Associations between accelerometer‐derived physical activity complexity, cognitive function, and plasma biomarkers of Alzheimer's disease

Abstract

INTRODUCTION

While higher physical activity (PA) levels are linked to lower Alzheimer's disease (AD) risk, traditional PA measures overlook temporal activity patterns.

METHODS

We developed the activity complexity measure in 164 older adults (mean age = 65.3 years) who completed wrist‐worn accelerometry and neuropsychological tests. PA complexity was derived using multiscale entropy (MSE) and averaged across days. Day‐to‐day variability was also computed. Plasma biomarkers (Aβ42, Aβ40, p‐tau181, p‐tau217, p‐tau231) were measured.

RESULTS

Participants with subjective cognitive decline and impaired scores without complaints showed lower PA complexity than cognitively unimpaired participants, but not those with mild cognitive impairment or AD. Low PA complexity was associated with poorer cognitive function (p < 0.03). Greater day‐to‐day variability in PA complexity was associated with higher plasma p‐tau levels.

DISCUSSION

Lower PA complexity and greater instability across days were associated with poorer cognition and AD pathology. Longitudinal studies should examine whether low PA complexity serves as digital biomarkers of preclinical AD.

Highlights

• Prior studies have mainly focused on quantity or intensity of physical activity.

• Poorer cognitive function was associated with lower complexity of daily activity.

• Lower complexity of physical activity may be an early indicator of dementia.

1. BACKGROUND

Dementia is a common neurodegenerative disorder that leads to profound loss of functional independence in older adults. ¹ Progression to Alzheimer's disease (AD) is hypothesized to begin with a preclinical phase of normal cognition that can precede dementia diagnosis by up to 20 years. ² , ³ Recent diagnostic framework highlighted AD as a continuum spans from clinically asymptomatic to severely impaired stages. ² Thus, identifying early changes prior to clinically recognizable symptoms of dementia is crucial for purposes of targeting interventions to slow pathological progression at the early stage.

Substantial work has sought to identify risk factors of cognitive decline and incident AD and related dementias (ADRD). Higher daily physical activity (PA) levels have been associated with slower cognitive decline, a lower risk of ADRD, ⁴ , ⁵ , ⁶ and less amyloid and tau burden. ⁷ , ⁸ However, most studies relied on self‐report PA which introduces recall bias and misclassification especially in older adults. ⁹ , ¹⁰ , ¹¹ Studies using accelerometer assessments have mainly focused on summary measures (e.g., steps/day, moderate to vigorous PA), ⁶ , ¹⁰ , ¹¹ which may underestimate associations due to measurement error and misclassification by using cutpoints. These conventional measures do not fully capture the temporal characteristics inherent in continuous accelerometer data. Novel accelerometer‐derived measures that capture variability, rhythmicity, or complexity of movement may better reflect day‐to‐day functioning and be more sensitive to early neurodegeneration than conventional PA measures.

Recently developed accelerometer metrics, including rest‐activity rhythms (e.g., intradaily variability, interdaily stability, and relative amplitude) quantify how strongly activity patterns follow a circadian pattern and how fragmented the activity pattern is within a day. Lower amplitude, higher intradaily variability, and lower interdaily stability have been linked to cognitive impairment and greater dementia risk. ¹² , ¹³ These measures, however, do not characterize the temporal irregularity of daytime movement and how activity levels fluctuate across multiple time scales (e.g., 1–5 minutes) independent of overall volume. Our group developed a novel metric ‐ “PA complexity” using multiscale entropy (MSE) method to quantify daily activity patterns in the Baltimore Longitudinal Study of Aging. ¹⁴ MSE captures the complexity and irregularities of time series across multiple temporal scales. ¹⁵ Although widely applied to neurophysiological signals (e.g., postural sway, ¹⁶ , ¹⁷ blood pressure ¹⁸ ), using MSE to characterize continuous accelerometer data is relatively novel. Higher PA complexity reflects more diverse and non‐repetitive patterns of movement (e.g., alternating activities and breaks across the day), whereas lower complexity reflects more uniform and monotonous activity patterns (e.g., prolonged sedentary bouts). In prior work, low PA complexity was associated with higher dementia risk and showed a bidirectional association with executive function. ¹⁴ PA complexity also outperformed existing metrics (e.g., total activity counts [TAC], time spent in active states, activity fragmentation) in predicting mild cognitive impairment (MCI) or dementia over 4 years of follow‐up. ¹⁴ These findings suggest that PA complexity may capture subtle loss of complexity in daily behaviors that is not detectable with conventional volume or fragmentation measures. However, the earlier cohort included relatively few MCI and dementia cases, limiting power to evaluate different stages of cognitive status, and did not examine AD‐related biomarkers.

RESEARCH IN CONTEXT

1. Systematic review: The authors reviewed the literature using traditional sources (e.g., PubMed) and meeting abstracts and presentations. Previous studies suggest that older adults with Alzheimer's disease (AD) had lower physical activity (PA) levels. These relevant citations are appropriately cited. However, traditional PA measures do not capture the temporal patterns of daily activity. We developed a novel PA measure that quantifies complexity of daily activities and examined whether this measure was associated with cognitive status and AD pathology.

2. Interpretation: Our study found that lower daytime PA complexity and greater day‐to‐day instability of complexity are linked to subtle cognitive decline and elevated plasma p‐tau in community‐dwelling older adults, which may act as early indicators of AD and related dementias (ADRD).

3. Future directions: Validation of this novel metric in other cohorts of older adults are needed. Longitudinal studies with repeated biomarker, cognitive, and neuroimaging assessments are needed to establish temporal and causal associations and examine whether declined PA complexity may serve as a preclinical indicator of ADRD.

Blood‐based biomarkers, such as amyloid‐β (Aβ42/Aβ40), phosphorylated tau, have been developed and demonstrated to detect AD at early stages. ¹⁹ , ²⁰ , ²¹ These biomarkers are less invasive and costly, which provide a time‐efficient approach for large‐scale screening than cerebrospinal fluid or neuroimaging biomarkers. ²² Higher self‐reported PA has been associated with lower plasma Aβ and p‐tau217, ²³ , ²⁴ but it is unknown whether accelerometer‐derived PA complexity relates to these emerging plasma AD biomarkers. Understanding these relationships could help determine whether PA complexity may serve as an accessible digital marker of early AD risk.

Therefore, we aimed to characterize PA complexity using the MSE method in older adults enrolled in the Human Connectome Project (HCP) ‐ Connectomics in Brain Aging and Dementia (CoBrA) study and examine cross‐sectional associations between PA complexity, cognitive diagnosis, cognitive performance, and AD‐related plasma biomarkers. We hypothesized that lower PA complexity is associated with greater odds of MCI and AD, poorer cognitive function, and higher levels of AD pathology.

2. METHODS

The HCP ‐ CoBrA is a longitudinal, community‐based study of brain structural and functional connectivity among cognitively normal and cognitively impaired individuals aged 50–89 years. The detailed study design and participants have been described elsewhere. ²⁵ Briefly, the study recruited cognitively normal and cognitively impaired individuals aged 50–89 years in Pittsburgh, Pennsylvania. The study was approved by the University of Pittsburgh Institutional Review Board and written informed consent was obtained from all participants. The current study included 190 adults who have completed both neuropsychological tests and 7‐day wrist‐worn accelerometer assessment from 2016 to 2021.

2.1. PA assessment

The CoBrA study used the ActiGraph GT9X Link (ActiGraph, Pensalcola, FL), a triaxial wrist‐worn accelerometer with a sampling frequency of 30Hz that provides access to raw acceleration data. Participants were instructed to wear the accelerometer on the nondominant wrist for 7 days in the free‐living environment. Accelerometer data was processed using the ActiLife software (version 6.13.4) to derive activity counts in 1‐minute epochs. Participants with at least 3 valid days (≤10% of missing data in the day) of accelerometer data were included in the analysis.

The concept ‐ “complexity” measured by MSE ¹⁷ , ²⁶ is based on the traditional sample entropy (SampEn), which is defined by the negative natural logarithm of the conditional probability that a time series, having repeated itself within a tolerance r for m points, will also repeat itself for m+1 points without self‐matches. ¹⁵ , ²⁷ Minute‐level vector magnitude counts between 05:00 and 23:00 were used to compute MSE which quantifies how unpredictable activity fluctuations are across multiple short time scales. For each valid day, we first z‐standardized the time series and then calculated sample entropy (embedding dimension m = 2, tolerance r = 0.15×SD, time delay τ = 1) on coarse‐grained versions of the series using scale factors 1–5 minutes. This yields an MSE curve (sample entropy vs. scale). We summarized it by averaging entropy across scales 1–5, with higher values indicating more irregular, less repetitive patterns of daytime activity (“PA complexity”). We then averaged daily PA complexity across valid days and computed its day‐to‐day standard deviation (day‐to‐day variability of PA complexity) as our primary PA complexity metrics. Full algorithmic details are provided in Supplementary Methods.

We also calculated total daily activity counts (TAC), active‐to‐sedentary transition probability (ASTP) as activity fragmentation, and time spent in sedentary states. The sum of activity counts per day was averaged across valid days to derive TAC. Each minute was labeled as active if the activity counts for that minute were ≥1,853 or sedentary if <1,853. ²⁸ Active bouts were calculated as the sum of consecutive active minutes. We calculated ASTP as the reciprocal of the average activity bout duration for each participant. ²⁹ Sedentary minutes were summed to obtain the total time spent in sedentary states.

2.2. Cognitive assessment and cognitive classifications

The neuropsychological (NP) assessments included the Word List Learning from the Consortium to Establish a Registry in AD (CERAD) battery, ³⁰ Logical Memory Story A from the Wechsler Memory Scale ‐ Revised, ³¹ modified Rey Osterrieth [R‐O] figure recalls, ³² semantic and letter fluency, ³³ Boston Naming Test (BNT), ³⁴ Trail Making Test (TMT) A and B, ³⁵ Digit Symbol test from the Wechsler Adult Intelligence Scale ‐ Revised (WAIS‐R), ³¹ and the Clock Drawing test. ³⁶ Cognitive diagnostic adjudication was determined via multidisciplinary consensus conference, ³⁷ in which NP testing, medical and social history, daily functioning, and reported cognitive symptoms. Structural MRI was reviewed for participants with AD. The ADRC classification scheme was used for AD and MCI (both amnestic and nonamnestic), and subjective cognitive decline. ³⁸ Cognitively normal participants were classified into two subgroups: those who reported no limitations in their cognition (cognitively unimpaired [CU]) and those with normal‐range test scores but reporting significant concerns [subjective cognitive complaints (SCC)]. Participants with test scores in the mildly impaired range were also classified into two subgroups: those who reported no concerns or decline of cognition [impaired scores without complaints (IWOC)], and those who reported decline of cognitive abilities (i.e., MCI). ²⁵

2.3. Plasma biomarker

Blood collections were centrifuged within 2 hours of collection at 2,000 xg for 10 min at 4°C and the transparent supernatant (plasma) transferred to new tubes for storage at ‐80°C until needed. Plasma biomarkers were quantified on the single‐molecule array (Simoa) HD‐X platform (Quanterix, Billerica, MA, USA). ³⁹ Plasma Aβ42 and Aβ40 assays were measured using the Neurology 4‐Plex E kit. P‐tau217 was measured using the ALZpath Simoa^® p‐Tau 217 V2 Assay. P‐tau181 was measured using p‐tau181 antibody AT270 for capture and anti‐Tau6‐18 antibody (Tau12) for detection. ⁴⁰ P‐tau231 was analyzed using the same procedure but with AT180 for capture and Tau12 for detection. ⁴⁰ Quality control (QC) samples were analyzed at the start and the end of each run in two to three different concentrations per assay to assess the reproducibility of each assay. The within‐ and between‐plate coefficients of variation (CV) were <5% and <6% for p‐tau181 and the p‐tau217 assay, generally <10% and <16% for the Neurology 4‐plex analytes, respectively. ⁴¹

2.4. Covariates assessment

Covariates included demographics (e.g., age, sex, race, and years of education), self‐reported chronic conditions (e.g., cardiovascular diseases, hypertension, diabetes, stroke, high cholesterol, lung disease, kidney disease, cancer), physical functioning (e.g., 2‐min walk distance, 4‐m walk gait speed, grip strength), total daily activity counts (TAC) from accelerometer measurements, and apolipoprotein E (APOE) e4 allele status.

2.5. Statistical analysis

Data distributions were checked for normality. Sociodemographic characteristics, health conditions, physical performance, and PA complexity were summarized using mean and standard deviation (SD) or frequency and percentage. Kruskal–Wallis tests for continuous variables and Fisher's exact tests for categorical variables were used to examine differences of sociodemographic characteristics, comorbidities, physical performance, and PA complexity by cognitive diagnosis. Average PA complexity and variability of PA complexity were treated as both a continuous measure and tertiles in regression models. Logistic regression models were used to estimate odds ratios (ORs) and 95% confidence intervals (CIs) for cognitive status by PA complexity. Receiver operating characteristic (ROC) curve analyses were conducted to evaluate the ability of PA complexity to distinguish each cognitive diagnostic group (SCC, IWOC, MCI, AD) from CU, and we estimated the area under the curve (AUC) at 95% CIs.

We used confirmatory factor analysis (CFA) to derive a composite score for each cognitive domain. We tested a correlated‐factors model with three latent factors (memory, executive function, language) and a model with one latent factor ‐ global cognitive function to ensure proper fit. Continuous observed cognitive variables were used to construct three latent factors: (1) memory: CERAD Word List Learning number of words correctly recalled in immediate and delayed recall, Logical Memory Story A score in immediate and delayed recall, and Memory Rey figure immediate recall; (2) executive function: TMT‐B time to completion (seconds), Clock Drawing test at a 15‐point scale, and Digital Symbol test number of correct matches; (3) language: BNT number of correct responses, count of correct words generated for semantic and letter fluency. Items yielding high measurement errors were removed from the model to improve parsimony. Potential correlations between the latent factors and between tests were considered. The final CFA model was refined through an iterative process, where modifications (e.g., adjusting factor structure, residual correlations, or bifactor model) was guided by both theoretical considerations and statistical diagnostics. The goodness‐of‐fit statistics used to assess model fit included the χ² statistic, root mean square error of approximation (RMSEA), the standardized root mean square residual (SRMR), Comparative Fit Index (CFI), Tucker‐Lewis index (TLI), and the Akaike and Bayesian Information Criterion (AIC and BIC). RMSEA and SRMR ≤ 0.08, CFI and TLI ≥ 0.95 signified good model fit. ⁴² Linear regression models were used to examine the association between PA complexity and factor scores for global cognitive function and each cognitive domain separately. From scatter plots and loess curves, we observed a knot at 0.6 where PA complexity was only associated with cognitive function when PA complexity is ≤0.6. Thus, we additionally treated PA complexity as a dichotomous variable with two categories: low PA complexity (≤0.6) and high PA complexity (>0.6).

AD‐related plasma biomarkers were treated as both continuous variables and tertiles since there were no standard cutoff points for these biomarkers. Linear regression models were used to examine the associations between PA complexity and continuous plasma biomarkers (p‐tau181, p‐tau231, p‐tau217, and Aβ42/Aβ40 ratio). p‐Tau181, p‐tau231, and p‐tau217 were log transformed due to skewed distribution. Multinomial logistic regression models were used to examine the associations between PA complexity and tertiles of blood biomarkers. Multivariable models were adjusted for age, sex, race, years of education, comorbidities, and daily total activity counts. We additionally adjusted for APOE e4 status and gait speed as sensitivity analyses.

We tested the potential modifying effects of sex and APOE e4 status by adding interaction terms in the models and all models were further stratified by sex and APOE e4 status to explore whether the associations between PA complexity, cognitive diagnosis, cognitive function, and AD‐related biomarkers differ by sex and APOE e4 status. We additionally tested the potential moderating effects of MCI or AD status on the associations between PA complexity, cognitive function, and AD‐related biomarkers. The significance level α was set as 0.05. All statistical analyses were conducted using SAS 9.4 (SAS Institute, Cary, NC) except for the CFA which was performed using Mplus 8.4 (Muthén & Muthén, Los Angeles, CA).

3. RESULTS

A total of 164 participants with valid accelerometer assessment were included in the analysis. The average age was 65.3 (SD [standard deviation] = 8.9) years, and 108 (65.9%) participants were women. About half (56.0%) of participants were White. The average PA complexity was 0.59 (SD = 0.18) for all participants. Age, race, years of education, and walk endurance were significantly associated with cognitive diagnosis (p < 0.02 for all; Table 1).

TABLE 1.

Sample characteristics.

Parameter	All participants (n = 164)	Cognitively unimpaired (n = 84)	Impaired without complaints (IWOC) (n = 23)	Subjective cognitive decline (SCC) (n = 13)	Amnestic MCI (n = 33)	Non‐amnestic MCI (n = 4)	Alzheimer's disease (AD) (n = 7)	p‐value*
Age (years)	65.3 (8.9)	64.6 (8.4)	61.4 (6.4)	68.6 (7.0)	66.5 (10.8)	67.5 (11.3)	72.6 (8.3)	0.009
Female	108 (65.9)	61 (72.6)	15 (65.2)	7 (53.9)	20 (60.6)	3 (75.0)	2 (28.6)	0.183
White	93 (56.0)	52 (61.9)	4 (17.4)	11 (84.6)	22 (66.7)	1 (25.0)	3 (42.9)	<0.001
Education (years)	15.2 (3.4)	15.9 (3.1)	12.9 (2.8)	16.8 (4.9)	14.6 (2.8)	15.3 (4.1)	14.3 (2.1)	<0.001
No. of comorbidities	1.4 (1.3)	1.6 (1.3)	0.8 (1.2)	1.5 (0.9)	1.6 (1.4)	0.8 (1.5)	1.4 (1.4)	0.104
Grip strength (n = 155)	59.8 (20.0)	59.1 (19.9)	66.0 (17.7)	61.2 (18.9)	59.2 (23.2)	50.7 (13.1)	59.4 (15.5)	0.149
Gait speed (m/s) (n = 145)	1.2 (0.2)	1.2 (0.2)	1.2 (0.2)	1.2 (0.2)	1.2 (0.3)	1.3 (0.4)	1.1 (0.4)	0.455
Walk endurance – 2 min distance (n = 139)	522.8 (108.6)	528.8 (106.9)	521.9 (69.7)	566.8 (80.0)	513.1 (142.6)	504.6 (43.6)	448.6 (86.4)	0.034
PA complexity (%)	0.59 (0.18)	0.62 (0.18)	0.52 (0.15)	0.51 (0.11)	0.61 (0.20)	0.61 (0.18)	0.52 (0.17)	0.081
Variability of PA complexity across days	0.17 (0.12)	0.19 (0.13)	0.16 (0.10)	0.15 (0.11)	0.16 (0.11)	0.17 (0.03)	0.13 (0.07)	0.544
TAC (x1000)	1997.5 (604.2)	2067.8 (579.2)	2045.5 (568.6)	1938.4 (641.5)	1933.5 (704.7)	1713.6 (384.2)	1700.1 (567.7)	0.544
APOE e4 status (n = 137)	43 (31.4)	15 (22.4)	8 (38.1)	4 (33.3)	10 (37.0)	3 (75.0)	3 (50.0)	0.137

Note: Values in the table indicate mean (standard deviation) for continous variables and n (%) for categorical variables. *p‐values from Kruskal–Wallis tests for continuous variables and Fisher's exact test for categorical variables. Bolded values indicate statistically significant results (p < 0.05).

Abbreviations: APOE, apolipoprotein E; PA, physical activity. TAC, total daily activity counts.

As illustrated in Supplementary Figure 1, MSE curves by diagnostic group were broadly parallel across scales 1 to 5, indicating that groups mainly differed in overall level of entropy rather than in qualitatively different scale‐specific patterns. The Pearson correlation r between average PA complexity and total activity counts, ASTP, and sedentary time (minutes) were 0.33, ‐0.05, and ‐0.35, respectively (Supplementary Figure 2). The associations between day‐to‐day variability of PA complexity and other PA measures were similar to average PA complexity (Supplementary Figure 2). We plotted minute‐level activity counts to visualize different PA complexity from 3 participants with similar daily TAC (Supplementary Figure 3). The participant with higher complexity (MSE = 0.8) tended to have more frequent and irregular fluctuations throughout the day. These visual patterns are consistent with the interpretation that MSE captures temporal irregularity in activity rather than the overall intensities.

Overall, there was no significant difference in average PA complexity and day‐to‐day variability of PA complexity by cognitive diagnosis groups (p = 0.08 and p = 0.66, respectively; Supplementary Table 1). In post‐hoc analyses, IWOC and SCC groups had significantly lower PA complexity compared to cognitively unimpaired participants (t = ‐2.54, p = 0.012; t = ‐2.21, p = 0.028, respectively). Other differences were not statistically significant. In ROC analyses, average PA complexity showed modest discrimination between SCC and CU (AUC = 0.68, 95% confidence interval [CI]: 0.54–0.82) and between IWOC and CU (AUC = 0.66, 95% CI: 0.53–0.78). Discrimination for AD vs CU was similar (AUC = 0.69, 95% CI: 0.46–0.91), although the confidence interval was wide because of the small AD group. In contrast, PA complexity had little ability to distinguish MCI from CU (AUC = 0.52, 95% CI: 0.40–0.64; Supplementary Figure 4).

Logistic regression models showed that participants in the IWOC and SCC groups had significantly lower PA complexity compared to cognitively unimpaired groups. Every 0.1 unit higher in PA complexity was associated with 36% lower odds of being classified in the IWOC group (OR = 0.64, 95% CI = 0.43–0.94, p = 0.024) and 39% lower odds of being in the SCC group (OR = 0.61, 95% CI = 0.37–1.00, p = 0.050) in the multivariable models adjusted for sociodemographic characteristics, comorbidities, and total activity counts per day (Table 2, Model 3). We found similar results after additionally adjusting for APOE genotype but the associations were diminished after further adjusting for usual gait speed (Supplementary Table 2). In terms of categorical PA complexity, we found that high PA complexity (>0.6) had 83% lower odds of being in the IWOC group (OR = 0.17, 95% CI = 0.05–0.61, p = 0.007) compared to the CU group in the fully adjusted models (Table 2; Model 3). The associations remained significant after adjusting for APOE e4 carrier status but diminished after further adjusting for gait speed (Supplementary Table 2). For variability of PA complexity, we found that every 0.1 unit higher in the variability was associated with 49% lower odds of the IWOC group (OR = 0.51, 95% CI = 0.27–0.98, p = 0.043) in the fully adjusted models (Table 2; Model 3). These associations did not differ by APOE e4 carrier status (p for interaction terms > 0.05). However, we found that sex significantly modified the association between variability of PA complexity and cognitive diagnosis (p for interaction term for SCC = 0.045). In the stratified analyses by sex, average PA complexity was significantly associated with IWOC only in females and not in males (Supplementary Table 3).

TABLE 2.

Physical activity complexity and cognitive diagnosis.

	Model 1			Model 2			Model 3
Parameter	OR	95% CI	p‐value	OR	95% CI	p‐value	OR	95% CI	p‐value
Continuous PA complexity (every 0.1 unit higher)
CU	ref	ref	ref	ref	ref	ref	ref	ref	ref
IWOC	0.67	0.48‐0.93	0.015	0.72	0.51‐1.02	0.066	0.64	0.43‐0.94	0.024
SCC	0.63	0.41‐0.97	0.035	0.61	0.38‐0.99	0.046	0.61	0.37‐1.00	0.050
MCI	0.95	0.77‐1.19	0.667	0.97	0.77‐1.22	0.762	0.99	0.77‐1.27	0.933
AD	0.68	0.40‐1.17	0.162	0.79	0.45‐1.37	0.395	0.78	0.44‐1.36	0.374
High vs. low PA complexity
CU	ref	ref	ref	ref	ref	ref	ref	ref	ref
IWOC	0.32	0.12‐0.89	0.030	0.26	0.08‐0.82	0.022	0.17	0.05‐0.61	0.007
SCC	0.27	0.07‐1.06	0.061	0.28	0.07‐1.12	0.072	0.28	0.07‐1.17	0.082
MCI	0.77	0.36‐1.68	0.515	0.69	0.30‐1.55	0.365	0.72	0.31‐1.67	0.440
AD	0.15	0.02‐1.31	0.087	0.15	0.02‐1.41	0.096	0.13	0.01‐1.33	0.086
Variability of PA complexity (every 0.1 unit higher)
CU	ref	ref	ref	ref	ref	ref	ref	ref	ref
IWOC	0.85	0.52‐1.38	0.502	0.64	0.36‐1.14	0.128	0.51	0.27‐0.98	0.043
SCC	0.76	0.40‐1.44	0.397	0.80	0.41‐1.53	0.493	0.84	0.41‐1.71	0.630
MCI	0.84	0.56‐1.27	0.413	0.78	0.51‐1.20	0.258	0.79	0.50‐1.27	0.332
AD	0.56	0.21‐1.48	0.242	0.57	0.21‐1.55	0.273	0.53	0.19‐1.51	0.234

Note: Model 1 was the unadjusted model. Model 2 was adjusted for age, sex, race, and education years. Model 3 was additionally adjusted for comorbidities and total activity counts. Low complexity was defined as ≤0.6. Bolded values indicate statistically significant results (p < 0.05).

Abbreviations: AD, Alzheimer's disease; CI, confidence interval; CU, cognitively unimpaired; IWOC, impaired without complaints; MCI, mild cognitive impairment; OR, odds ratio; PA, physical activity; SCC, subjective cognitive complaints.

Regarding cognitive performance measures, both CFA models had a good fit with RMSEA = 0.073, 90% CI: 0.053–0.092, CFI = 0.950, and SRMR = 0.050 for global cognitive function, and RMSEA = 0.075,90% CI: 0.056–0.095, CFI = 0.947, and SRMR = 0.053 for three cognitive domains. The associations of PA complexity with global cognitive function and each cognitive domain are presented in Table 3 and Figure 1. Unadjusted linear regression models showed that every 0.1‐unit higher in PA complexity was associated with 0.23‐unit higher (β = 0.23, SE = 0.11, p = 0.041) in global cognitive function, 0.10‐unit higher (β = 0.10, SE = 0.04, p = 0.014) in executive function, and 0.09‐unit higher (β = 0.09, SE = 0.04, p = 0.014) in language (Table 3; Model 1). The association was diminished after adjusting for covariates. However, using the dichotomous PA complexity measure, we found that participants with high PA complexity (>0.6) had 0.88‐unit higher in global cognitive function (β = 0.88, SE = 0.34, p = 0.011), 0.29‐unit higher in memory (β = 0.29, SE = 0.12, p = 0.020), 0.38‐unit higher in executive function (β = 0.38, SE = 0.13, p = 0.003), and 0.34‐unit higher in language (β = 0.34, SE = 0.12, p = 0.005) compared to participants with low PA complexity in the fully adjusted models (Table 3; Model 3). No significant associations were found after further adjustment for APOE genotype and gait speed (Supplementary Table 4). We did not find significant association between day‐to‐day variability of PA complexity and cognitive function in the fully adjusted models (Table 3; Model 3). For sensitivity analyses, the associations between average PA complexity and cognitive function did not differ by APOE e4 status. However, APOE e4 status significantly modified the association between variability of PA complexity and executive function (p for the interaction term = 0.042). The results stratified by APOE e4 status are presented in Supplementary Table 5. Sex and MCI/AD diagnosis did not modify the association between PA complexity and cognitive function (p for interaction terms>0.05; Supplementary Figure 5).

TABLE 3.

Physical activity complexity and cognitive function.

	Model 1			Model 2			Model 3
Parameter	β	SE	p‐value	β	SE	p‐value	β	SE	p‐value
Global cognitive function
Continuous PA complexity (every 0.1 unit higher)	0.23	0.11	0.041	0.14	0.10	0.160	0.12	0.11	0.246
High vs. low PA complexity	0.90	0.39	0.022	0.92	0.33	0.007	0.88	0.34	0.011
Variability of PA complexity (every 0.1 unit higher)	0.17	0.20	0.387	0.26	0.17	0.142	0.22	0.19	0.243
Memory
Continuous PA complexity (every 0.1 unit higher)	0.07	0.04	0.078	0.04	0.03	0.261	0.03	0.04	0.347
High vs. low PA complexity	0.29	0.14	0.037	0.30	0.12	0.014	0.29	0.12	0.020
Variability of PA complexity (every 0.1 unit higher)	0.05	0.07	0.463	0.08	0.06	0.189	0.07	0.07	0.277
Executive function
Continuous PA complexity (every 0.1 unit higher)	0.10	0.04	0.014	0.06	0.04	0.071	0.05	0.04	0.151
High vs. low PA complexity	0.41	0.14	0.004	0.40	0.12	0.001	0.38	0.13	0.003
Variability of PA complexity (every 0.1 unit higher)	0.09	0.07	0.233	0.10	0.06	0.111	0.08	0.07	0.240
Language
Continuous PA complexity (every 0.1 unit higher)	0.09	0.04	0.014	0.06	0.03	0.066	0.05	0.04	0.156
High vs. low PA complexity	0.37	0.13	0.007	0.37	0.12	0.002	0.34	0.12	0.005
Variability of PA complexity (every 0.1 unit higher)	0.10	0.07	0.143	0.13	0.06	0.042	0.10	0.07	0.119

Note: Model 1 was the unadjusted model. Model 2 was adjusted for age, sex, race, and education years. Model 3 was additionally adjusted for comorbidities and total activity counts. Low complexity was defined as ≤0.6. Bolded values indicate statistically significant results (p < 0.05).

Abbreviations: PA, physical activity; SE, standard error.

FIGURE 1.

Physical activity complexity and cognitive domains. The cognitive domain factor scores were derived from confirmatory factor analysis. EF, executive function; LAN, language; MEM, memory.

The associations between PA complexity and AD‐related plasma biomarkers were visualized in Supplementary Figure 6. For continuous blood‐based biomarkers, no significant results were found in linear regression models (Supplementary Table 6). For tertiles of blood biomarkers, no associations were found with average PA complexity (p > 0.05; Figure 2), but we found that greater variability of PA complexity was associated with more p‐tau181, p‐tau231, and p‐tau217 in multinomial logistic regression models (Figure 2). Specifically, every 0.1 unit higher in variability of PA complexity was associated with 65% greater odds of the highest levels of p‐tau181 (OR = 1.65, 95% CI = 1.01–2.69, p = 0.046), 70% greater odds of the highest levels of p‐tau231 (OR = 1.70, 95% CI = 1.03–2.80, p = 0.040), and 67% greater odds of the highest levels of p‐tau217 (OR = 1.67, 95% CI = 1.03–2.71, p = 0.037) after adjusting for sociodemographics, comorbidities, and total activity counts (Figure 2). The association between variability of PA complexity and p‐tau181 remained significant after further adjusting for APOE genotype and gait speed (Supplementary Table 7). We did not find significant moderating effects of sex or APOE e4 carrier status on the association between PA complexity and plasma biomarkers. The stratified analyses are presented in Supplementary Table 8 to Supplementary Table 11. The variability of PA complexity was associated with p‐tau levels in females but not in males (Supplementary Table 11). The associations between PA complexity measures and plasma biomarkers did not differ by cognitive diagnosis. However, the association between day‐to‐day variability of PA complexity and p‐tau181 levels remained significant after further restricting the sample to participants without MCI or AD diagnosis (Supplementary Table 12). Every 0.1 unit higher in day‐to‐day variability of PA complexity was associated with approximately two‐fold higher odds of being in the highest tertile of p‐tau181 (OR = 2.44, 95% CI = 1.12–5.34, p = 0.025) and p‐tau217 (OR = 2.26, 95% CI = 1.10–4.67, p = 0.027) (Supplementary Table 12).

FIGURE 2.

Forest plot for multinomial logistic regression models of average PA complexity (A) and day‐to‐day variability of PA complexity (B) with blood‐based AD‐related biomarkers. Multinomial logistic regression models were adjusted for age, sex, race, education years, comorbidities, and total activity counts. AD, Alzheimer's disease; CI, confidence interval; OR, odds ratio; PA, physical activity.

4. DISCUSSION

The study demonstrated that lower daytime PA complexity was associated with subtle subjective cognitive concerns and poorer cognitive performance and showed modest ability to discriminate these groups from cognitively unimpaired controls. Lower daytime PA complexity was significantly associated with poorer performance in global cognition, memory, executive function, and language. Greater day‐to‐day variability in PA complexity, but not the mean level, was modestly associated with higher plasma p‐tau levels. These associations remained significant after excluding participants who had already developed MCI or AD. These findings suggest that the low complexity of daytime activities is associated with poorer cognition beyond conventional volume or intensity measures and may emerge in older adults at the early stages of the cognitive continuum. More stable activity patterns across days may be linked to less AD‐related biomarker burden. Future longitudinal studies are needed to establish temporality and causality between PA complexity and ADRD risk.

Our study extends the emerging evidence that lower movement complexity predicts worse cognition and elevated risk of incident MCI and dementia. ¹⁴ We used a novel metric to quantify the complexity of continuous accelerometer signals. Based on the complexity theory of aging, ²⁶ , ⁴³ age‐related decrease in complexity of neurophysiological signals reflect impaired physiological function and reduced adaptability to stressors. ²⁶ , ⁴⁴ Such reductions in complexity of neurophysiological signals have also been consistently associated with age‐related diseases. ¹⁶ , ¹⁸ For instance, decreased complexity in standing postural sway predicts a higher risk for falls, ¹⁶ and reduced complexity in continuous beat‐to‐beat systolic blood pressure signals has been linked to increased dementia risk. ¹⁸ Applying the complexity theory to daily movement, we hypothesized that cognitively healthy older adults had diverse and variable activity patterns within a day, whereas individuals with cognitive impairment generally demonstrate more repetitive and uniform activity patterns which lead to lower complexity of PA. However, in the current study, we did not observe reduced complexity in participants already diagnosed with MCI or AD. This may suggest that complexity loss emerges earlier in the disease progression and may be a sensitive marker in the transition from cognitively normal status to SCC or IWOC, but not to an MCI diagnosis. The observed selective complexity reduction in groups with SCC and IWOC, but not in those with cognitive impairment, supports the notion that declining movement complexity may serve as an early behavioral marker preceding clinically significant cognitive decline.

We found that lower PA complexity was significantly associated with poorer cognitive performance in multiple domains, including memory, executive function, and language. This evidence aligns with our previous work demonstrating that reduced PA complexity was associated with poorer performance in global cognitive function, particularly executive function. ¹⁴ Blodgett and colleagues used detrended fluctuation analysis (DFA) to quantify fractal complexity of daily PA and found that higher fractal complexity was associated with higher composite z‐score and the associations were strongest for verbal fluency and weakest for immediate and delayed recall. ⁴⁵ Executive function is essential for initiation of goal‐directed behaviors, planning, and sequencing of complex tasks. ⁴⁶ The observed link between low PA complexity and executive function may reflect diminished cognitive flexibility and reduced capacity to organize varied daily activities. With declined executive function, older adults tend to engage in more monotonous or repetitive routines, which is reflected by reduced movement complexity. ⁴⁷ Similarly, memory impairment, especially in episodic memory, can disrupt the ability to recall and organize activities across time, which may also lead to reduced diversity of daily behaviors. These findings suggest that accelerometer‐derived complexity can provide meaningful insights into cognitive status, and future longitudinal studies are needed to examine whether reduced PA complexity is associated with cognitive decline over time.

To our knowledge, this is the first study to link day‐to‐day instability of PA complexity to Aβ and tau pathology in plasma. Previous studies have linked variability in activity patterns to AD biomarkers and found that elevated brain Aβ was associated with lower total amount and lower within‐day variability of PA in overnight/early evening in cognitively unimpaired older adults. ⁴⁸ , ⁴⁹ Higher day‐to‐day variability in accelerometer‐derived circadian and sleep measures has also been associated with cerebrospinal fluid (CSF) and plasma Aβ and tau. ⁵⁰ In the current study, we found that higher day‐to‐day variability in PA complexity is associated with elevated plasma tau levels. These plasma tau isoforms are considered among the most sensitive and specific peripheral markers of AD pathology, particularly p‐tau231, which has been shown to increase early in preclinical AD prior to overt amyloid positron emission tomography (PET) positivity. ⁵¹ , ⁵² However, these effects were modest, and we did not find strong dose‐response relationships when biomarkers were examined as continuous variables. We also found mixed findings for cognitive diagnostic groups where higher day‐to‐day variability in PA complexity was associated with lower odds of being in the IWOC group. Given the small sample size for each diagnostic group and the cross‐sectional design, we interpret these findings cautiously and future studies are needed to replicate and validate our findings. Associations appeared stronger in APOE e4 non‐carriers than in carriers and in females compared with males. Given the small sample sizes in the stratified analyses, these results should be interpreted with caution. There has been limited evidence on sex differences in the link between PA complexity and AD biomarkers. Several studies have focused on rest‐activity rhythms or fractal motor activity regulation using DFA, but the results on sex differences have been inconsistent. Van Egroo and colleagues have reported that the association between disrupted circadian rest‐activity patterns and AD pathophysiological processes was more evident in cognitively unimpaired males. ⁵³ One study did not find sex‐differences in these associations, ⁵⁴ while another study found that the association between degradation in fractal motor activity regulation and PET‐derived Aβ burden or CSF p‐tau181/Aβ42 ratio was largely driven by females. ⁵⁵ These inconsistent results may be driven by the heterogeneity of sample characteristics, various assessments of AD biomarkers, and different measures of daily activities. Future longitudinal studies are needed to better understand the sex differences in the associations between daily activities and preclinical AD.

The strengths of the study included objective measurements of daily activity, novel approaches of quantifying activity patterns from continuous accelerometer data, use of novel AD plasma biomarkers, and detailed cognitive classifications. There were some limitations that need to be considered. The temporality and causality cannot be determined due to the cross‐sectional study design. The small sample size also limited the statistical power to detect associations between biomarkers and PA complexity, especially in the APOE e4 status‐ and sex‐stratified analyses. Due to the lack of sleep diary, we estimated the typical sleep hours from 11pm to 5am which may not be accurate to capture sleep time. Future research should examine PA complexity during waking and sleep hours and examine how sleep measures play a role in these associations.

In conclusion, lower daytime PA complexity and greater day‐to‐day instability of complexity are linked to subtle cognitive impairment and higher plasma p‐tau, respectively, in community‐dwelling older adults. These findings suggest that PA complexity at free‐living movement may differ by stages of cognitive status. Longitudinal studies with repeated biomarker, cognitive, and neuroimaging assessments are needed to establish temporal and causal associations and examine whether declines in PA complexity may serve as a preclinical indicator of ADRD.

CONFLICT OF INTEREST STATEMENT

The authors do not have conflict of interest to disclose.

CONSENT STATEMENT

The studies involving human participants were reviewed and approved by Human Research Protection Office, University of Pittsburgh. The patients/participants provided their written informed consent to participate in this study.

Supporting information

Supporting Information

Supporting Information

Cai Y, Snitz B, Cohen AD. Associations between accelerometer‐derived physical activity complexity, cognitive function, and plasma biomarkers of Alzheimer's disease. Alzheimer's Dement. 2026;22:e71181. 10.1002/alz.71181