average 24‐h activity was reported with the compositional mean (Dumuid et al. 2020, 2022). For each participant, time spent in sleep, sedentary behaviour, light PA, and moderate‐to‐vigorous PA was transformed to three isometric log ratios using the package Compositions (Dumuid et al. 2019; van den Boogaart et al. 2023). We fitted Cox proportional hazards models with 24‐h activity—reflected by the isometric log ratios—as exposure, and stroke and dementia as outcomes. Following standard methods (McGregor et al. 2020), we used a likelihood ratio test as an “omnibus test” to assess whether adding 24‐h activity to an age‐ and sex‐adjusted base model significantly improved the fit in order to assess whether 24‐h activity is associated with the outcome. Next, we first fit a model adjusted for age and sex, and a second model additionally adjusted for education, smoking, paid employment, body mass index, and coronary heart disease. Given that results were similar, only fully adjusted models are reported. The proportional hazards assumption was met for each behaviour.

To facilitate interpretation of the effect sizes, we calculated the risk of the outcome associated with a substitution in the composition, an approach called isotemporal substitution (Dumuid et al. 2019; Walmsley et al. 2021). First, we calculated the hazard ratio associated with substituting 30 min spent in one behaviour (e.g., sedentary behaviour) with another behaviour (e.g., sleep), while the other two behaviours remained constant. The reference composition was the average 24‐h activity in our study population. Next, we visualised pair‐wise substitutions of behaviours up to 1 h, using the Epicoda package (Walmsley 2020). Mathematical derivation of hazard ratios and 95% confidence intervals can be found in Walmsley et al. (2021). As a sensitivity analysis, we repeated our main analyses while including sedentary behaviour out of bed and time spent awake in bed as separate components of 24‐h activity to assess whether associations of sedentary behaviour were driven by one of these two components. We additionally repeated our main analyses while censoring cases that occur within the first year, to explore the role of reversed causation in our findings.

Results

Characteristics of the study population can be found in Table 1. The average age of the 1925 included participants was 71.3 years (standard deviation (SD): 9.26) and 51.6% of the population was female. The average participant spent 28 h sleeping, 45 h in sedentary behaviour, 27 h in light PA, and 19 h in moderate‐to‐vigorous PA. During a median of 4.5 (Q1–Q3 3.6–6.2) years of follow‐up, 50 (2.6%) participants were diagnosed with incident dementia. Moreover, 75 (4.0%) experienced a first‐time stroke during a median of 5.1 (Q1–Q3 4.2–6.8) years of follow‐up.

Composition of 24‐h activity was significantly associated with the risk of dementia (omnibus χ ^2^ (df): 29.51(3), p < 0.001). Substitution plots (Figure 2) indicate that those who spent more time in moderate‐to‐vigorous PA and less in other behaviours had a lower risk of dementia even at small amounts (i.e., 5 min) compared to the reference composition. For example, those who spend 30 min less in sedentary behaviour and 30 min more in moderate‐to‐vigorous PA have a lower risk (hazard ratio (HR): 0.36; 95% confidence 0.24–0.55) of dementia compared with the reference composition (Table 2). Spending more time in light PA and less in other behaviours was associated with a higher risk of dementia (e.g., HR 1.78; 1.19–2.66 for 30 min more light PA and 30 min less sedentary behaviour). Risk of dementia was higher in those who spent more time in sedentary behaviour and less in sleep (HR 1.15; 1.04–1.28) or moderate‐to‐vigorous PA (HR 4.44; 2.39–8.23). Those who spend more time in sedentary behaviour and less in light PA (HR 0.49; 0.30–0.80) had a lower risk of dementia. Last, those who spent more time asleep had a lower risk of dementia if they spent less time in sedentary behaviour (HR 0.87; 0.79–0.97) or light PA (HR 0.43; 0.27–0.68), but a higher risk when they spent less time in moderate‐to‐vigorous PA (HR 3.88; 2.07–7.28). Conclusions remained similar when incident cases that occurred within the first year after actigraphy were censored (Table S1).

The overall association of 24‐h activity composition with the risk of stroke did not reach statistical significance (omnibus χ ^2^ (df): 7.34(3), p = 0.06). Given that the p‐value was close to 0.05, we decided to inspect effect sizes for individual substitutions while remaining cautious in the interpretation. When specific substitutions were modelled, those who spent 30 min more asleep and 30 min less in sedentary behaviour (Table 2) had a higher risk of stroke (HR: 1.14; 1.03–1.27). Vice versa, those who spent 30 min more in sedentary behaviour and 30 min less in sleep had a lower risk of stroke (HR 0.87; 0.78–0.97). No other 30‐min substitutions were statistically significantly associated with the risk of stroke (Figure 3). Again, conclusions were similar when incident cases that occurred in the first year after actigraphy were censored (Table S1).

As a sensitivity analysis, we included time spent awake in bed and sedentary behaviour out of bed as separate components of 24‐h activity (Table S2). For dementia, spending relatively more time in either sedentary behaviour out of bed or being awake in bed, and less time sleeping, was associated with a higher risk of dementia. However, the association was statistically significant only for sedentary behaviour out of bed. For stroke, spending more time in sleep and less in sedentary behaviour out of bed, but not time awake in bed, was associated with a higher risk of stroke. For both dementia and stroke, those who spent less time in light or moderate‐to‐vigorous PA and more time awake in bed had a similar risk to those with more time in sedentary behaviour out of bed.

Discussion

In this study, we found that spending more time in moderate‐to‐vigorous PA and sleep and less time in sedentary behaviour or light PA was associated with a lower risk of dementia. There were no clear associations between 24‐h activity and the risk of stroke.

We observed that spending more time in moderate‐to‐vigorous PA is associated with a substantially lower risk of dementia. This is in line with results from several meta‐analyses (e.g (Feter, Leite, et al. 2023; Iso‐Markku et al. 2022; Zhang et al. 2023).) and previous cross‐sectional compositional studies assessing cognition as an outcome (Feter, de Paula, et al. 2023; Dumuid et al. 2022; Hyodo et al. 2022; Wu et al. 2023; Mitchell et al. 2023). Additionally, intervention studies confirm that there might indeed be a small beneficial effect of PA on cognition (Livingston et al. 2020; World Health O 2019). We confirm this association might extend to dementia as an outcome and show that this holds regardless of the replaced behaviour and for small amounts of moderate‐to‐vigorous PA. PA is thought to prevent dementia directly via mechanisms including decreased amyloid‐β production and improved blood flow and vasculature in the brain and indirectly via improving other cardiovascular risk factors (De la Rosa et al. 2020). Given our relatively short follow‐up, findings might be partly attributable to reverse causation bias, because PA tends to decline in the prodromal phase of dementia, even though our findings were not driven by cases that occurred in the first year after actigraphy (Kivimäki et al. 2019). Future studies with extended follow‐up time and repeatedly measured actigraphy are needed to disentangle these causal relations. Nevertheless, our results support the need for large‐scale intervention efforts to assess whether small increases in daily levels of moderate‐to‐vigorous PA might attenuate the risk of dementia.

Those who spent more time in sleep and less in sedentary behaviour or light PA had a lower risk of dementia, in line with previous evidence linking short sleep duration and sleep disturbances to a higher risk of dementia (Shi et al. 2018; Sabia et al. 2021). Previous compositional studies that used self‐reported sleep (opposite to objective estimates) did not find associations of sleep with cognition (Feter, de Paula, et al. 2023; Dumuid et al. 2022; Hyodo et al. 2022; Wu et al. 2023; Mitchell et al. 2023), while we do find this association for dementia. This could be explained by the use of accelerometry, stressing the importance of incorporating objective measures of sleep in studies of 24‐h activity. Indeed, in previous work we noted associations of objectively measured sleep characteristics, but not subjectively reported sleep characteristics, with the risk of dementia (Lysen et al. 2018, 2020). Suggested mechanisms linking short sleep to dementia include impaired amyloid‐β clearance, increased neuroinflammation, and disrupted neurogenesis (Shi et al. 2018; Wang and Holtzman 2020). Again, reversed causation might partly explain our findings, as objectively measured short sleep might reflect sleep disturbances that are common in the prodromal phase of dementia (Leng et al. 2019). Nevertheless, our findings underscore the importance of including sleep as a part of 24‐h activity in multidimensional interventions and activity guidelines that aim to improve brain health (Ross et al. 2020).

Our finding that those who spend more time in light PA, such as walking, and less in sedentary behaviour have a higher risk of dementia is unexpected, as it contrasts with the widely held hypothesis of a dose‐dependent protective effect of PA on brain health (e.g (Nguyen et al. 2022).), which underpins many guidelines and recommendations (Ross et al. 2020). However, it is similar to previous findings from compositional studies on 24‐h activity and cognition (Feter, de Paula, et al. 2023; Dumuid et al. 2022; Mitchell et al. 2023). These findings may partly result from residual confounding. Higher levels of light PA could, for example, reflect an inability to perform moderate‐to‐vigorous PA as a result of frailty. We could also speculate that light PA might not reach the intensity needed to positively affect brain health (Zhang et al. 2023) or that replacing sedentary time with light PA might be most beneficial in those with low levels of light PA (Bennett et al. 2017). However, our effect sizes reflect a 30 min substitution from the average composition, containing 2.5 h light PA, which is less than recommended for this age group (Ross et al. 2020) and what was previously reported in this age group (e.g., Walmsley et al. 2021). Last, a possible explanation for the observed harmful effect of light PA is that certain sedentary activities (e.g., social activities or puzzling) are cognitively challenging and can have a protective effect when replacing light PA (Su et al. 2022). This hypothesis can be explored by integrating different types of sedentary behaviour within the compositional framework.

Unexpectedly, we did not find an association between time spent in PA and the risk of stroke, contrasting previous research showing that increased PA is associated with a lower risk of stroke, based on both subjective (Kivimäki et al. 2019; Bennett et al. 2017; Lear et al. 2017) and objective (Nguyen et al. 2022; Hooker et al. 2022; Peter‐Marske et al. 2023) estimates of PA. Studies of 24‐h activity composition also reported decreased risks of cardiovascular disease when moderate‐to‐vigorous PA substituted other behaviors (Yerramalla et al. 2021; Niemelä et al. 2023; Walmsley et al. 2021), but stroke was not studied separately. We could speculate that vigorous PA might increase the risk of stroke in people with a high cardiovascular risk, which could explain an overall null effect (Bennett et al. 2017). Unfortunately, we could not explore this hypothesis, given our limited sample size. We did observe that spending more time asleep and less in sedentary behaviour—but not other behaviours—is associated with a higher risk of stroke. Sensitivity analyses revealed that this was driven by sedentary behaviour out of bed and not time spent awake in bed. This finding should however be interpreted with caution as 24‐h activity was overall not associated with risk of stroke. Previously, self‐reported short and long sleep duration have been linked to a higher risk of stroke (Leng et al. 2015; Wang et al. 2022), but studies based on objective measures of sleep are scarce and show mixed results (Zhao et al. 2021; Hoepel et al. 2024). Future studies linking 24‐h activity composition to biomarkers of cerebrovascular disease, for example, small vessel disease or intracranial atherosclerosis, are needed to provide more insight in this association.

Strengths of the current work include our population‐based sample with low attrition and objective measures of PA and sleep. Importantly, by applying compositional Cox regression with isotemporal substitutions we can accurately account for the relative nature of 24‐h activity data. Nevertheless, a limitation of this framework is that it only considers the duration of PA and sleep. PA and sleep are multidimensional constructs and other domains beyond duration (e.g., time‐of‐day, sleep quality, etc.) are known to affect health as well (Buysse 2014). 24‐h activity measured during 1 week might also not reflect habitual behaviour. Moreover, the current study lacked power to study possible non‐linear associations (e.g., U‐shaped for sleep duration, with short and long sleep duration linking to a higher risk) (Leng et al. 2015) or differentiate between clinical subtypes of dementia and stroke. Neither could we explore the potential interaction with APOE‐ε4 status (Dumuid et al. 2022) nor adjust for a wider range of confounders. Applying compositional analysis in larger samples, potentially multi‐cohort efforts (Chastin et al. 2021), will give more nuanced insights in the relation of 24‐h activity with brain health. We could also not account for potential changes over time in the composition of 24‐h activity, as activity was not repeatedly measured. Future research should employ novel advances in the field of compositional data analysis (e.g., (Le et al. 2025; Lund Rasmussen et al. 2025)) to investigate how day‐to‐day variability within the 24‐h composition or changes across a longer time period are related to the risk of dementia and stroke. Last, wrist‐worn accelerometers might misclassify behaviours, specifically sedentary behaviour and light PA (Peter‐Marske et al. 2023; Atkin et al. 2012), but also sleep duration (Plekhanova et al. 2023), which could be improved with emerging machine learning algorithms, validated in free‐living conditions (Walmsley et al. 2021).

Conclusion

In this population‐based cohort of middle‐aged and elderly adults, we observed that spending more time in moderate‐vigorous PA and sleep and less in other behaviours is associated with a lower risk of developing dementia. We did not observe a similar consistent effect for stroke. Our findings suggest that increasing time spent in moderate‐to‐vigorous PA and sleep might contribute to healthy brain ageing and support the need for intervention studies. Further application of the compositional framework is needed to disentangle the roles of sedentary behaviour and light PA and to understand how 24‐h activity affects stroke risk.

Author Contributions

Sanne J. W. Hoepel: conceptualization, methodology, writing – original draft, data curation, formal analysis. M. Kamran Ikram: conceptualization, supervision, formal analysis, methodology, writing – review and editing. M. Arfan Ikram: writing – review and editing, resources. Trudy Voortman: writing – review and editing, resources, funding acquisition. Annemarie I. Luik: funding acquisition, supervision, conceptualization, writing – review and editing, writing – original draft.

Ethics Statement

The Rotterdam Study has been approved by the Medical Ethics committee of Erasmus MC (registration number MEC 02.1015) and by the Dutch Ministry of Health, Welfare and Sport (Population Screening Act WBO, licence number 1071272‐159521‐PG) and has been entered into the Netherlands National Trial Register (NTR; www.trialregister.nl) and the WHO International Clinical Trials Registry Platform (ICTRP; www.who.int/ictrp/network/primary/en/) under shared catalogue number NTR6831. All participants provided written informed consent to participate in the study.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information