Time spent sitting increases risk of developing major chronic diseases and is associated with multiple adverse health outcomes (Dempsey et al., 2020;Katzmarzyk et al., 2019;Saunders et al., 2020). Sitting is most often operationalized as amount of time accumulated during a day or week. However, emerging epidemiologic and experimental evidence shows that the temporal pattern of accumulating sitting time in a comparatively prolonged versus more interrupted manner (henceforth referred to as sitting patterns; Tremblay et al., 2017) is also important to public health inquiry (Bellettiere et al., 2017;Diaz et al., 2016;Dunstan et al., 2012;Hartman et al., 2018;Healy et al., 2011;Owen et al., 2020;Saunders et al., 2018). The mounting evidence has led several national and international public health guideline development committees to call for more research on the topic, both to establish its importance in chronic disease etiology and subsequently to support development of quantitative public health recommendations (Dempsey et al., 2020;Katzmarzyk et al., 2019;Saunders et al., 2020).

Accurate, scalable measurement of sitting time and sitting patterns is critically needed to inform future iterations of public health guidelines. Most prospective cohorts (Wijndaele et al., 2015) have used hip-worn accelerometers and processed the data using the 100 counts per minute (cpm) cut point (Migueles et al., 2017). However, this method has low accuracy for measuring posture and postural transitions (Barreira et al., 2015) and consequently leads to low-quality measures of sitting patterns. The estimated percent error in measuring sitting patterns using the 100 cpm is evident across age groups, and ranges from 40% to over 300% depending on the sedentary metric being used (Bellettiere et al., 2021;Carlson et al., 2019). The degree of misclassification generated by this cut point classification leaves some uncertainty around quantitative recommendations for sitting patterns. A few large cohorts (Stamatakis et al., 2020) use devices worn on the thigh that are optimized and often programmed specifically to detect posture and can accurately measure both sitting time and sitting patterns when compared with direct observation (Lyden et al., 2012). However, data from these studies are limited, and the pool of available information regarding sitting patterns could be rapidly expanded by developing new data processing methods for hip-worn accelerometer data to generate measures of sitting patterns that converge with estimates from thigh-worn devices. The application of these new processing methods to both newly collected hipworn accelerometer data and the large amount of archival data could produce accurate and homogenous norms and risk estimates related to sitting patterns and health, both of which are needed to generate quantitative public health recommendations.

Our research group has recently demonstrated that deep learning models can be used to improve the utility of hip-worn monitors (Greenwood-Hickman et al., 2021). Using a combined Convolutional Neural Network (CNN) + bidirectional long shortterm memory network (BiLSTM) model architecture in a sample of older adults, we developed a novel method for predicting sitting postures from hip-worn accelerometer data (e.g., ActiGraph) in older adults by training against concurrently thigh-worn activPAL inclinometers. This algorithm, which we named CHAP (CNN Hip Accelerometer Posture), substantially out-performed previous, widely used data processing methods (e.g., 100 cpm cut point based method) in older adults (≥65 years), achieving 93% agreement in predicted minute-level posture (i.e., sitting vs. upright/ nonsitting) and 83% sensitivity for detecting sit to stand postural transitions (Greenwood-Hickman et al., 2021). CHAP was trained for use in an older adult population, but norms and risk estimates for sitting patterns and how they relate to health are needed across the age spectrum. Especially since many existing cohort studies that collect accelerometer data include younger, middle-aged, and older adults. Therefore, accurate posture classification is needed across a broader age range. Ideally, a single postural detection algorithm could be used to make accurate predictions for all adults.

Building on CHAP to fashion a procedure suitable for adults of a broader age range, we created the CHAP-Adult algorithm by training a new and more generalizable CNN+BiLSTM model using data from adults across a broader age range (35-99 years). We leveraged data from participants of two cohort studies who concurrently wore ActiGraph and activPAL accelerometers for up to 7 days while going about their usual (i.e., free living) behavior patterns (N = 1,397). (Healy et al., 2015;Rosenberg et al., 2020) The objective of this study was to assess the validity and reliability of CHAP-Adult, which we did using a separate, randomly selected validation sample of 419 adults. We also assessed reliability and validity separately by subgroups of age, gender, and body mass index (BMI).

Data for this study come from two distinct studies among adults who concurrently wore an activPAL on their thigh and an Acti-Graph GT3X+ on their hip: the Australian Diabetes, Obesity, and Lifestyle study (AusDiab) and the Adult Changes in Thought (ACT) study.

The AusDiab is a population-based epidemiologic cohort that initially enrolled adults aged 25 years and older throughout Australia. The original complex survey sampling methods were previously published (Dunstan et al., 2002). In brief, three data collection points (1999-2000, 2004-2005, and 2011-2012) were undertaken with each assessment including questionnaires and biomedical assessments conducted at a local testing site. Of the 11,247 adults who completed the baseline assessment in 1999-2000, 4,562 (all ≥ 35 years) attended one of 46 testing centers across Australia in the 2011-2012 follow-up (Tanamas et al., 2013). From this group, a subsample of 1,014 ambulatory community-living participants were invited to join an ancillary activity monitor study, described in detail elsewhere (Healy et al., 2015). Participants were asked to wear the activPAL3 ™ (thigh) and the ActiGraph GT3X+ (hip) for seven consecutive days, with the ActiGraph removed for sleeping and the activPAL worn continuously. Of those approached, 77% (n = 782) provided informed written consent and participated. Protocols for the study were approved by the Alfred Health Human Ethics Committee (project no. 39/11).

The ACT began in 1994 as an ongoing longitudinal cohort study investigating risk factors for the development of dementia and a wide range of cognitive and noncognitive factors of healthy aging. ACT recruited adults aged over 65 years randomly sampled from the King County membership of Group Health Cooperative of Puget Sound (now Kaiser Permanente Washington). As of 2005, ACT continually replaced participants who have died, were lost to follow-up, or who were diagnosed with dementia, in order to maintain a consistent enrollment of approximately 2,000 participants. Study participants attend biennial assessment visits. Beginning with biennial visits in April 2016, ACT participants were invited to participate in an activity monitor substudy (ACT-AM). Those who consented were invited to wear an ActiGraph GT3X+ accelerometer on the hip (ActiGraph LLC), an activPAL micro3 on the thigh (PAL Technologies), or, if willing, both continuously for seven consecutive days. Participants were excluded from ACT-AM if they were wheelchair bound, or receiving hospice or care for a critical illness, or resided in a nursing home, or, if memory problems became evident during testing. In total, 1,211 ACT participants provided written informed consent to participate in ACT-AM and wore at least one device. Details on the ACT-AM cohort were published elsewhere (Rosenberg et al., 2020) and protocols for the study were approved by the Kaiser Permanente Washington Institutional Review Board.

ActiGraph data originally collected at 30 Hz from both studies were obtained so that identical data processing methods could be applied. The 30-Hz data were subsequently converted to 1-min epochs using ActiLife software. ActiGraph nonwear was detected using the commonly used (Migueles et al., 2017) Choi algorithm applied to vector magnitude cpm using a 90-min window, 30-min streamframe, and 2-min tolerance (Choi et al., 2011(Choi et al., , 2012)). For each night of device wear, participants completed sleep logs and the resulting data were used to identify in-bed and out-of-bed time. The activPAL data were processed using PALbatch (version 7.2.32; PAL Technologies, original VANE algorithm) with the default 10-s minimum sitting/upright period and converted to event level files. The activ-PAL "sitting/lying" events are referred to hereinafter as "sitting." ActiGraph and activPAL data were then plotted together with sleep log and nonwear data using heat maps (Greenwood-Hickman et al., 2021) that were then visually inspected by trained analysts.

Figure 1 shows a STROBE diagram detailing participant allocation into the training and validation subsamples. Analyses excluded any time either device was not worn or the person was in bed, and no exclusions were made based on minimum wear time. A total of 51 (6.5%) participants from AusDiab and 231 (19.1%) from ACT did not have concurrently worn devices and were excluded from the analyses. Because ACT participants could choose to wear a single device, this exclusion rate is higher for ACT participants. Data were visually inspected for all AusDiab and ACT participants who wore both devices. In some cases, there was evidence that the devices gradually fell out of sync with each other in a phenomenon called time drift (Steel et al., 2019), possibly due to malfunction or different internal clocks (see the supplemental digital content in Greenwood-Hickman et al., 2021, for an example and description of drift). This occurred in 38 AusDiab participants and 271 ACT participants, with differential time drift rates possibly due to different internal clocks used in the activPAL3 ™ and the activPAL micro. To ensure that our prediction algorithm used the most accurate and optimal criterion data, we excluded samples with time drift for training and for validation.

Data from the resulting 688 AusDiab and 709 ACT participants were then randomly assigned within each cohort to the fortraining data set (70%) or the for-validation data set (30%). To ensure model accuracy was not biased by overfitting that occurs when the model is trained and validated using the same data, the team who trained the CHAP-Adult algorithm used only the fortraining data set and had no access to the for-validation data set.

Sitting Time, Prolonged Sitting Time, Breaks in Sitting, and Sitting Accumulation Total sitting time was measured as the average number of minutes per day sitting. There is no agreed-upon metric for measuring sitting patterns, so we present validation results for postural transitions, prolonged sitting time, and three common pattern indicators based on the duration of sitting bouts. Time spent in ≥30-min bouts (herein termed prolonged sitting) was computed as the average minutes per day spent in bouts ≥30 min. The number of breaks in sitting (sometimes termed sit-stand transitions, or interruptions) is approximately equal to the number of sitting bouts and the number of nonsitting bouts. For convenience, in this study we have used the term "number of breaks" and report the average number of sitting bouts per day. Mean sitting bout duration was computed as the arithmetic mean of all sitting bout durations in minutes. Using procedures outlined by Chastin and Granat (2010) and Chastin et al. (2015), we calculated usual bout duration (in minutes) via nonlinear regression, and alpha (see Supplementary Methods [available online] for equations). Alpha is a unitless parameter that indicates how steep the distribution of bout duration is, which is assumed to be power-law (see Supplemental Figure 1 in Bellettiere et al., 2017). Unlike mean and usual bout duration, higher values of alpha indicate a more interrupted sitting pattern. All measures were based on strict bouts that contained only sitting and did not include modified bouts with a tolerance for upright behaviors.

Details of the machine learning architecture and training procedures have been previously published (Greenwood-Hickman et al., 2021;Nakandala et al., 2021). Briefly, the input data (raw triaxial 30-Hz ActiGraph data) and the ground truth data (activPAL event-level data classified as sitting or not sitting using majority rules for each epoch) during awake wear time were split into 10-s nonoverlapping windows. Features in the ActiGraph data that could differentiate sitting from not sitting were automatically learned using a CNN model. These features were then input into a BiLSTM to learn the features required to recognize temporal patterns that could further distinguish sitting from not sitting. This network was specifically designed to identify the timing of transitioning from sitting to not sitting.

Participant characteristics were summarized for each cohort and for the for-training and for-validation groups and compared between groups with two-sample t tests (continuous variables) or chi-square tests (categorical variables). Validation statistics were calculated and reported only from the for-validation data set. To assess how well CHAP-Adult predicted each instance of sitting versus not sitting at the 10-s epoch, sensitivity, specificity, positive predictive value (PPV), and negative predictive value were computed separately for each individual, then summarized using boxplots (Figure 2). To assess how well CHAP-Adult captured the timing of postural transitions, for each individual separately, we used the transition pairing method (Hibbing et al., 2020) to pair activPAL-labeled transitions and CHAP-Adult-labeled transitions, with a 1min lag time tolerance. From these matches sensitivity (fraction of activPAL transitions matched to a CHAP-Adult transition) and PPV (fraction of CHAP-Adult transitions matched to an activPAL transition) were calculated and summarized using boxplots (Figure 2). Agreement between activPAL and CHAP-Adult measures was assessed for sitting time, prolonged sitting time, breaks in sitting, and sitting accumulation by computing mean error with 95% limits of agreement, percent error, absolute mean error, mean absolute percent error (MAPE), Pearson correlation coefficient, and the Lin's (1989) concordance correlation coefficient. Agreement was assessed for overall and also reported within subgroups based on age (35-49, 50-64, and ≥65 years), sex (male and female), and BMI (<30 and ≥30 kg/m 2 ).

Overall, ages ranged from 35 to 90+ years old in the combined ACT and AusDiab data set. Most participants were women (58.5% of ACT and 56.3% of AusDiab), and approximately 23% had a BMI ≥30 kg/m 2 (Table 1). There were notably more adults with high school education or less in AusDiab (29.6%) than in ACT (8.7%). There was no evidence of randomization failure or imbalanced data sets, with no meaningful differences on measured variables between the for-training and for-validation data sets.

When applied to the for-validation data set, the CHAP-Adult algorithm correctly classified as sitting an average (SD) of 95.3% (4.7%) of the activPAL sitting time (sensitivity), while the algorithm correctly classified 89.8% (8.1%) of activPAL not sitting (specificity). On average 93.5% (6.0%) of the predicted sitting instances and 92.2% (8.4%) of predicted nonsitting instances were correct (PPV and negative predictive value respectively). To accurately measure sitting patterns, it is critical to get the timing of postural transitions correct as well as the total number. An average of 74.4% (10.2%) of the activPAL transitions were detected within a 1-min window by CHAP-Adult (sensitivity) and 77.6% (12.4%) of the transitions that CHAP-Adult predicted were true postural transitions recorded on the activPAL within a 1-min window (PPV).

The estimates of sitting, prolonged sitting, breaks in sitting, and sitting patterns produced using CHAP-Adult correlated strongly with those derived from the activPAL, with Pearson's correlations (r) ranging between .78 and .96. Correlations were lower for alpha (r = .78) and the number of breaks per day (r = .86) than for total sitting time (r = .96), time spent in prolonged (≥30 min) sitting bouts (r = .96), usual bout duration (r = .92), and mean bout duration (r = .92). Concordance correlation coefficients were nearly identical to the Pearson's correlations.

There was good agreement in sitting time, prolonged sitting time, breaks in sitting, and sitting patterns (Table 2), with MAPE <5% for total sitting time and alpha, between 5% and 10% for prolonged sitting time and usual bout duration, and only slightly higher than 10% for mean bout duration and postural transitions. Mean differences (activPAL -CHAP-Adult) were also modest, ranging from 0% to 5.4% of the activPAL value, and indicating on average a small amount of overestimation of total sitting time, mean bout duration, usual bout duration, and prolonged sitting time and underestimation of postural transitions.

Validation results in each population subgroup (Table 3 and Supplementary Table S1 [available online]) resembled those for the overall for-validation sample with consistently high correlations and reasonably modest error that was similar though not identical in each subgroup. For example, BMI-stratified Pearson's correlations were: total sitting time, r <30 kg/m 2 = .96 and r ≥30 kg/m 2 = .95; breaks in sitting time, r <30 kg/m 2 = .88 and r ≥30 kg/m 2 = .79; and mean bout duration, r <30 kg/m 2 = .95 and r ≥30 kg/m 2 = .88, with MAPEs ranging from 3.7% to 16.7% across the different sedentary pattern metrics. Importantly, across age categories, Pearson's correlations were ≥.95 for total sitting time, ≥.82 for breaks in sitting time, and ≥.9 for mean bout duration, with MAPEs ranging from 3.5% to 13.9% for these sedentary pattern metrics across age categories. Error across age categories was larger for usual bout duration MAPE 35-49 = 5.8%, MAPE 50-64 = 6.4%, and MAPE ≥65 = 11.6% and for prolonged sitting MAPE 35-49 = 7.3%, MAPE 50-64 = 8.6%, and MAPE ≥65 = 10.4%, although remaining relatively low with Pearson's correlations ranging between .90 and .99.

Our findings demonstrate the CHAP-Adult algorithm can provide accurate classification of sitting and sitting patterns for men and women over a wide range of adult ages, expanding on our previous successful work with the CNN and BiLSTM model architecture in older adults (Greenwood-Hickman et al., 2021). The algorithm performed well at distinguishing sitting from not sitting (mean balanced accuracy = 93% ± 5%) and detected 74% ± 10% of postural transitions. Consequently, the measures derived using this algorithm displayed good validity not only for total sitting time (MAPE = 4%) but also for breaks in sitting time (MAPE = 11.6%), prolonged sitting time (MAPE = 9.6%), and sitting accumulation patterns (MAPE = 1%-12%), which have historically been more difficult to capture accurately. Agreement metrics were consistent across age, gender, and BMI subgroups, with mean error approximately three times lower than when metrics are computed using the common 100 cpm data processing method (discussed below). Overall, CHAP-Adult, a single algorithm that can be used to detect sitting behaviors across a wide range of adult age groups, represents a significant advancement in the ambulatory measurement of sitting and sitting using hip-worn monitors. If validity in an external adult population can also be demonstrated, it would be suitable for extracting valid measures from the numerous historic studies that have collected raw triaxial hip-worn acceleration data (Wijndaele et al., 2015), and could extend our understanding of the epidemiology and health consequences of sitting.

Sedentary behavior, by definition, is waking time spent sitting or reclining with energy expenditure below 1.5 metabolic equivalents (Tremblay et al., 2017). While hip-worn ActiGraph devices cannot accurately distinguish posture, estimates of total time per day spent sedentary using the most common 100 cpm data processing method are typically close to the estimates of sitting time derived from direct observation or activPAL-with mean errors from several studies across a wide age range of adults ranging from 5% to 20% (Aguilar-Farías et al., 2014;Bellettiere et al., 2021;Koster et al., 2016;Kuster et al., 2020;Lyden et al., 2012). This mean error can be further reduced to 2%-3% by increasing the cut point used to classify time as sedentary to 200 cpm (Aguilar-Farías et al., 2014;Koster et al., 2016), or by implementing either the Kuster algorithm (4% error; Kuster et al., 2020), the original CHAP (0% error; Greenwood-Hickman et al., 2021), or CHAP-Adult (2% error).The similar error rates mean that many of these data processing methods would provide similar results, on average, for measuring total sitting time.

Measuring sitting patterns, on the other hand, has been notoriously challenging. Primary because it is predicated on the accurate identification of breaks between sitting or reclining and being upright, which happen rapidly and relatively infrequently in real life. Accordingly, hip-worn accelerometer data processed using the 100 cpm method have historically overestimated the frequency of sit-stand transitions (also known as breaks in sedentary behavior) in younger (Barreira et al., 2015;Lyden et al., 2017), middle-aged (Kuster et al., 2020), and older (Bellettiere et al., 2021) adults with mean error ranging from 63% to 99%. There are two primary sources of error using hip-worn accelerometers processed using the 100 cpm cut point: (a) long sitting bouts are artificially broken by seated movement such as fidgeting and wiggling; and (b) standing without ambulation gets misclassified as sitting, and then any subsequent stepping while already standing is incorrectly registered as a "sitting break." Both sources of error stem from an inability to distinguish posture when relying solely on the accelerometer signal from vertical acceleration at the hip. To more accurately assess posture using data from hip-worn devices, Kuster et al. (2020) trained a random forest machine learning algorithm (using ground truth data from the activPAL) using 26 predefined signal features from hip-worn ActiGraph data as inputs (Kuster et al., 2020). This algorithm reduced error in the number of sitstand transitions to just 18%, thereby improving the accuracy of sitting pattern measurement. Notably, the deep learning characteristics of CNN, adopted by our CHAP models, automatically determine signal features, thus obviating the need to define a priori features. Furthermore, adding BiLSTM models to the CNN architecture leverages the concept that when predicting human behavior (i.e., here, transitions from one posture to the next), modeling the behavior that occurred directly before and after the behavior aligns the prediction process with theoretical models of behavior. (Hovell et al., 2009) This also effectively leverages the within-person correlated data to improve prediction accuracy. This advancement resulted in CHAP-Adult reducing error in the number of breaks to 4%. Importantly, the low error for estimating the number of breaks using CHAP-Adult was consistent across age, gender, and BMI, demonstrating consistent performance across subgroups of adults.

The U.S. Physical Activity Guideline Advisory Committee and the World Health Organization Sedentary Behavior Guideline Committee recognized heterogeneity across studies in the measurement of bout lengths and breaks in sedentary time as an "important area of future research" and a reason for the lack of evidence regarding how bouts and breaks in sedentary behavior are related to health outcomes (Dempsey et al., 2020;Katzmarzyk et al., 2019). CHAP-Adult was intentionally designed to fill these gaps in two ways: first by being trained and validated among community-living adults from two continents across a wide age range (35-99 years); and second by being designed for application to a very common accelerometer and wear protocol (hip-worn ActiGraph GT3X+) for free-living physical behavior measurement (Wijndaele et al., 2015). With CHAP-Adult, many large epidemiologic cohorts that use either hip-worn GT3X+ or thigh-worn activPAL accelerometers can produce estimates of sitting patterns that are in high agreement. This broad applicability will enable consistent and accurate estimates that will help the field of sedentary behavior epidemiology generate norms for sitting patterns. It will also facilitate generation of estimates of the health risks associated with sitting patterns from various, heterogenous cohorts around the world-estimates that can more easily be harmonized for metaanalytic purposes since they would be generated using the same pretrained model (CHAP-Adult). Finally, it is our hope that this widespread availability of accurate data on sitting patterns will spur rigorous investigation of how sitting patterns and total sitting volume are jointly and independently associated with health. Ultimately, the new data generated using CHAP-Adult can therefore help the field progress toward specific quantitative (time based) recommendations that include a threshold for limiting sitting time and recommendations regarding how sitting can be broken up throughout the day.

CHAP-Adult was trained using 30-Hz data from the triaxial ActiGraph GT3X+ as inputs and activPAL-classified postures as the ground truth, both set up strengths and limitations. Use of triaxial data preclude the application of this pretrained algorithm to data collected using uniaxial accelerometers, limiting applicability, although new deep learning algorithms could be developed for uniaxial accelerometer data streams, provided a large volume of raw acceleration data as well as data from a concurrent ground truth assessment were also available. We also are unsure how well CHAP-Adult will work with triaxial data collected using other devices that collect raw acceleration data (e.g., GENEActiv and Axivity), but we suspect it would perform well as long as a similar wear protocol was followed. CHAP-Adult was not designed for use with data from wrist-worn accelerometers; new algorithms are needed for wrist-based devices. While the activPAL is highly accurate for measuring sitting and sit-stand transitions compared with direct observation (Lyden et al., 2012), it is not without error. Thus, we have shown convergent and concurrent validity, but future work should also look at criterion validity. However, the error in using this ambulatory monitor must be viewed against the benefits of including a large and diverse sample from the United States and Australia, and measuring physical behavior during free-living environments for an average of 7 days. This greatly enhances the utility and therefore the potential impact of CHAP-Adult. Other limitations are that we could not assess specificity and negative predictive value for transitions because the class is highly imbalanced and the validation metrics would be artificially inflated. This algorithm was also developed using data on adults ≥35 years of age, and future similar studies among children and younger adults are warranted. Analyses did not consider clustering or stratification from the multiple cohorts or within cohorts (e.g., AusDiab stratified multistage sampling), however, since this validation is concerned with pairwise comparisons, the impact on the validation metrics is likely negligible.

The CHAP-Adult is a pretrained deep learning algorithm that leverages CNN and BiLSTM architecture to classify sitting and postural transitions in adults over a wide age range. Our validation results demonstrate that CHAP-Adult produces estimates of sitting time and sitting patterns that are highly convergent with the activ-PAL, and does so within subgroups based on age, gender, and BMI. Additional external validation is needed to increase confidence in these findings. Ultimately, we will measure the success of CHAP-Adult by how often it is applied to data around the world. Application to data collected in many established and emerging national and international cohort studies will more strongly position the field of sedentary behavior epidemiology to establish scientifically sound population and subgroup sitting norms and determine how sitting and sitting patterns may be causally related to adverse health outcomes. Better identification of sitting patterns will also enable more accurate classification of the types of activities used to break up sitting (Blankenship et al., 2021), which is an emerging area of public health inquiry. Ultimately, widespread use of CHAP-Adult could generate data that contribute to national and international quantitative (as distinct from current broadly qualitative) recommendations for sitting and sitting patterns.

Open-sourced code and detailed vignettes for implementing CHAP-Adult are available at the following website: https:// adalabucsd.github.io/DeepPostures/.