Validity of ActivPAL CREA software detection of sitting and lying during free-living conditions

Objective. Approaches to differentiate sitting and lying are available within the default activPAL software from a single thigh-worn monitor. Dual-monitor methods use multiple monitors positioned on the thigh and torso to characterize sitting versus lying. We evaluated the validity between these two methods to measure waking, sitting, and lying time in free-living conditions. We also examined if the degree-threshold distinguishing sitting/lying for the dual-monitor (<30° and <45°) impacted results. Approach. Thirty-five young adults (24 ± 3 years, 16 females) wore an activPAL 24 h per day on their thigh and torso during free-living conditions (average: 6.8 ± 1.0 d, 239 total). Data were processed using the default activPAL software (thigh-only) or a custom MATLAB program (dual-monitor). Main results. The single-monitor recorded less lying time (59 ± 99 min d−1) and more sitting time (514 ± 203 min d−1) than the dual-monitor method regardless of 30° (lying: 85 ± 94 min d−1; sitting: 488 ± 166 min d−1) or 45° lying threshold (lying: 170 ± 142, sitting: 403 ± 164 min d−1; all, p < 0.001). The single monitor lying time was weakly correlated to the dual-monitor (30°: ρ = 0.25, 45°: ρ = 0.21; both, p < 0.001), whereas sitting was moderate-strong (30°: ρ = 0.76, 45°: ρ = 0.58; both, p < 0.001). However, the mean absolute error was 81 min d−1 (30°) and 132 min d−1 (45°) for both lying and sitting. Significance. The method of differentiating sitting/lying from a single thigh-worn activPAL records more sitting time and less lying time compared to a dual-monitor configuration (regardless of degree-threshold) that considered the position of the torso. A further refinement of algorithms or implementation of multiple-monitor methods may be needed for researchers to derive detailed sedentary positions.


Introduction
Sedentary time is included in (inter)national guidelines (Bull et al 2020, Ross et al 2020 and is defined as any waking time spent expending low energy in a seated, reclining, or lying posture (Tremblay et al 2017). Sedentary time may be objectively measured using thigh-worn inclinometry that distinguishes sedentary postures from upright time. Despite the term, 'sedentary' encompassing multiple postures, sitting and lying might exhibit unique cardiovascular (Atuk et al 1959, Lu et al 2008, Vranish et al 2018 and musculoskeletal effects (Haynes and Williams 2008). In a laboratory setting, lower-limb vascular function is lower in a sitting versus lying posture (Vranish et al 2018) and arterial blood pressure is elevated in sitting versus the lying position (Lu et al 2008). Further, post-myocardial infarction, sitting (or Armchair treatment) has been demonstrated to yield better clinical and psychological outcomes than strict bedrest (Levine andLown 1952, Atuk et al 1959). Our Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI. understanding of the health impacts of specific sedentary positions is largely limited by the lack of established threshold to characterize the sitting versus lying posture and measurement techniques and algorithms available to distinguish sitting and lying.
The activPAL is a valid measure of physical activity  and postures . Specifically, a single thigh-worn activPAL has previously been validated to accurately quantify stationary time (i.e. sedentary and standing time) (Grant et al 2006, and stepping (Johns et al 2020, Wu et al 2021 using the commonly used software. Edwardson et al (2017) has extensively demonstrated the prevalent use of the activPAL within field-based research for its ability to measure sitting/lying via posture (Edwardson et al 2017). However, differentiating between sedentary postures (i.e. sitting versus lying) using this software is yet to be explored. Using a single activPAL monitor, Lyden et al (2016) presented a method to derive lying from sitting, based on the premise that the thigh is more rotated during lying versus sitting. Their method characterized sedentary periods where the thigh was rotated (along the long axis) >65°or <−65°as bouts of lying. We (O'Brien et al 2023a) and others have implemented multiple activPAL (Taraldsen et al 2011, Bassett et al 2014, Smits et al 2018 and other objective monitor (Awais et al 2019) configurations whereby monitors are positioned on the thigh and torso to indicate if participants are sitting (horizontal thigh and vertical torso) or lying (horizontal thigh and horizontal torso). Awais et al (2019) found that a dual monitor approach was more accurate than a single monitor for identifying activities of daily living for older adults (Awais et al 2019). Additionally, the dual-monitor configuration accurately characterized sitting and lying compared to direct observation (Bassett et al 2014, Smits et al 2018. Smits et al (2018) demonstrated moderate agreement between single-and dual-monitor derived lying time, with the single monitor reporting less lying time than the dualmonitor. Accordingly, this lone study of healthy adults over a 24 h period (Smits et al 2018) supported that a single and multiple activPALs do not produce interchangeable outcomes.
Initially, the original default activPAL software, termed VANE, was used to derive time spent in sedentary postures, standing, and stepping (Buchan andUgbolue 2022, Montoye et al 2022). More recently, PAL Technologies implemented the CREA and GHLA (same as CREA with improved sensor calibration for AP4 and AP5 devices) software versions to provide more detailed movements (e.g. cycling) and sedentary posture information (e.g. sitting versus lying). The CREA and GHLA algorithms examine all periods of non-upright wear time and classify lying as periods longer than 1 h with a rotated thigh. Some breaks in lying are permitted (uninterrupted sitting or upright <15 min). The CREA and GHLA algorithms report equivalent sitting/lying time (Buchan and Ugbolue 2022). Given that the activPAL is a very common accelerometer to characterize posture (Edwardson et al 2017), and most users likely rely on the provided software to analyze their data, it is important to understand the accuracy of the outcomes. While prior work demonstrated that dual and single monitors exhibit moderate agreement (Smits et al 2018), the agreement between methods using the available algorithms that differ from the original study to denote lying is unclear. The method readily available to activPAL users has not been evaluated.
The purpose of this study was to examine the agreement between a single monitor determined sitting and lying time using the activPAL CREA algorithm with a multiple monitor configuration that can differentiate sitting versus lying. Based on previous literature (Smits et al 2018) and the newly implemented 1-h minimum duration thresholds implemented in the CREA algorithm, it was hypothesized that the single monitor with CREA will record less lying time during waking hours than the multiple monitor configuration. By recording less waking lying time at a fixed total sedentary time, it was hypothesized that the single monitor would record more sitting time. Since there is not a defined hip-angle to distinguish sitting from lying, we approached our research question implementing either <30°and <45°as thresholds for lying.

Methods Participants
Thirty-five healthy, young adults (<40 years, 16 non-pregnant females) consented to participate in the study. All participants were non-hypertensive (seated resting systolic blood pressure 139 and diastolic blood pressure 89 mmHg) via automated vital signs monitor (Carescape V100; General Electric Healthcare, Mississauga, ON, Canada). ActivPAL derived hip-angles have been presented previously (O'Brien et al 2023a), but the analyzes conducted answered an independent, novel research question. Height and weight were measured using a calibrated stadiometer/physician's scale (Health-O-Meter, McCook Il, USA) to the nearest 0.5 cm and 0.1 kg, respectively. These measures were then used to calculate body mass index (kg m −2 ). Prior to testing, verbal and written informed consent were acquired. All protocols and procedures conformed to the Declaration of Helsinki and were approved by the Dalhousie University Health Sciences Research Ethics Board. All participants provided written informed consent.
Habitual activity monitoring Participants were equipped with an activPAL inclinometer (AP3 and AP4, PAL Technologies Ltd., Glasgow, UK) positioned on their torso and thigh that were waterproofed via a nitrile finger cot and secured using Tegaderm TM medical dressing (3M, London, ON, Canada) (figure 1). Based on recommended guidelines (Edwardson et al 2017), the thigh monitor was positioned on the right anterior thigh, one third of the way between the hip and the knee. The torso monitor was positioned on the right side of the torso just below the ribcage, parallel to the thigh monitor (O'Brien et al 2023a). Participants wore the activPAL monitors 24 h per day for a minimum of 5 d including a Saturday/Sunday day (mean ± SD: 6.8 ± 1.0 d). Participants self-reported their waking hours via a diary of wake-up and sleep times (including naps) to distinguish bouts of sleep from sedentary time.
For the dual-monitor method, the activPALs were synchronized to start recording simultaneously at midnight of the first day of monitoring. A detailed description of the analysis program is presented elsewhere (O'Brien et al 2023a). Acceleration data were sampled at 20 Hz, downloaded, and processed (PALanalysis V7) to provide '.csv' files, which were used in a custom program to calculate hip angles. The acceleration data were lowpass filtered at a 0.18 Hz cut-off (i.e. 1st order zero-lag digital Butterworth filter) (Pickford et al 2019). The sedentary joint-angle positions algorithm was developed using MATLAB (R2022a, The MathWorks Inc. Natick, MA, USA). The analysis program is available from GitHub (O'Brien et al 2023a). Relative hip flexion angles between activPALs were determined during the sedentary bouts using the dot-product method between the torso-thigh. Definitions of what angle constitutes sitting versus lying are unclear. Accordingly, two hip angles of <30°and <45°were considered lying for this study and analyzes were conducted separately for each threshold.
The single monitor on the thigh was analyzed using the conventional PALanalysis software that exported time spent sitting and lying. While the thigh rotation used by the CREA software to denote lying is not indicated, thresholds of >65°or <−65°used in the Lyden et al (2016) study might be implemented into the software, but this was not confirmed. For the software algorithm to characterize the bout as lying, it must be >1 h in duration with a rotated thigh. Sitting (non-rotated thigh) or upright bouts >15 min are needed to demark the end of the lying period. Any sedentary time that does not meet the duration and thigh rotation criteria, are characterized as sitting using this algorithm.

Statistical analysis
Statistical analyzes were conducted following standardized reporting guidelines for physical activity monitor validation studies (Welk et al 2019). All variables were assessed for normality using the Shapiro-Wilk test and determined to be non-normal (all, p < 0.05). To characterize the average time spent in postures derived from the dual-monitors corresponding to <30°(i.e. laying), 30°−45°, and >45°(i.e. sitting) a one-way ANOVA with Bonferroni post-hoc was conducted between the three thresholds. Wilcoxon-signed rank tests compared time spent sitting and lying between the single-monitor versus both the 30°or 45°dual-monitor hip angle thresholds. Mean absolute error was calculated by averaging the absolute difference (single-dual monitor) for all derived measures. Mean absolute percent error was not calculated due to frequent instances of the single monitor recording 0 min d −1 of lying time (61% of days).
Spearman's Rank correlational analyzes were conducted between measures. Bland-Altman analyzes were used to identify whether fixed and/or proportional biases existed. If the 95% confidence interval of the mean difference included '0', fixed bias was absent. If the 95% confidence interval for the slope of the mean differences included '0', proportional bias was absent.
Rather than determine whether measures were statistically different, equivalence testing was used to assess the level of similarity (Dixon et al 2018). Instead of selecting an arbitrary equivalence zone, we determined the minimum percentage required for the measures to be equivalent, as described previously (O'Brien 2021). Generally, a threshold of ±10% is used in studies comparing device-based measures, but this threshold may vary between studies as there have not been any standardized guidelines published (O'Brien 2021). Equivalence testing identified when the range of the dual-monitor-derived outcome included the 90% confidence interval (i.e. 100%-2α[5%]; for upper and lower confidence intervals) of the single monitor outcome. Statistics were completed in SPSS (Version 28.0. IBM Corp., Armonk, NY). Statistical significance was accepted as α < 0.05. Data are presented as means ± standard deviations.
Lying time recorded by the single monitor (59 ± 99 min d −1 ) was less than the dual-monitor method when either the 30°(85 ± 94 min d −1 ) or 45°hip-angle threshold was used (170 ± 142 min d −1 ; both, p < 0.001). The mean absolute error was 81 min d −1 (30°) and 132 min d −1 (45°). The single monitor lying time was weakly correlated to the dual-monitor ( figure 2). Similarly, when expressed as a percentage, large equivalence zones were required for the measures to be statistically equivalent (30°: ± 43%, 45°: ± 72%), indicating low equivalence. As demonstrated in figure 3, a negative fixed bias was observed in the Bland-Altman analyzes showing that the single monitor records less lying time than the dual-monitor, but a negative proportional bias was observed for the 45°only (β = −0.56, p < 0.001) whereby underprediction increased as a function of the  Given that the single and dual-monitors recorded the same total sedentary time (573 ± 192 min d −1 ), all non-lying time were considered sitting time, which was larger in the single monitor (514 ± 203 min d −1 ) versus the dual-monitor 30°(488 ± 166 min d −1 ) and 45°hip-angle thresholds (403 ± 164 min d −1 ; both, p < 0.001). Despite differences in absolute time, the single monitor sitting time was moderate-strongly correlated to the dual-monitor (figure 4). The mean absolute error was 81 min d −1 (dual-monitor: 30°) and 132 min d −1 (dualmonitor: 45°). Based on the group-level data, the 30°threshold exhibited a high level of equivalence with the single-monitor ( ±10%), but the 45°threshold did not ( ±33%). As demonstrated in figure 5, positive fixed biases and positive proportional biases (all, p < 0.001) were observed, with the single monitor resulting in more sitting time than the dual-monitor that increased as a function of the average.

Discussion
We compared the current activPAL software-determined lying and sitting time from a single thigh-worn monitor versus a validated dual-monitor method that positions an additional monitor on the torso. Consistent with our hypothesis, the single monitor using the CREA software recorded less lying time and thus, more sitting time than the dual-monitor method. This difference was even larger if the threshold used to denote lying from the dual-monitor was increased from 30°to 45°. Despite the integration of deriving lying from sitting in the  available activPAL software used by researchers to process their data, the algorithm implemented underestimates lying time and overestimates sitting relative to a dual-monitor approach. Importantly, compared to using the dual-monitor approach, researchers who uses the CREA algorithm will be subject to a mean absolute error of ∼81 min d −1 (30°) and 132 min d −1 (45°). The discrepancy in measuring the duration that participants are spending in these postures could incorrectly lead to type I or II errors regarding the relationship of sitting/lying with physiological, biomechanical, and/or health outcomes.
Utilizing multiple activity monitors to derive detailed body positions has been implemented in previous studies to better characterize specific sedentary positions (Taraldsen et al 2011, Norvang et al 2018, Smits et al 2018, with existing laboratory-based validation work demonstrating that dual-monitors improve the distinction of sitting/lying compared to direct observations (Bassett et al 2014, Smits et al 2018. While the position of the thigh may be more externally rotated during lying versus sitting (Lyden et al 2016), our study calls into question the full-proofness of this method to quantify lying, as sitting may be conducted with the thigh in a more externally rotated thigh position or lying conducted with the thigh in a more internally rotated position. Plausibly, there are instances that a rotation-based criteria may misclassify activity, and this is likely exacerbated when additional criteria are included by the CREA software. Specifically, the requirement of 1 h bouts to be considered lying has limitations and does not include shorter lying sessions (e.g. lying to watch a television show, read, meditate, etc). Such strict duration-based and rotation-based criteria explain the frequent instances where there were 0 min d −1 of lying time recorded by the single monitor algorithm (146/239 d), whereas the dualmonitor detected much fewer days (30°: 6; 45°: 1). Conversely, there were some, albeit relatively fewer instances of the single monitor estimating more lying time than the dual-monitor method (figure 3). This is likely due to individuals sitting with their thigh externally rotated for a duration long enough to constitute lying by the algorithm. Altogether, caution is warranted for researchers interested in quantifying awake lying time using a single thigh-worn monitor as it records a lower duration and is only weakly correlated to the dual-monitor approach.
Research on sedentary time has grown exponentially over the past 3 decades and is now included in international activity guidelines (Bull et al 2020). Despite both being classified as sedentary, sitting and lying are unique postures. While thigh-worn monitors provide measures of sedentary time due to their distinction of true sedentary bouts from quiet standing (O'Brien et al 2022), some researchers may be interested in understanding their participants' sitting time specifically (Norvang et al 2018). A moderate-strong correlation was observed between the single monitor and dual-monitor sitting time, with a slight overprediction of the single monitor at the group level ( figure 4). However, absolute mean error reflects the individual level error and demonstrated that the single and dual-monitor methods differed in sitting time, challenging the agreement between methods on an individual level. This level of error may pose issues for researchers who are trying to understand the impact of specific sedentary postures on outcomes (e.g. physiological, mortality, mobility, etc) and/or developing more detailed sedentary guidelines. A further refinement of algorithms may be needed to improve the detection of sitting time specifically, or other strategies such as using dual monitors, may be needed to measure detailed sedentary postures.
Integrating detailed sedentary postures into software is helpful for activity researchers. Given that the software is readily available for those using the activPAL, it is plausible that many researchers will present sitting and lying in forthcoming studies. However, the current evidence is preliminary for trying to ascertain these metrics based on a single thigh monitor. More development and validity studies are required to demonstrate that the existing algorithm can accurately detect these postures. As evident by their software updates, there is an interest in better understanding sedentary postures, but perhaps the ability to provide detailed postures is not possible by only knowing the orientation of the thigh. Rather, providing analysis options for researchers who wish to implement dual-monitor (Taraldsen et al 2011, Bassett et al 2014, Norvang et al 2018, Smits et al 2018 or tri-monitor (O'Brien et al 2023a) configurations would negate the need to spend resources to program their own in-house analysis programs. However, it is appreciated that the software provides raw accelerometer outputs as it provides an avenue for researchers independently explore ways of expanding the capabilities of the monitors. In agreement with others, an open sharing of methods used and internal testing processes by activity monitor companies would be helpful and reduce the 'blackbox' nature of activity monitors ( We acknowledge that our study is limited to the population of interest and that our findings may not apply to hospitalized or bedridden participants, who engage in more waking lying time than the young adults investigated. Similarly, the current participant sample is young (24 ± 3 years) with a relatively normal body mass index (24.4 ± 3.4 kg m −2 ), which may not be representative of the general population. It is unclear whether age and body mass influence time spent in detailed sedentary postures, with evidence that older adults spend more time sedentary in general (Harvey et al 2013). Given the criteria that needs to be met for the CREA algorithm to detect lying, it is likely that similar results would be observed in older or obese populations. Future studies should consider these populations. While the theoretical basis for the dual-monitor configuration is strong, in that knowing the position of the torso is important in distinguishing sitting versus lying, video-recorded body positions would be a superior criterion in free-living conditions. Studies using gold-standard criterion measurements would be worthwhile, but methods to gather and process multiple days of video-recorded data are time-intensive (Johnston et al 2021), require high participant burden, and lack clear processes demarcations of sitting/lying. The CREA software may pose useful for identifying bedtime, but our study was interested in sedentary time specifically, which is accumulated while awake. Our study is strengthened by its rigorous statistical evaluation of validity, inclusion of free-living assessed sedentary positions, and testing the available activPAL software that sedentary researchers currently have available to them. Using the validated dual-monitor approach, future studies should investigate the relationships between sitting versus laying postures on various outcomes. This will be a particularly important contribution towards creating detailed sedentary guidelines for adults and informing clinical population (e.g. bed rest patients).
In conclusion, the single-monitor outcomes analyzed using the CREA algorithm and a dual-monitor configuration do not produce interchangeable time spent lying or sitting, with the single monitor recording less lying time and more sitting time. This observation remained regardless of the dual-monitor threshold implemented (30°or 45°). Our study documented the challenges with the currently available activPAL algorithm to derive sitting versus lying, as evident by the frequent instances of 0 min d −1 of any waking time spent in lying postures. We propose that the orientation of the torso is likely needed to truly distinguish sitting/lying and that caution is advised if using available methods to measure waking lying time.

Data availability statement
All data files are available from the corresponding author upon reasonable request. Authors must agree to make any data required to support or replicate claims made in an article available privately to the journal's editors, reviewers and readers without undue restriction or delay if requested.