Background : Temporal movement behavior patterns (i.e., physical activity and sedentary behavior patterns) consist of combinations of activity dimensions. The aim of this work is to develop the largest inventory to date of metrics and methods for describing patterns in movement behaviors. In addition, it seeks to establish a taxonomy for categorizing these metrics and methods, enabling researchers to better understand their similarities and differences. Methods : A systematic search was conducted across four databases—Ovid Medline, Ovid Embase, Web of Science, and Scopus—from their inception to June 10, 2021. Studies that reported metrics or methods describing physical activity and sedentary behavior patterns measured with wearable devices were included. An inventory of these metrics and methods was created, which informed the development of a taxonomy for categorizing temporal movement behavior patterns. Results : From 22,074 unique records, 2,541 were included in the review. The review identified 161 standalone metrics and methods, which were mapped onto a taxonomy consisting of four stages and nine nodes: (I) data preparation: (1) data format and (2) input derivation method; (II) input measures: (3) posture, (4) volume, (5) frequency, (6) duration, and (7) intensity; (III) output measures: (8) pattern definition; and (IV) output derivation method: (9) computational methods on input. Conclusion : The derived taxonomy and inventory provide an overview of commonly used metrics and methods for examining temporal patterns and offer guidance for their application and future research. An online version of the taxonomy and inventory resource is available.
Wearable devices are essential to physical activity research, yet heterogeneous calibration protocols yield noncomparable estimates of key metrics (sedentary time, step counts, and moderate to vigorous physical activity). We introduce the PhasED (Phased Evaluation and Development for Wearable Algorithms) framework, an evolution of a prior phase-based framework, that embeds measurement theory principles to improve comparability and real-world performance of algorithms. PhasED integrates three elements: (a) establish acceptable reference measurements and procedures with the goal of real-world application for key behaviors; (b) harmonize development and calibration procedures to enhance comparability of algorithm output across devices and wear locations; and (c) conduct rigorous real-world testing according to Phase III of PhasED, which requires the use of established reference measures in independent study populations to quantify algorithm accuracy and precision. We propose the use of several reference measures by summarizing evidence for the accuracy and precision of acceptable reference measures for three core outcomes: sedentary time, moderate to vigorous physical activity, and step counts. We also identify key research gaps and priorities: consensus operational definitions, commutable calibration protocols, and open-source benchmarking data sets for evaluation that span diverse populations and devices. Broad adoption of PhasED can align algorithm development across research groups and devices, yielding estimates that are accurate, precise, and comparable across studies, thereby strengthening translation into guidelines and public health surveillance.
Aim: This study compared posture and movement classification accuracy and agreement of two thigh-worn accelerometers (SENS and Axivity) among children aged 3-14 years using the same classification algorithm. Method: Forty-eight children attended a structured 1-hr laboratory session, of which 46 provided valid data. The session was video recorded and accelerometer data were collected from accelerometers worn on the front of the thigh. Human-coded video provided the reference standard to calculate accuracy of ActiMotus algorithm posture and movement classifications. Agreement between devices was compared using multiple metrics. Accelerometer variables (acceleration variation and sensor orientation metrics) used for classification were also compared between devices. Results: The devices showed comparable classification accuracy (F1-score = 66%-68%; balanced accuracy = 82%-83%) with a trend for SENS to have slightly higher accuracy, but the only significant difference was found for running where SENS had lower metrics (F1-score = 55%; balanced accuracy = 77%) than Axivity (F1-score = 56%; balanced accuracy = 81%). Agreement in posture and movement duration estimates was highest for lying, sitting, and standing (85%-88%), but lower for dynamic movements, particularly running (65%) and stair climbing (53%). When comparing the accelerometer variables, the greatest relative difference was observed for the SD of x-axis for running (mean absolute percent error = 18.4%, SD = 8.9). Conclusions: We observed differences between SENS and Axivity for posture and movement classification accuracy and agreement which may be linked to differences in hardware features and thereby derived accelerometer variables. The results suggest that hardware choice might influence estimates of postures and movements among children even when using the same placement and algorithm.
Purpose: To assess the criterion validity of chest-worn wearable camera (WC) still-image (IMG) annotations to capture the frequency and duration of physical activity and sedentary behavior against direct observation video within a sample of healthy adults. Methods: Forty-eight participants (Mage: 48 +/- 25 years, 42% women) were asked to wear a WC while performing their habitual activities in a free-living setting and be video-recorded for 2 hr, for up to three study visits. Videos and IMGs were annotated with platforms derived from the Compendium of Physical Activities, with platforms containing one schema focused on detailing behavior and another schema detailing posture. Annotation codes were collapsed into behavior domains, posture domains, and 13 behavior types. Classification metrics were computed to assess the validity of WC IMGs to capture the frequency of domains and behavior, whereas validity to capture duration was done using equivalency testing. Results: 253 hr of data from 131 study visits were collected. Sensitivity, specificity, precision, and F1 score for WC IMG annotations within household, leisure, occupation transportation, sit, and five behavior types were all above 0.9. Stand and five other behavior types had metrics above 0.8. Bias and 90% confidence intervals were within +/- 5% of direct observation video for household, transportation, sit, electronics, sports/exercise, and vehicle travel; +/- 10% for leisure, occupation, and seven behavior types; and +/- 15% for stand. Conclusions: Annotations from WC IMGs can provide criterion measures of physical activity and sedentary behavior across all behavior domains, sedentary and upright posture, and 10 different behaviors.
Workplace health promotion programs target office-based workers to sit less. Measuring posture and location at work, to assess the effectiveness of these programs, is generally undertaken using wearable monitors or video, adding burden or privacy concerns for workers. Millimeter wave is a technology that uses millimeter wavelength electromagnetic waves, which can provide a highresolution 3D point cloud representation of humans and, when processed using artificial intelligence, has the potential to reconstruct human movement and location. This pilot study assessed the accuracy of millimeter wave predictions of posture (sitting/standing) and presence at desks (1 m) for workers in an open-plan office room (4.6 & times; 3.6 m, three desks) alone or with other workers present (up to three). The referent was an observed timed protocol with 2,340 conditions assessed. The data sets were concatenated and used for training classification models using a leave-one-file-out cross-validation method, partitioning remaining files into training (80%) and validation (20%) sets. Accuracy for prediction of whether a person was present or not was high (F1 score > 0.95) but lower for prediction of sitting (F1 score= 0.65-0.70) and standing (F1 score= 0.40-0.86). This pilot data provides evidence that millimeter wave could be used as a low burden method of assessing presence at the desk in the workplace. More research into identifying if workers are sitting or standing is needed. Field-based evaluations in real office situations should be undertaken.
Background: Physical activity (PA) is a crucial modifiable risk factor for mortality among older adults. While most studies use total PA volume as a summary metric, continuously monitored accelerometer data can capture more detailed daily PA patterns. This study compares the ability of classical, machine learning, and deep learning survival models to predict mortality using temporal-pattern-based PA features. Methods: We analyzed data from 3,032 individuals aged >= 50 years from the 2011 to 2014 cohort of the National Health and Nutrition Examination Survey (NHANES), linked with mortality data in the National Death Index. Minute-level accelerometer data were transformed into functional principal components to capture the temporal dynamics of daily activity. We evaluated four survival models: Cox Proportional Hazards, Random Survival Forest, Gradient Boosted Survival Model, and DeepSurv for mortality prediction using Harrell's C-index and time-dependent area under curve (AUC) over 100 random train-test splits. Results: Based on the Cox Proportional Hazards model, a higher daily PA during the daytime was found to be associated with a reduced hazard of all-cause mortality, while a higher PA during night, indicating disrupted sleep, was found to be associated with a higher hazard of all-cause mortality. The Random Survival Forest model exhibited the highest predictive performance, with a mean C-index of 0.80 and a mean time-dependent AUC of 0.91. This was superior to DeepSurv (C-index: 0.78, AUC: 0.85), Gradient Boosted Survival Model (C-index: 0.78, AUC: 0.83), and the traditional Cox Proportional Hazards model (C-index: 0.77, AUC: 0.82). Notably, machine learning models Random Survival Forest and Gradient Boosted Survival Model demonstrated significant gains when using functional principal components-based features compared to simpler mean activity-based models. Conclusions: Functional features derived from daily activity patterns provide richer interpretation of time-of-day patterns than traditional summary statistics. When combined with flexible machine learning models, they can uncover the complex, nonlinear relationships between PA and survival, improving mortality prediction for older adults.
Background: The minimum upright period (MUP) and minimum nonupright period (MNUP) in PAL Technologies processing rules define time-based posture-classification thresholds for activPAL recordings. Although activPAL shows high validity for posture classification, the influence of these thresholds on detecting brief sit-to-stand (STS) transitions remains unclear. Research Question: Do different MUP/MNUP settings (1, 2, 5, and 10 s) affect the accuracy and equivalence of STS counts, and do they influence walking-related outcomes? Methods: Twenty-three healthy men performed 10 STS transitions at fixed interstand intervals (1, 3, 5, and 10 s) while wearing a thigh-mounted activPAL4. Data were reprocessed using symmetric MUP/MNUP settings (1, 2, 5, and 10 s). STS count, walking time during a standardized 100-step task, and step count were compared with reference values using Friedman tests, Bland-Altman analyses, mean absolute error, and equivalence testing with two one-sided tests (equivalence bounds +/- 1 STS). Results: STS detection was strongly dependent on MUP/MNUP selection. At interstand intervals >= 3 s, MUP/MNUP 1-2 s yielded STS counts equivalent to the reference (two one-sided test p < .001, mean absolute error 5 0.1), whereas longer thresholds (5-10 s) substantially undercounted STS. At 1-s intervals, all settings underestimated STS. Walking time and step count showed minimal bias and were unaffected by MUP/MNUP selection. Significance: STS quantification from activPAL data depends on processing thresholds rather than device hardware. Short MUP/MNUP settings (1-2 s) are required for accurate STS detection when transitions are brief, whereas walking-related outcomes are robust. Explicit reporting of MUP/MNUP is essential for reproducibility.
Large language models (LLMs) are a new paradigm in artificial intelligence. Their training on vast data sets enables them to learn complex patterns and relationships, resulting in a high degree of generalization and adaptability across a wide range of tasks, often without requiring extensive retraining. Since the emergence of LLMs, their adoption has rapidly expanded across various domains and applications. Recently, there has been a growing trend toward integrating LLMs to interpret and analyze wearable-generated physiological and behavioral data streams for tasks such as health prediction, data interpretation, and the generation of personalized recommendations. Here, we highlight four important characteristics of LLMs and discuss how these could apply to the analysis and interpretation of wearable-measured data on physical behaviors. The synergy between LLMs and wearable sensor data appears to hold a promise to open exciting opportunities for advancing research at the intersection of physical behaviors, health, wearables, and behavioral science.
Background:Consumer Technology Association guidelines recommend two independent researchers review video recordings of treadmill bouts to verify accuracy of the number of steps taken. Performing multiple post hoc validation checks requires significant resources for personnel time and costs, but it is unknown if the burden is necessary. The purpose of this study was to evaluate accuracy and efficiency of step counts assessed during direct observation compared with multiple reviews of video recordings. Methods:Participants ages 18-20 years (n = 72) completed up to twelve 5-min, zero-incline treadmill bouts, speeds ranging 0.2-2.7 m/s (0.5-6.0 mph). Directly observed steps were counted and hand-tallied in real-time. A sex-age stratified random sample of 30 participants' bouts were selected for analysis. Two independent researchers recounted steps from video recordings, producing three independent counts. Intraclass correlation coefficient determined interrater reliability. Mean average percent deviation computed deviation magnitude among the three counts. Correlation tests determined the relationship between absolute count differences and average counts for each pairwise comparison. Results:Interrater reliability was excellent with intraclass correlation coefficients ≥ .995 (p < .001) across all speeds. Average mean average percent deviation was 0.29% with bout-specific deviations ranging from 0.10% to 0.54%, indicating < 1% difference among the three counts for all bouts. Conclusions:A single real-time step count represents a sufficient and efficient approach to measure steps taken in a controlled treadmill setting, while video recounts serve as back-up to allay potential data loss concerns and/or for random selection verification. These data-driven findings are useful for optimizing personnel time and reducing costs associated with collecting quality data in large-scale treadmill-based step-counting studies.
Background: This study explored the acceptability of an intensive ambulatory protocol combining four wearable devices and ecological momentary assessment (EMA) to assess physical and eating behaviors. Methods: EMA and wearable data were collected over 7 days in a convenience sample of 622 participants (56.1% women, Mage=38.2 years) in Ireland, Germany, France, and the Czech Republic. Short EMA questionnaires (eight to 17 items) assessing current activity, context, and mood were triggered randomly (seven per day) and in response to behavioral patterns detected by Fitbit, such as prolonged sitting (max four per day), walking, or running (each max three per day). EMA compliance was assessed by the proportion of completed questionnaires. Among four wearables worn (Fitbit/ActivPAL/Actigraph/LifeQ), compliance was measured using Fitbit heart rate recordings. Acceptability for EMA and wearables was assessed using 5-point Likert-scale questionnaires on daily burden, ease of use, interference, and ethical satisfaction. Associations with sociodemographic and health factors were analyzed. Results: Participants received on average 11.5 (+/- SD 1.9) EMA questionnaires per day. EMA acceptability scores ranged from 3.8 (burden) to 4.6 (ethical satisfaction), while wearable acceptability scores ranged from 4.1 (burden/reactivity) to 4.6 (ease of use/ ethical satisfaction). Low perceived daily burden was associated with higher compliance for EMA wearables (odds ratio: 1.88, 95% CI [1.30, 2.72] and 1.74, 95% CI [1.08, 2.80], respectively), while smartphone interference increased with age (odds ratio: 0.62, 95% CI [0.50, 0.78]). No association was found with sex and educational level. Conclusions: Acceptability of EMA and wearables was generally satisfactory. As EMA compliance may be affected by daily burden, well-designed protocols are key to leveraging ambulatory assessment of real-time behavioral and contextual data.
Background: Various summary metrics have been proposed to standardize the processing of raw accelerometer data including Data were from the Free-Living Activity Study for Health, in which participants (N= 78, age: 23.2 +/- 8.4 years, 64% female, 87% white, body mass index: 24.4 +/- 3.5 kg/m2) wore an ActiGraph wGT3X-BT monitor on each wrist, plus GENEActiv and Axivity AX3 monitors (one on each wrist, counterbalanced for equal placements on dominant and nondominant sides). Participants wore devices for 24 hr and reported their physical activity via the Activities Completed over Time in 24 hr web survey. Summary metrics were obtained using open-source methods. The agreement was evaluated by equivalence testing (reference, ActiGraph) and Bland-Altman plots. The mean absolute percent error was calculated to assess relative individual errors. Results: The AC, AI, and MIMS yielded 10% statistical equivalence across all brands. Trends were less consistent for ENMO, MAD, and ROCAM. Lowest overall mean absolute percent error was seen for MIMS (13.2 +/- 16.7%), with higher values for AI (22.7 +/- 50.5%), AC (25.1 +/- 146.9%), ENMO (44.1 +/- 79.2%), MAD (45.3 +/- 66.3%), and ROCAM (98.5 +/- 115.1%). Similar patterns were observed within specific intensities segmented by the Activities Completed over Time in 24 hr. Conclusion: The results support the use of MIMS for device-agnostic physical behavior monitoring, whereas interdevice agreement was weaker for the other summary metrics.
Aim: To evaluate the criterion validity of the SENS Motion Activity algorithm to classify postures and movements among children aged between 3 and 14 years in a laboratory setting by comparing with human-coded video. Method: Data were collected on 48 Australian children who attended a structured similar to 1 hr data collection session at a laboratory with their caregivers. The session was video recorded, and thigh acceleration was measured using a SENS accelerometer. Data from the accelerometer were processed and classified into four postures and movements using the SENS algorithm. Human-coded video provided the reference standard to calculate the performance metrics sensitivity, specificity, precision, F1-score, and balanced accuracy. Results: Overall, the SENS Motion Activity algorithm classified postures and movements with performance metrics F1-score of 73% and balanced accuracy of 87%. Lying/sitting had the highest F1-score and balanced accuracy (99%). Standing, walking, and running had somewhat lower performance metrics (F1-score and balanced accuracy of 69%-77%, 57%-92%, and 68%-82%, respectively). Difficulties with human coding likely contributed to lower performance. A higher balanced accuracy was found for boys compared with girls for running. The algorithm had better performance for children older than 11 years compared with younger children overall and for walking and running. Conclusion: Our results suggest that the SENS Motion Activity algorithm can provide accurate measurements for detecting lying/sitting, standing, walking, and running among children. Further research could usefully explore algorithm development for classifying movement in sitting and standing postures and brief sporadic movements, especially among children younger than 11 years.
Background: Accurate and reliable measurements of activity type and intensity are important for monitoring health and wellbeing. Purpose: Determine the validity and interrater reliability of the pocket-worn Fibion device for classifying activity type and intensity in a laboratory setting. Methods: Forty-nine participants (22.6 +/- 4.5 years, 56% male) wore a Fibion in each pocket (dominant and nondominant) while performing approximately 10 activities for 5 min each. To assess validity, classification accuracy for Fibion-predicted activity type (sitting, standing, walking, cycling, and high intensity) and intensity (sedentary, light, and moderate to vigorous) were compared with direct observation using percent agreement and kappa at both the epoch and activity levels. To assess interdevice reliability, predicted activity type and intensity for each epoch and activity were compared between devices worn in the dominant and nondominant pockets using percent agreement, kappa, equivalence testing, linear mixed models, and correlations. Results: For validity, classification accuracies for single activities ranged from 1.0% to 98.1% and 0.5% to 95.0% for activity type and 22.7% to 98.1% and 10.3% to 98.8% for intensity for dominant and nondominant devices, respectively. For interdevice reliability, there were no significant differences in time spent in any activity type or intensity between the two devices. Activity- and epoch-level correlations were moderate to high for all variables (r = .56-.99) except for cycling (r= .32). Conclusion: In the laboratory, pocket-worn Fibion devices accurately classified sitting, walking at lower speeds, stair climbing, running activities, and most sedentary and moderate to vigorous activity intensities and had moderate-to-high interdevice reliability for all activity types except for cycling.
Adequate handling of missing data on physical activity assessments is crucial in longitudinal accelerometer studies. This study aimed to evaluate the effectiveness of various imputation methods for handling missing data in an empirical application which utilizes wearable accelerometers. We employed a simulation approach to assess performance under different missing data scenarios including Missing Completely at Random, Missing at Random, and Missing Not at Random for a longer study period (6 weeks). Our findings revealed that mean imputation and hot-deck imputation applied with a fine degree of matching criteria (participant, day of the week, and time of day) outperformed discard-based methods under Missing Completely at Random and Missing at Random conditions as they produced the smallest bias and best precision. Notably, no imputation methods performed well under Missing Not at Random scenarios. We recommend conducting simulation studies tailored to specific study designs to compare imputation methods, implement strategies for improving data quality, gather information on nonwear periods, and ensure continuous monitoring and participant compliance thereby reducing bias in activity level estimates. If a simulation study is not feasible, we recommend to impute data relying on mean or hot-deck approaches with the finest possible degree of matching criteria.
Background:The ActiGraph GT9X Link wearable accelerometer has been widely utilized for physical activity (PA) measurement but is no longer in production. The performance of the ActiGraph LEAP, regarded as its successor and released in 2023, remains unevaluated in comparison to other accelerometers. This study aimed to establish concurrent validity of the ActiGraph LEAP device through PA measurements and comparison with the GT9X Link. Methods:A total of 26 young adults (mean age = 21 ± 4 years) wore the ActiGraph LEAP and GT9X Link simultaneously for 7 consecutive days on their nondominant wrist. Monitor Independent Movement Summary units were utilized for calculation of daily sedentary time, light PA, and moderate to vigorous PA across valid days of wear (≥10 hr). Nonwear time was determined using a preexisting algorithm. Intraclass correlation coefficient, mean absolute percent error, equivalence tests (with GT9X Link as reference), and Bland-Altman plots assessed agreement between daily values of sedentary time, light PA, and moderate to vigorous PA recorded by each device across all valid days recorded among participants. Results:A total of 149 days of data were compared between the two devices. Across all PA levels, acceptable-excellent agreement was found (mean absolute percent error = 3.86%-13.86% and intraclass correlation coefficient = .94-.99), and Bland-Altman plots showed that at least 90% of recorded data fell within the 95% limits of agreement. Conclusions:The findings suggest that sedentary time and PA measurements using Monitor Independent Movement Summary units from the ActiGraph LEAP compare well with the GT9X Link, indicating that the LEAP performs similarly to the GT9X Link.
We present iPlayer, a video annotation tool for direct observation of human activity. In contrast to other tools generally designed for continuous or intermittent annotation, iPlayer is optimized for annotations of specific time windows. The tool is cross-platform (macOS, Windows, and Linux), open-source, locally run, and configurable to support different labels and time window sizes. Users can annotate multiple activities per time window, and the video can be looped over the user-selected time window (optional), reducing annotation error, particularly for small window sizes. We developed and used the tool to annotate toddler activity and position in 1-s windows to be synchronized with accelerometer data. We also provide examples of how iPlayer could be used with other annotation frameworks. While the tool was optimized for annotations of human activity, we believe it may be of use to any researchers looking to precisely annotate segments of any video recordings.
Accelerometry provides information on an individual’s habitual physical activity and postures. The transition from sit-to-stand involves the cardiovascular, neural, and muscular systems, and is linked with balance disorders when tested in laboratory conditions. The methodology for quantifying sit-to-stand velocity (STSv) in free-living conditions and the association with participant-level and health factors are unclear. A scoping review of the available literature on free-living STSv was performed to investigate methodology, participant, and health-related factors in relation to STSv. A literature search was conducted across Scopus, EMBASE, MEDLINE, CINAHL, and Academic Search Premier databases (initial 2,098 articles), yielding n = 15 articles that measured STSv using an accelerometer in a free-living condition or in preparation for use in free-living conditions ( n = 10 methodological, n = 5 participant/health-related factors). Sensor type, sampling rate, and filtering characteristics were all heterogeneous, while most sensors were placed on the anterior thigh. There was no consensus for algorithm development to identify sit-to-stand transitions or for the quantification of STSv. Among studies, STSv in individuals with frailty, stroke, and older adults were slower compared with healthy controls. The evidence regarding the utility of free-living STSv is limited but encouraging. The physiological and health underpinnings of preserving or improving STSv using interventional designs are highly warranted. This scoping review identifies literature gaps and recommendations for future investigations. STSv is not limited to controlled conditions only and wearable monitors provide insight into this metric during free-living condition, but harmonized sensor data collection and analytical approaches to quantifying STSv are needed.
Physical activity (PA) and sedentary behavior (SB) are deeply ingrained within everyday life. It is important to capture all dimensions of PA and SB given their impact on health where the ideal measurement tool is able to measure the frequency, intensity, duration, and type of PA and SB when they occur. Media from wearable cameras (WCs) have seen increased use within the scientific literature for capturing health behavior, but their use for capturing PA and SB has not been systematically reviewed before. Therefore, the purpose of this scoping review was to determine the extent of WC media use for detailing PA and SB. A variety of record types from seven databases was searched from database inception to November 2023. Seventy-four sources were included, with purposes of WC media for capturing PA and SB for first-person activity recognition, to provide primary or secondary study outcome measures to be used as the ground truth measure for other measurement tools and for validation of WC media estimates of PA and SB. Both still images and video from WCs were used to capture the frequency, intensity, and duration of PA and SB, with 100% of included sources using WC media to capture types of PA and SB. Despite the extensive use of WC media to capture PA and SB, few studies have rigorously established the concurrent validity of WC media estimations of PA and SB against gold-standard measures, and participant characteristics within the first-person activity recognition field are seldom reported.
Few traditional physical behavior estimates capture accumulation patterns. This study introduces and applies motif probability, a novel metric quantifying the occurrence of a specific sequence of bouts ("motif"), enhancing physical behavior pattern analysis. Motif probability is calculated using a forward algorithm with learned parameters from the hidden semi-Markov model, segmenting accelerometer data into physical behavior sequences. Accelerometer data from 517 participants were processed using hidden semi-Markov models to derive physical behavior states and construct transition matrices, using these parameters to calculate motif probability. In addition, traditional volume-based estimates (e.g., daily average acceleration and time spent in specific intensities) and complexity metrics were calculated. Clustering analysis grouped participants with similar physical behavior sequences, volume-based estimates, complexity metrics, sex, and body mass index z-score. We identified influential estimates with principal component analysis and assessed associations with body mass index z-score and sex with linear and logistic regression, respectively. Five behavioral clusters emerged, characterized by: (a and b) low average acceleration and complexity, (c) highest volume-based estimates, (d) highest complexity, and (e) lowest physical activity and highest sedentary behavior motif probabilities, despite similar volume-based estimates to Cluster 3. Four main components explained 67.4% of the variance: (a) motif probability, (b) volume-based estimates, (c) sedentary behavior and physical activity motif probabilities with complexity, and (d) sedentary behavior estimates. Motif probabilities correlated moderately with complexity and weakly with volume-based estimates, thus capturing complementary aspects of physical behaviors. Motifs were significantly associated with sex but not body mass index z-score. Motif probability captures the occurrence of specific temporal physical behavior sequences, adding temporal detail beyond traditional metrics.
