Assessment of children’s sleep using thigh-worn Axivity AX3 accelerometers

Assessment of children’s sleep using thigh-worn Axivity AX3 accelerometers | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Assessment of children’s sleep using thigh-worn Axivity AX3 accelerometers Maja Sulstad Johansen, Esben Høegholm Lykke, Anders Grøntved, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7935022/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 24 Apr, 2026 Read the published version in Sleep Science and Practice → Version 1 posted 12 You are reading this latest preprint version Abstract Background Accurate assessment of sleep is vital, but the gold standard, polysomnography, is costly and impractical for large-scale studies. An alternative is wearable accelerometers, which reduce participant burden and eliminate potential recall biases. This study aimed to develop and validate a method for estimating time in bed (TIB), total sleep time (TST), sleep efficiency (SE), sleep onset latency (SOL), and wake after sleep onset (WASO) utilizing machine learning applied to accelerometry data. Methods Data on 309 nights from 134 children aged 4–17 years was used to develop a method utilizing two machine learning models applied to data from thigh-worn accelerometers to estimate sleep metrics. Inputs were collected simultaneously from the Zmachine Insight + and raw data from thigh-worn accelerometers, and validated using k-fold cross-validation. The method was then cross-validated against polysomnography in an independent sample of 136 children aged 8–16 years. Results The independent validation showed overestimations of 28.0 minutes for bedtime and 11.2 minutes for wake time, with ICC of 0.59 and 0.55. TIB and TST were overestimated by 13.8 and 3.4 minutes with ICC of 0.59 and 0.56, respectively. The correlation for estimating SE, SOL and WASO was weak with ICC of 0.21, 0.01 and 0.04, respectively. Conclusions This method demonstrated sufficient accuracy for assessing bedtime, wake time, TIB and TST at the group level when validated in an independent sample against polysomnography, although wide limits of agreement suggest limited precision for individual-level assessments. Low agreement for SE, SOL and WASO indicated insufficient accuracy of the method for these metrics. accelerometry actigraphy validation polysomnography sleep assessment sleep/wake cycles children adolescents machine learning wearable sensors Figures Figure 1 Figure 2 Figure 3 Background It is well-recognized that sleep, including duration, quality and timing, plays a critical role in physical and mental health and development of children and adolescents (1, 2). Consequently, accurate assessment of sleep is crucial for tracking sleep patterns and improving our understanding of the sleep-health relationship, with methods required to facilitate assessment in large-scale, longitudinal studies. The gold standard for sleep assessment, laboratory-based polysomnography (PSG), is impractical in large-scale observational and experimental studies due to the high cost, need for professional administration, participant burden and susceptibility to rater bias (3, 4). Alternatively, diaries and questionnaires have been used due to their cost-effectiveness and simplicity, although they are subject to recall bias (5). The Zmachine®ฏ Insight+ (ZM), is a wearable EEG-based sleep monitoring device validated against PSG (6, 7), offering an objective assessment of sleep, but can be implemented in a home-based environment without the need for professional monitoring. Crucially, the ZM facilitates multi-night analysis in free-living conditions due to its unobstructive and ease of use (8), capturing natural variation in sleep patterns. This makes it advantageous over single-night PSG, particularly as a data source in machine learning (ML), as it provides multiple nights of measurements without inter-rater bias. Despite these benefits, the ZM still poses a significant participant burden and cost. An innovative approach involves device-based measurement. These tools, which estimate several sleep metrics, are advantageous due to their reduced participant burden and elimination of potential recall bias. Prominent examples of such tools are body-worn accelerometers, which offer a practical and affordable means of assessing sleep patterns at home for extended periods. These devices also enable the capture of natural day-to-day variations in sleep that may be oversimplified by subjective methods. Their use in sleep and wake classification began with a wrist movement-based algorithm developed by Webster et al. in 1982 (9). This algorithm was refined in 1992 by Cole et al. (10), leading to the widely adopted Cole-Kripke model. With advancements in the field, a variety of techniques, including heuristic algorithms, ML models, regression, and deep learning, are now used to analyze data from hip and wrist-worn accelerometers (10–15). While wrist and hip-worn devices have benefited from extensive methodological development, thigh-worn accelerometers have not seen the same level of advancement. Existing studies mainly focus on distinguishing sleep from wakefulness, with emphasis on defining ‘waking time’ and ‘bedtime’ (16–19). While recent advances in estimating sleep duration using thigh-worn devices include the promising algorithm by Johansson et al. (20), the application of ML in this area is still underexplored. Considering the potential of thigh-worn accelerometers for accurate physical behavior assessment and identification of postural transitions such as standing and sitting (21–23), there is a significant research gap. Thus, this study aimed to develop a method using two machine learning models to estimate bedtime, wake time, and sleep metrics, utilizing raw data collected from thigh-worn Axivity AX3, a tri-axial accelerometer, and validate the method against PSG in an external dataset. Methods Dataset and Participants This study used data from the SCREENS trial (24, 25), conducted from June 2019 to March 2021 in the Region of Southern Denmark, that evaluated the effect of limiting screen media usage within families. For our analysis, we focused on data from child participants aged 4 to 17 years. Our primary sources of data were accelerometer readings from Axivity AX3 devices (Axivity Ltd, Newcastle, United Kingdom), and EEG-derived sleep states and sleep metrics from the ZM device, from three consecutive days during baseline and follow-up. The Axivity AX3, an unobtrusive 3-axis accelerometer, was attached midway between the hip and knee on the right anterior thigh using elastic belts, and worn 24-hours a day for the full three days. Sleep state information was extracted using the ZM, a product of General Sleep Corporation. The ZM, which utilizes EEG hardware and signal processing algorithms, employs three self-adhesive, disposable sensors placed outside the hairline for EEG signal acquisition. The participants of the SCREENS trial were instructed to attach the ZM device when they went to bed and remove the device upon getting out of bed. The ZM uses two proprietary algorithms: Z-ALG and Z-PLUS. The Z-ALG is utilized for sleep detection, showcasing its suitability for in-home monitoring (6), while the Z-PLUS differentiates sleep stages, as evidenced by its alignment with expert evaluations using PSG data (7). In the current study, we treated all sleep stages (light sleep (N1 & N2), deep sleep (N3), and REM sleep) as a single category effectively deducing the output of the ZM to “awake” and “asleep”. Previous research has not been successful in accurately distinguishing between sleep stages when using accelerometry, and this approach was chosen because such differentiation is not necessary to derive the sleep metrics of interest and helps simplify the learning process of the ML algorithms. We chose to incorporate a 10-minute median-filtering of the raw predictions from the ZM device into our modeling process. Children typically experience around five to eight sleep cycles per night, with awakenings most likely at the end of each cycle (26). In examining the raw ZM predictions, we observed a notable overestimation in the number of awakenings for the children in our study, surpassing expected counts based on typical sleep cycle patterns illustrated in Fig. 1. The average sleep efficiency determined by the raw ZM predictions for our sample was 83%, which is slightly below the threshold of ≥85% indicating good sleep quality (1). In fact, prior research has indicated sleep efficiencies of over 90% in similar child cohorts (27, 28). This discrepancy suggests that the raw ZM predictions might be overestimating awake periods. Many of these brief awakenings could be considered as noise, which when present in the data, can potentially hinder the learning process of ML algorithms by obscuring the underlying patterns that the algorithms are trying to learn, leading to less accurate predictions. Figure caption 1: Impact of 5- and 10-minute median filtering on raw ZM awakening predictions. (Fig. 1) Figure legend 1: The difference in number of awakenings between the raw ZM predictions vs. 5-minute, and 10-minute median filtered predictions for a random night (boy, 9 years). Grey line is the raw predictions; black line is the median filtered predictions. A: 5-minute median filter on raw ZM predictions, B: 10-minute median filter on raw ZM predictions. Only ZM recordings accompanied by complete accelerometer data and lasting between 7 and 14 hours were considered. These cutoffs were set to exclude very short or long periods of estimated sleep, as several children did not adhere to the protocol e.g. removing the ZM during the night or failing to turn off after removal. All nights were manually curated and nights during which the ZM reported sensor issues were excluded. Data Preprocessing and Feature Extraction Processing of the raw accelerometer data began with a low-pass filtration step using a 4th order Butterworth filter with a 5 Hz cut-off frequency to eliminate high-frequency noise similar to methods described by Skotte et al. (21). Any non-wear data was identified using previously described methods (29) and nights with non-wear were excluded in the development and validation of the method. The acceleration was processed in two second windows with one second overlap, providing a resolution of one second. For each window, the mean, standard deviation (SD), inclination, angle and turn angles were calculated. The data were then aggregated using a mean into 30-second epochs, so every sample corresponds to a 30-second epoch scored during the ZM recordings. For each of the inclination, angle and turn features a 30 second window SD was calculated and added. A minimum 5 step walking bout circadian rhythm feature (WalkCirc) was generated by using a wavelet step detection (30, 31) and subsequently enveloping from the last step bout at 11 PM and the first step bout after 7 AM. The feature defines the primary period of the day for which the subject is moving around (the first bout of walking until the last bout of walking). The feature is binary and pre-defines what is day and what is potentially night, thereby directing the algorithm on identifying the circadian rhythm. Together with a categorical weekday feature (weekday or weekend) the complete data set contains thirteen unique features describing the movement behavior and is presented in Table 2 . Model training - Classification of sleep metrics The sleep metrics included in this study are defined as follows: Bedtime – The start of the ZM recording, representing the timepoint when sleep is initiated. Wake time – The ending of the ZM recording, representing the end of sleep. Time In Bed (TIB) - The total duration of time in bed, which is defined as the time from the start to the end of the ZM recording. Sleep Onset Latency (SOL) - The time it takes to transition from wakefulness to sustained sleep. It is calculated as the time from the beginning of the ZM recording until the first period when 10 out of 12 minutes are scored as sleep. Wake After Sleep Onset (WASO) – Time spent awake after initially falling asleep and before the final awakening. In our analysis, a period is counted as ‘awake’ only if it consists of 3 or more contiguous 30-second epochs, which is also how the ZM summarizes WASO. Total Sleep Time (TST) - The time spent asleep within the TIB, without sleep onset latency and wake after sleep onset. Sleep Efficiency (SE) - The ratio between TST and TIB, representing the proportion of the time in bed that was actually spent asleep. The classification of TIB, and sleep/wake from acceleration measured at the thigh is a simple binary classification. We chose a simple classical decision tree for both classification tasks due to its transparency and interpretability – particularly relevant in this context, as we wanted to trace how specific input features influence the output and avoid potential overfitting and the black-box issues often associated with highly flexible models like gradient boosting or neural networks. We employed a two-model strategy to assess TIB/time out of bed (TOB) and sleep/wake from the thigh-mounted accelerometer data. The algorithm approach is outlined in Fig. 2. The approach was used to simplify the prediction task by decomposing the multiclass problem of classifying out-of-bed-awake, in-bed-awake, and in-bed-asleep into two binary stages: first predicting TIB, then using the suggested TIB data to estimate sleep and wake time while in bed. Figure caption 2: The two-model machine learning method. (Fig. 2) Figure legend 2: Graphical overview of the two-model machine learning method to classify bedtime, wake time, time in bed, sleep onset latency, wake after sleep onset, total sleep time, and sleep efficiency. Independent validation There are numerous options for validating the generalizability of a supervised ML model and we utilized k- fold cross-validation during the development. However, a more realistic estimate of the external validity involves using an independent dataset collected in a different setting in children, using a similar device and placement, along with a reference measurement with sufficient accuracy like PSG or EEG. The independent validation was performed using data from a study conducted in New Zealand (NZ), in which simultaneous thigh-worn accelerometry (Axivity AX3) and PSG (Embletta MPRPG, ST + Proxy and TX Proxy, Natus, California, USA) data over one night from healthy children and adolescents aged 8–16 years. Detailed procedures for the overnight PSG assessment, including device specifications, signal processing methods, and scoring of sleep stages according to American Academy of Sleep Medicine guidelines, have been described elsewhere (32). Participants in the NZ study were recruited via social media (i.e. Facebook), schools, and word of mouth. Children did not have any history of sleep disturbance (32). Statistical All raw accelerometery processing and model development were done in Matlab (Version 24.1.0 R2024a, Mathworks Inc., Natick, Massachusetts, US) and performance evaluation using R version 4.3.0 (2023-04-21) (R Core Team 2023) and the blandr package. To assess performance of each model on an epoch-to-epoch basis the following standard evaluation metrics are included: $$\:accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$ $$\:sensitivity=\frac{TP}{TP+FN}$$ $$\:specificity=\frac{TN}{TN+FP}$$ $$\:precision=\frac{TP}{TP+FP}$$ $$\:NPV=\frac{TN}{TN+FN}$$ $$\:{F}_{1}=2\cdot\:\frac{precision\cdot\:sensitivity}{precision+sensitivity}$$ where \(\:NPV\) is negative predictive value, \(\:{F}_{1}\) is the F1 score, \(\:TP\) is true positives, \(\:FP\) is false positives, \(\:TN\) is true negatives, and \(\:FN\) is false negatives. TIB and sleep are defined as positive labels and TOB and wake as the negative labels in accordance with previous research (33, 34). To assess the performance of our models in deriving sleep metrics, we utilized Bland-Altman plots, Pearsons’s correlation ( r ) and Intra Class correlation (ICC). The Bland-Altman method was employed specifically to determine the level of agreement between the ZM and accelerometry derived predictions in the development and the agreement between the PSG and accelerometry predictions in the independent validation. We first calculated the mean difference with 95% CI between the measurement methods and then defined the limits of agreement (LOA) as the mean difference plus or minus 1.96 times the SD of these differences. Results Based on the curation of 906 nights, 597 nights were excluded for failing to meet the cut-off criteria, and 13 children were completely excluded due to no eligible nights, leaving 309 nights from 134 children with a mean of 2.3 nights per child (SD 0.98) eligible for model development. The ZM predictions encompassed 369,708 epochs, each 30 seconds long. Table 1 presents the characteristics of the children in the development dataset (SCREENS), and children in the validation dataset (NZ). The development group had a mean age of 9.51 (SD 2.17) compared to a mean age of 11.66 (SD 2.25) in the independent validation group. The Danish sample was also leaner than the New Zealand sample, with 18.7% overweight compared with 36.0% respectively. Table 1 Participants characteristics. SCREENS Development data NZ Validation data Sex, n (%) Boys 55 (41.04%) 69 (50.74%) Girls 79 (58.96%) 67 (49.26%) Age (years) 9.51 (2.17) 11.66 (2.25) BMI-z*, n (%) Thinness for age 5 (3.73%) 2 (1.47%) Normal weight for age 104 (77.61%) 85 (62.50%) Overweight for age 25 (18.66%) 49 (36.03%) Nights 309 136 Day, n (%) Monday 83 (26.86%) 38 (27.94%) Tuesday 102 (33.01%) 23 (16.91%) Wednesday 79 (25.57%) 17 (12.50%) Thursday 54 (17.48%) 35 (25.74%) Friday 36 (11.65%) 13 (9.56%) Saturday 52 (16.83%) 4 (2.94%) Sunday 58 (18.77%) 6 (4.41%) Characteristics of children participating in the SCREENS and NZ study. Age is presented as mean (SD). *Based on WHOs interpretation of BMI z-score cut-offs. Model development Permutation feature importance A total of 13 features were used in the development of ML I (TIB classification) whereas only 11 were used in ML II (sleep classification). ML II was applied only to data classified as TIB, which is why the features Steps and WalkCirc were not relevant in this context and were therefore excluded from this model. The importance of the individual features used in both models is presented in Table 2 . For TIB classification, the WalkCirc feature was the most important with temperature, inclination angle and SD of the inclination having minor contributions. The most important features in the sleep classification were the SD of the acceleration angle and maximum SD of the acceleration with minor contribution from SD of the acceleration, temperature and SD of the acceleration turn angle. Table 2 Permutation feature importance. Feature ML I Time In Bed ML II Sleep/wake Tempure 0.4 2.1 Macc 0.1 0.0 SDacc 0.2 3.7 SDmax 0.1 15.3 Inclination 0.4 0.5 Angle 0.1 1.0 Turn 0.0 0.0 InclSD 0.5 0.4 AnglSD 0.0 73.9 TurnSD 0.0 2.4 Steps 0.0 - Weekday 0.2 0.6 WalkCirc 98.1 - Overview of the features included in the ML I for classifying time in bed and the ML II classifying sleep/wake, along with the percentage importance of each individual feature. Macc represents the mean acceleration, SDmax denotes the maximal standard deviation in acceleration in all three axes, while inclSD, anglSD and turnSD refer to the standard deviations of the inclination, angle, and turn, respectively. Epoch level performance The performance of the epoch level classification of TIB and sleep with the model development is presented in Table 3 . All epoch level classification performance metrics for the identification of TIB were close to perfect and suggests that even though the model was simple it was possible to identify the time when children were in bed. By contrast, while accuracy and sensitivity were also high for detecting sleep, the ability to detect wake was low as indicated by a specificity of 0.45. Table 3 Epoch-by-epoch performance on development data. Accuracy Sensitivity Specificity F1 Precision ML I (time in bed) 0.95 0.93 0.99 0.96 0.99 ML II (sleep/wake) 0.93 0.99 0.45 0.59 0.85 Model accuracy, sensitivity, specificity, F1 and precision for the epoch level classification of ML I (classification of time in bed) and ML II (classification of sleep/wake while in bed). Sleep outcomes on development Mean difference and estimated LOA for the predicted TIB classification and sleep outcomes are presented in Table 4 . For the bedtime and wake time mean difference was close to zero and upper and lower LOA were less than 3.8 minutes. For TIB mean difference was < 9 minutes and the lower and upper LOA was − 67.8 and 85.7 minutes respectively. From the combined outputs of ML I and ML II, we derived TST, SE, SOL, and WASO. Mean difference for TST was zero with lower and upper LOA of -15.3 and 13.4 minutes respectively. The r suggested a strong correlation between the classification of bedtime, wake time, TIB and TST. Table 4 Agreement on development data. Mean difference LOA+ LOA- r Bedtime (min) -0.3 -3.8 3.2 0.79 Wake time (min) 0.6 -1.3 2.6 0.86 TIB (min) 8.9 -67.8 85.7 0.72 TST (min) 0.0 -15.3 13.4 0.99 SE (%) -1.5 -11.9 9.03 0.69 SOL (min) 1.7 -51.0 54.4 0.49 WASO (min) 4.7 -16.1 25.4 0.94 Agreement between the ZM measured and accelerometry predicted bedtime, wake time, time in bed, total sleep time, sleep onset latency, sleep efficiency and wake after sleep onset for the subjects included in the development dataset. Independent validation Mean difference, LOA, r and ICC of bedtime and wake time, TIB, TST, SE, WASO and SOL using the independent dataset is presented in Table 5 and with the Bland Altman plot of both TIB and TST presented in Fig. 3. There was a weak r and ICC for SE, SOL and WASO. The r for bedtime, wake time, TIB and TST were moderate to strong, while the ICC for these metrics were moderate. The mean difference for estimating bedtime and wake time was 28.0 minutes (95%CI 19.5, 36.5) and 11.2 minutes (95%CI 1.2, 21.3), respectively. The mean difference for TIB and TST was 14.0 minutes (95%CI 3.3, 24.8) and 3.3 minutes (95%CI -7.3, 13.9) when validated against PSG. Table 5 Agreement on validation data. Mean difference (CI) LOA- (CI) LOA+ (CI) r ICC p-value Lower Upper Bedtime (min) 28.0 (19.5, 36.5) -64.8 (-79.3, -50.3) 120.7 (106.2, 135.3) 0.66 0.59 0.00 0.32 0.74 Wake time (min) 11.2 (1.2, 21.3) -98.7 (-115.9, -81.5) 121.2 (104.0, 138.4) 0.57 0.55 0.00 0.42 0.66 TIB (min) 14.0 (3.3, 24.8) -103.6 (-122.1, -85.2) 131.7 (113.3, 150.2) 0.62 0.59 0.00 0.46 0.69 TST (min) 3.3 (-7.3, 13.9) -112.2 (-130.3, -94.1) 118.8 (100.7, 136.9) 0.57 0.56 0.00 0.43 0.67 SE (%) -1.6 (-2.6, -0.6) -12.8 (-14.5, -11.0) 9.5 (7.8, 11.3) 0.23 0.21 0.01 0.05 0.37 SOL (min) 14.2 (10.1, 18.3) -30.2 (-37.1, -23.2) 58.6 (51.6, 65.5) 0.08 0.01 0.04 -0.01 0.05 WASO (min) 18.0 (14.5, 21.6) -20.9 (-27.0, -14.8) 56.9 (50.8, 63.0) 0.11 0.04 0.29 -0.09 0.17 Agreement between the PSG measured and accelerometry predicted bedtime, wake time, time in bed, total sleep time, sleep onset latency, sleep efficiency and wake after sleep onset for the subjects included in the independent validation dataset. Figure caption 3: Agreement between PSG-measured and accelerometer-predicted time in bed and total sleep time. (Fig. 3) Figure legend 3: Bland Altman plots of the agreement between the PSG measured and accelerometry predicted time in bed and total sleep time including confidence intervals with mean difference (purple), upper limits of agreement (green) and lower limits of agreement (red) for the subjects included in the independent validation dataset. Discussion In this study, a method utilizing two ML models to predict TIB and sleep/wake was developed and validated based on features derived from acceleration and temperature data collected using thigh-worn Axivity AX3 devices. Based on the ML I model to predict TIB and the ML II to predict sleep/wake commonly derived sleep metrics including TST, SE, SOL and WASO were estimated and validated both during model development and using an independent dataset. Overall, thigh-worn estimates of bedtime, wake time, TIB and TST demonstrated moderate to strong validity compared to PSG reference estimates in the independent sample; however, the wide limits of agreement suggested considerable individual variability, and the method appeared more appropriate for analyzing group-level averages or for use in regression models where multiple days of data are available per participant, thereby reducing the impact of individual measurement error. The estimation of SE demonstrated low mean difference but poor correlation whereas for SOL and WASO both mean difference, LOA and correlation indicated that these outcomes were not valid. Carlson and colleagues evaluated a third-party algorithm, “ProcessingPAL,” and a proprietary one, “CREA,” which achieved epoch-by-epoch accuracies of 89.4% and 89.7%, respectively, when classifying TIB and TOB using thigh-worn accelerometry. These algorithms, evaluated against self-reported measures among adolescents and adults, produced F1 scores of 87.2% and 86.6% (16). These scores were slightly lower than the performance of the proposed ML I model, which achieved F1 score and accuracy exceeding 95% in identifying TIB. The validation revealed a mean difference of 14.0 minutes in estimating TIB with our ML I model, reflecting trends observed in previous research by Winkler et al. conducted in young- middle-aged and older adults (19). They developed an algorithm that, despite a moderate correlation ( r = 0.67) between their algorithmic results and diary-recorded waking times, overestimated waking time by more than 30 minutes, resulting in an underestimation of TIB (19). This trend was further confirmed when Inan-Eroglu et al. examined Winkler et al.’s algorithm, revealing an underestimation of TIB with 9.8 minutes compared to self-reported measures in middle-aged adults (17). Similarly, a study by van der Berg et al. reported a slight underestimation of TIB in a sample of middle-aged and older adults. They employed a unique approach with their algorithm, which relied on quantifying the number and duration of sedentary periods to determine TIB, and active periods (standing or stepping) to identify wake times (18). Important to note is that predictive performance in determining TIB does not necessarily translate to accurate predictions of broader sleep quality metrics. The crucial task of detecting wake periods during TIB, a key factor in assessing further derived sleep quality metrics, is not effectively captured by TIB predictions alone. The distinction between actual sleep and time spent in bed awake, is critical for a comprehensive understanding of sleep quality. Johansson and colleagues (20) presented in their study epoch-to-epoch performance metrics for sleep scoring using thigh-worn accelerometers in adults. They achieved a mean sensitivity of 0.84, specificity of 0.55, and accuracy of 0.80, using a single-night evaluation dataset of 71 adult subjects. Despite our ML II model achieved a sensitivity of 99% when identifying sleep, it, like Johansson’s et al.’s algorithm, struggled with detecting awake epochs. This was reflected in the low specificity score of 0.45, reported in our study. The challenge of low specificity is not unique to methods using data collected from thigh-worn accelerometers. Conley et al.’s meta-analysis (35) reported similar findings when estimating sleep using wrist-worn accelerometers among healthy adults, with a mean sensitivity, accuracy, and specificity of 0.89, 0.88, and 0.53, respectively. Furthermore, Patterson and colleagues (36) recently summarized the performance of various heuristic algorithms, ML, and deep learning models used to predict sleep. They found the mean sensitivity and specificity to be 93% (SD = 2.8) and 60% (SD = 11.1) respectively. These findings underscore the challenge of automating the detection of awake periods while in bed. Interestingly, despite low specificity on our ML II model, we observed an overestimation on SOL and WASO, contrasting with most previous research (11, 35). This overestimation of awake epochs indicated that only a small proportion of the awake predictions were correct. In a recent study by Meredith-Jones et al. (2024) the assessment of sleep using Axivity AX3 and ActiGraph GT3X devices worn on the wrist, back, hip, and thigh was compared to PSG among participants aged 8–16 years (32). Eleven count-based algorithms were investigated and the results from the study suggested that for device placement the wrist provided the most accurate sleep estimates. For acceleration measured at the thigh the count-scaled algorithm had mean differences of 14, 2, 16, 4, and 26 minutes for sleep onset, sleep offset, sleep period time, TST, and WASO, respectively, and a mean difference of 3.5% for SE (32). However, no single algorithm consistently provided the most accurate assessment across all evaluated sleep outcomes and none of the algorithms were originally developed for use with thigh, which might explain the findings. Nevertheless, all eleven algorithms did demonstrate an impressive accuracy with epoch-by-epoch sleep-wake classification ranging from 71.9% to 87.3%, sensitivity ranging from 88.1% to 95.5% and specificity ranging from 52.8% to 87.2% (32). The epoch-by-epoch sleep/wake accuracy found in the present study was 93.2% with a sensitivity of 99.0% and specificity of 45.0%. The epoch-by-epoch performance for sleep classification seemed comparable between studies suggesting that algorithms developed for other placements could provide similar performance as an advanced method targeted to the specific device placement. However, using PSG in the assessment of sleep is an obtrusive method which enforces a protocol which facilitates a restricted evening and morning routine and may interfere with normal sleep or normal routines. The obtrusiveness of the PSG method makes it potentially easier to classify sleep-wake correctly and might not provide an accurate validity of the algorithms with measurements made in a natural environment. A total of thirteen features were used in the ML I model when predicting TIB. The feature importance analysis showed the most important feature for classifying TIB was the WalkCirc. The WalkCirc feature was intended to describe the general walkable period during TOB. Other studies developing methods for identifying TIB have used circadian rhythm derived time features to improve the classification accuracy (37). However, time generated features might be problematic as these features are based on a priori knowledge that bedtime is expected under a strict statistical assumption not based on actual behavior. Including the WalkCirc feature allowed for the capture of behavior just prior to getting into bed and just after getting out of bed which often includes walking. In contrast, Skovgaard and colleagues (38) reported that a “sitting” feature provided the strongest signal when manually identifying TIB. Whereas physical activity intensity was not an accurate indicator of getting into or out of bed. Differences in the importance of features for identifying TIB might be explained by the physical activity intensity used to describe each feature. In the study by Skovgaard et al. (38) the features generated from acceleration at the thigh used activity type classifications developed by Skotte et al. (21) which also includes the identification of walking. This seems to suggest that the WalkCirc feature used in this study might be generated from the identification of walking. However, the identification of walking in the Skotte et al. method was based on intensity and not the identification of steps from heel strike and the requirement for at least 5 steps. Thus, using only intensity might be too sensitive and introduce misclassification. Swalve et al. suggested that using temperature data could improve the validity of sleep assessment with activity monitors (39). In our ML II model, the contribution from temperature was only 2.1% of the overall feature importance. This was to some extent disappointing and might be explained by the Axivity AX3 device being filled with composite material. The reason to fill the device with a composite material is most likely to ensure the IP68 water resistance, which provides the ability for the device to resist water at 1.5-meter depth for up to 30 minutes. However, the water resistance comes at the price of a slow response to temperature changes. The slow response introduces a delay in the estimated temperature and thus diminishes the ability of the device to provide detail about short duration changes in sleep to wake. For future device development it would be interesting to place the temperature sensor closer to the casing surface and to use a casing material with higher conductivity for temperature. This solution was used with the discontinued device SenseWear Armband introduced by BodyMedia (40). During TIB the volume of wake as compared to sleep is clearly imbalanced (41) suggesting that the classification of sleep/wake should take this into account. There are different methods available to over sample the wake data to ensure the optimal balance in the two classes. Synthetic Minority Over-sampling Technique (SMOTE) is a method to address class imbalance in datasets. It generates synthetic samples for the minority class by interpolating between existing samples and their nearest neighbors, improving model performance on imbalanced data. In this study we did attempt to use SMOTE but this did not improve performance as expected. This might be explained by the “wake” data per se and specifically that the interpolation simply does not add new information. This seemed to suggest that it was not the class imbalance which caused the poor performance of the wake identification but the lack of signal feature type or quality and thus the inherit challenge discriminating brain derived sleep from wake with acceleration and temperature measured at the thigh. The method developed in this study used an epoch length of 30 seconds which for physical activity assessment using accelerometry would typically underestimate high intensity activity (42). This could suggest that using a shorter epoch might add valuable information and thus improve the identification of wake during TIB. We conducted multiple sensitivity analyses (results not shown) in the development of the method and reducing the epoch length did not improve the model. This might be explained by the granularity of sleep data per se and that increasing the epoch length only adds more complexity to the data than adding valuable information due to the placement of the accelerometer. Leg movements during the night are typical during body relocation, which is quite different from arm movements during the night. The arms are commonly moved more during sleep as compared to the legs and full body which suggest that short epoch length with wrist worn devices might be valuable and improve performance. The increased complexity when using shorter epoch length and thigh acceleration might be solved by using ML methods, which are more flexible. However, it is important to acknowledge that linking PSG recorded sleep and wake with accelerometer-measured movements on the thigh is challenging and using too flexible ML methods with the classification of wake time during TIB is highly sensitive to overfitting which will cause poor generalizability with real world recordings (38). In this study we used an independent data set to validate the performance of the TIB and sleep/wake identification. Typically, the development of methods for TIB and sleep/wake classification are evaluated by splitting the dataset into 80% development and 20% validation. However, validation in an external independent dataset is better than a simple 80/20 split because it avoids potential data leakage or biases that may arise when training and testing data share the same source, ensuring a truly unbiased assessment of the methods performance in diverse real-world conditions. In the independent dataset, the sleep was determined using PSG, which is considered the gold standard for estimating sleep because it provides a comprehensive, objective assessment of sleep architecture and quality. It records multiple physiological signals simultaneously, including brain activity (EEG), eye movements (EOG), muscle activity (EMG), heart rate, and respiratory parameters, allowing precise identification of sleep stages and detection of sleep disorders. Its accuracy and breadth of data make it unmatched in sleep research and clinical diagnosis. However, the application of PSG is also quite complex and might interfere with the subject’s sleep and especially with the behavior occurring before and after getting into bed. In this proposed method the period before and after getting into bed was important for the accurate classification of bedtime and the validation performance was potentially not a true reflection of the performance with real world data. This further amplifies the importance of developing a method for the classification of TIB that avoids relying on a highly flexible ML method to prevent overfitting and ensure generalizability. Strength and limitations A major strength of this study is the use of multiple nights for each subject in the development of the ML method for the classification of TIB and sleep/wake. The use of multiple nights allowed the algorithm to capture important intra-individual night-to-night variability, likely enhancing generalizability when applied to new individuals. Also, the validation of the method in an independent sample is a major strength of this study. However, EEG is considered a strong reference for developing sleep algorithms for accelerometry, EEG and especially PSG is an obtrusive method which require trained staff to conduct and the behavior before going to bed might not reflect the subject’s typical behavior during the evening. Conducting PSG recordings will impact the movement and thus the acceleration measured before and after entering bed. The ZM device used in the development of the proposed method was administered by the family and not as intrusive to the behavior before sleep as PSG and might be a better reflection of the actual behavior preceding and succeeding bedtime. Subjects in the development dataset were asked to record three nights during baseline and follow-up using the ZM. This provided a large amount of training data for developing the method. However, there were numerous sensor problems observed with the recordings and from viewing the raw sleep staging it was clear that for some nights the instructed measurement protocol was not followed. To ensure sufficient data quality we needed to manually validate each night leading to the exclusion of 2/3 of the data, which was a limitation. Lastly, in some applications napping might be an important aspect to capture. Unfortunately, none of the available datasets contained napping, presenting a limitation to capture napping behavior reliably. Conclusion In this study, a method utilizing two machine learning models for the classification of acceleration measured at the thigh into time in bed, and sleep/wake was developed. The method was validated in an independent sample and demonstrated a sufficient accuracy for assessing bedtime, wake time, time in bed and total sleep time at the group level; however, wide limits of agreement indicated that caution is warranted when interpreting sleep estimates for individual participants. The accuracy of estimating sleep efficiency was low and the estimation of sleep onset latency and wake after sleep onset was not valid with the proposed method. Abbreviations TIB Time in bed TOB Time out of bed TST Total sleep time SE Sleep efficiency SOL Sleep onset latency WASO Wake after sleep onset PSG Polysomnography ZM The Zmachine Insight+ ML Machine learning SCREENS The SCREENS trial WalkCirc Walking bout circadian rhythm NZ Independent validation study, New Zealand ML I Machine learning model I, time in bed classification ML II Machine learning model II, sleep classification SMOTE Synthetic Minority Over-sampling Technique Declarations Ethics approval and consent to participate The SCREENS trial has been approved by the Ethical Committee of Southern Denmark (S- 20170213), and all data handling processes complied with the General Data Protection Regulation, ensuring the ethical and secure management of participant information. Ethical approval for the New Zealand study was obtained from the University of Otago Human Ethics Committee (ref H18/073). Informed consent was obtained on behalf of all participants in both studies. Consent for publication Not applicable Funding This work was supported by funding from TrygFonden (grant number ID 130081 and 115606) and the European Research Council (grant number 716657). Author Contribution AG acquired funding for the SCREENS trial, and conceptualization of the SCREENS trial was carried out by AG, PLK, JSP, SOS, and SRM. JSP, SOS, and SRM collected the data. Conceptualization, data curation, formal analysis, methodology development, and implementation of the two-model machine learning algorithm was done by JCB. RT and KMJ designed and conceptualized the NZ study used for independent validation. MSJ and JCB drafted the original manuscript. Visualization of the manuscript was performed by MSJ. EHL, AG, PLK, JSP, SOS, SRM, KMJ, and RT, conducted a critical review and provided substantial feedback on the manuscript. All authors read and approved the final manuscript. Data Availability The data underlying this study are not publicly available due to legal and ethical restrictions under the General Data Protection Regulation (GDPR) and Danish data protection laws. Researchers who wish to access the data must submit a formal application to the relevant steering committee(s) of the respective projects. Access will be granted only if the proposed use complies with applicable legal and ethical requirements.The computer code and models are shared in a GitHub repository: jbrond/ThighSleepIndentification (https://github.com/jbrond/ThighSleepIndentification.git) References 1. Ohayon M, Wickwire EM, Hirshkowitz M, Albert SM, Avidan A, Daly FJ, et al. National Sleep Foundation's sleep quality recommendations: first report. Sleep Health. 2017;3(1):6–19; doi:10.1016/j.sleh.2016.11.006 2. Hirshkowitz M, Whiton K, Albert SM, Alessi C, Bruni O, DonCarlos L, et al. National Sleep Foundation's updated sleep duration recommendations: final report. Sleep Health. 2015;1(4):233 − 43; doi:10.1016/j.sleh.2015.10.004 3. Van de Water AT, Holmes A, Hurley DA. Objective measurements of sleep for non-laboratory settings as alternatives to polysomnography–a systematic review. J Sleep Res. 2011;20(1 Pt 2):183–200; doi:10.1111/j.1365-2869.2009.00814.x 4. Lee YJ, Lee JY, Cho JH, Choi JH. Interrater reliability of sleep stage scoring: a meta-analysis. J Clin Sleep Med. 2022;18(1):193–202; doi:10.5664/jcsm.9538 5. Arnardottir ES, Islind AS, Óskarsdóttir M. The Future of Sleep Measurements: A Review and Perspective. Sleep Med Clin. 2021;16(3):447 − 64; doi:10.1016/j.jsmc.2021.05.004 6. Kaplan RF, Wang Y, Loparo KA, Kelly MR, Bootzin RR. Performance evaluation of an automated single-channel sleep-wake detection algorithm. Nat Sci Sleep. 2014;6:113 − 22; doi:10.2147/nss.S71159 7. Wang Y, Loparo KA, Kelly MR, Kaplan RF. Evaluation of an automated single-channel sleep staging algorithm. Nat Sci Sleep. 2015;7:101 − 11; doi:10.2147/nss.S77888 8. Pedersen J, Rasmussen MGB, Olesen LG, Kristensen PL, Grøntved A. Self-administered electroencephalography-based sleep assessment: compliance and perceived feasibility in children and adults. Sleep Science and Practice. 2021;5(1):8; doi:10.1186/s41606-021-00059-1 9. Webster JB, Kripke DF, Messin S, Mullaney DJ, Wyborney G. An activity-based sleep monitor system for ambulatory use. Sleep. 1982;5(4):389 − 99; doi:10.1093/sleep/5.4.389 10. Cole RJ, Kripke DF, Gruen W, Mullaney DJ, Gillin JC. Automatic sleep/wake identification from wrist activity. Sleep. 1992;15(5):461-9; doi:10.1093/sleep/15.5.461 11. Palotti J, Mall R, Aupetit M, Rueschman M, Singh M, Sathyanarayana A, et al. Benchmark on a large cohort for sleep-wake classification with machine learning techniques. NPJ Digit Med. 2019;2:50; doi:10.1038/s41746-019-0126-9 12. Sazonov E, Sazonova N, Schuckers S, Neuman M. Activity-based sleep-wake identification in infants. Physiol Meas. 2004;25(5):1291 − 304; doi:10.1088/0967-3334/25/5/018 13. Sadeh A, Sharkey KM, Carskadon MA. Activity-based sleep-wake identification: an empirical test of methodological issues. Sleep. 1994;17(3):201-7; doi:10.1093/sleep/17.3.201 14. van Hees VT, Sabia S, Anderson KN, Denton SJ, Oliver J, Catt M, et al. A Novel, Open Access Method to Assess Sleep Duration Using a Wrist-Worn Accelerometer. PLoS One. 2015;10(11):e0142533; doi:10.1371/journal.pone.0142533 15. Sundararajan K, Georgievska S, Te Lindert BHW, Gehrman PR, Ramautar J, Mazzotti DR, et al. Sleep classification from wrist-worn accelerometer data using random forests. Sci Rep. 2021;11(1):24; doi:10.1038/s41598-020-79217-x 16. Carlson JA, Tuz-Zahra F, Bellettiere J, Ridgers ND, Steel C, Bejarano C, et al. Validity of Two Awake Wear-Time Classification Algorithms for activPAL in Youth, Adults, and Older Adults. J Meas Phys Behav. 2021;4(2):151 − 62; doi:10.1123/jmpb.2020-0045 17. Inan-Eroglu E, Huang B-H, Shepherd L, Pearson N, Koster A, Palm P, et al. Comparison of a Thigh-Worn Accelerometer Algorithm With Diary Estimates of Time in Bed and Time Asleep: The 1970 British Cohort Study. Journal for the Measurement of Physical Behaviour. 2021;4(1):60 − 7; doi:10.1123/jmpb.2020-0033 18. van der Berg JD, Willems PJ, van der Velde JH, Savelberg HH, Schaper NC, Schram MT, et al. Identifying waking time in 24-h accelerometry data in adults using an automated algorithm. J Sports Sci. 2016;34(19):1867-73; doi:10.1080/02640414.2016.1140908 19. Winkler EA, Bodicoat DH, Healy GN, Bakrania K, Yates T, Owen N, et al. Identifying adults' valid waking wear time by automated estimation in activPAL data collected with a 24 h wear protocol. Physiol Meas. 2016;37(10):1653-68; doi:10.1088/0967-3334/37/10/1653 20. Johansson PJ, Crowley P, Axelsson J, Franklin K, Garde AH, Hettiarachchi P, et al. Development and performance of a sleep estimation algorithm using a single accelerometer placed on the thigh: an evaluation against polysomnography. J Sleep Res. 2023;32(2):e13725; doi:10.1111/jsr.13725 21. Skotte J, Korshøj M, Kristiansen J, Hanisch C, Holtermann A. Detection of physical activity types using triaxial accelerometers. J Phys Act Health. 2014;11(1):76–84; doi:10.1123/jpah.2011-0347 22. Arvidsson D, Fridolfsson J, Börjesson M, Andersen LB, Ekblom Ö, Dencker M, et al. Re-examination of accelerometer data processing and calibration for the assessment of physical activity intensity. Scand J Med Sci Sports. 2019;29(10):1442-52; doi:10.1111/sms.13470 23. Brønd JC, Grøntved A, Andersen LB, Arvidsson D, Olesen LG. Simple Method for the Objective Activity Type Assessment with Preschoolers, Children and Adolescents. Children (Basel). 2020;7(7); doi:10.3390/children7070072 24. Rasmussen MGB, Pedersen J, Olesen LG, Brage S, Klakk H, Kristensen PL, et al. Short-term efficacy of reducing screen media use on physical activity, sleep, and physiological stress in families with children aged 4–14: study protocol for the SCREENS randomized controlled trial. BMC Public Health. 2020;20(1):380; doi:10.1186/s12889-020-8458-6 25. Pedersen J, Rasmussen MGB, Sørensen SO, Mortensen SR, Olesen LG, Brønd JC, et al. Effects of Limiting Recreational Screen Media Use on Physical Activity and Sleep in Families With Children: A Cluster Randomized Clinical Trial. JAMA Pediatr. 2022;176(8):741-9; doi:10.1001/jamapediatrics.2022.1519 26. Galland BC, Taylor BJ, Elder DE, Herbison P. Normal sleep patterns in infants and children: a systematic review of observational studies. Sleep Med Rev. 2012;16(3):213 − 22; doi:10.1016/j.smrv.2011.06.001 27. Pedersen MJ, Leonthin H, Mahler B, Rittig S, Jennum PJ, Kamperis K. Two nights of home polysomnography in healthy 7-14-year-old children - Feasibility and intraindividual variability. Sleep Med. 2023;101:87–92; doi:10.1016/j.sleep.2022.10.027 28. Palm L, Persson E, Elmqvist D, Blennow G. Sleep and wakefulness in normal preadolescent children. Sleep. 1989;12(4):299–308; doi:10.1093/sleep/12.4.299 29. Skovgaard EL, Roswall MA, Pedersen NH, Larsen KT, Grøntved A, Brønd JC. Generalizability and performance of methods to detect non-wear with free-living accelerometer recordings. Sci Rep. 2023;13(1):2496; doi:10.1038/s41598-023-29666-x 30. Godfrey A, Del Din S, Barry G, Mathers JC, Rochester L. Instrumenting gait with an accelerometer: a system and algorithm examination. Med Eng Phys. 2015;37(4):400-7; doi:10.1016/j.medengphy.2015.02.003 31. Gustafsson M, Schiøttz-Christensen B, Wedderkopp N, Brønd JC. Step Count in Patients With Lumbar Spinal Stenosis: Accuracy During Walking and Nonwalking Activities. Spine (Phila Pa 1976). 2022;47(17):1203-11; doi:10.1097/brs.0000000000004385 32. Meredith-Jones KA, Haszard JJ, Graham-DeMello A, Campbell A, Stewart T, Galland BC, et al. Validation of actigraphy sleep metrics in children aged 8 to 16 years: considerations for device type, placement and algorithms. Int J Behav Nutr Phys Act. 2024;21(1):40; doi:10.1186/s12966-024-01590-x 33. HJORTH MF, CHAPUT J-P, DAMSGAARD CT, DALSKOV S-M, MICHAELSEN KF, TETENS I, et al. Measure of sleep and physical activity by a single accelerometer: Can a waist-worn Actigraph adequately measure sleep in children? Sleep and Biological Rhythms. 2012;10(4):328 − 35; doi: 10.1111/j.1479-8425.2012.00578.x 34. Kushida CA, Chang A, Gadkary C, Guilleminault C, Carrillo O, Dement WC. Comparison of actigraphic, polysomnographic, and subjective assessment of sleep parameters in sleep-disordered patients. Sleep Med. 2001;2(5):389 − 96; doi:10.1016/s1389-9457(00)00098-8 35. Conley S, Knies A, Batten J, Ash G, Miner B, Hwang Y, et al. Agreement between actigraphic and polysomnographic measures of sleep in adults with and without chronic conditions: A systematic review and meta-analysis. Sleep Med Rev. 2019;46:151 − 60; doi:10.1016/j.smrv.2019.05.001 36. Patterson MR, Nunes AAS, Gerstel D, Pilkar R, Guthrie T, Neishabouri A, et al. 40 years of actigraphy in sleep medicine and current state of the art algorithms. NPJ Digit Med. 2023;6(1):51; doi:10.1038/s41746-023-00802-1 37. Walch O, Huang Y, Forger D, Goldstein C. Sleep stage prediction with raw acceleration and photoplethysmography heart rate data derived from a consumer wearable device. Sleep. 2019;42(12); doi:10.1093/sleep/zsz180 38. Skovgaard EL, Pedersen J, Møller NC, Grøntved A, Brønd JC. Manual Annotation of Time in Bed Using Free-Living Recordings of Accelerometry Data. Sensors (Basel). 2021;21(24); doi:10.3390/s21248442 39. Swalve N, Harfmann B, Mitrzyk J, Montoye AHK. Utility of Activity Monitors and Thermometry in Assessing Sleep Stages and Sleep Quality. Journal for the Measurement of Physical Behaviour. 2018;1(3):108 − 21; doi:10.1123/jmpb.2018-0012 40. Lopez GA, Brønd JC, Andersen LB, Dencker M, Arvidsson D. Validation of SenseWear Armband in children, adolescents, and adults. Scand J Med Sci Sports. 2018;28(2):487 − 95; doi:10.1111/sms.12920 41. Galland BC, Short MA, Terrill P, Rigney G, Haszard JJ, Coussens S, et al. Establishing normal values for pediatric nighttime sleep measured by actigraphy: a systematic review and meta-analysis. Sleep. 2018;41(4); doi:10.1093/sleep/zsy017 42. Edwardson CL, Gorely T. Epoch length and its effect on physical activity intensity. Med Sci Sports Exerc. 2010;42(5):928 − 34; doi:10.1249/MSS.0b013e3181c301f5 Additional Declarations No competing interests reported. Cite Share Download PDF Status: Published Journal Publication published 24 Apr, 2026 Read the published version in Sleep Science and Practice → Version 1 posted Editorial decision: Revision requested 26 Nov, 2025 Reviews received at journal 21 Nov, 2025 Reviews received at journal 14 Nov, 2025 Reviews received at journal 10 Nov, 2025 Reviewers agreed at journal 07 Nov, 2025 Reviewers agreed at journal 06 Nov, 2025 Reviewers agreed at journal 05 Nov, 2025 Reviewers agreed at journal 05 Nov, 2025 Reviewers invited by journal 04 Nov, 2025 Editor assigned by journal 30 Oct, 2025 Submission checks completed at journal 30 Oct, 2025 First submitted to journal 23 Oct, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7935022","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":542831464,"identity":"e2cbb75b-1855-4d71-95b2-9b499da23f17","order_by":0,"name":"Maja Sulstad Johansen","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA70lEQVRIie3PsYoCMRCA4VkEqzzANrpPIAwIcoX4LLMsbBXQ3sLiCsttc+BDKIJ1loA2C7YBLRTBaouFa2Jncu2dOe0s8hdDivlIAhAKvWeRdKOI7TyBOxMAk34j7VL0JdwyvUKW+lnSKyosb2bUWR8+yyY1R0jmHKGuHpOB5qgYZf3NcZvFRFfAqsZooX2EoQJqpRtrLVGAMccWazxkbx9maJauxfjbOJKI/4jkKBmpdBnz9s8toB3xPiyfKJbv+kLngw/KFcPqOikXvu/v1epshtNOIbKLboaqm8yz1anePia/Ym7IF0AoFAqF/ugOZ2VhqmHYzIkAAAAASUVORK5CYII=","orcid":"","institution":"University of Southern Denmark","correspondingAuthor":true,"prefix":"","firstName":"Maja","middleName":"Sulstad","lastName":"Johansen","suffix":""},{"id":542831465,"identity":"0b5ec984-19c3-4a4c-8772-a6ad82068ebc","order_by":1,"name":"Esben Høegholm Lykke","email":"","orcid":"","institution":"University of Southern Denmark","correspondingAuthor":false,"prefix":"","firstName":"Esben","middleName":"Høegholm","lastName":"Lykke","suffix":""},{"id":542831466,"identity":"76bacaca-0c0c-4daf-9c4c-f4a9a2a6fbd8","order_by":2,"name":"Anders Grøntved","email":"","orcid":"","institution":"University of Southern Denmark","correspondingAuthor":false,"prefix":"","firstName":"Anders","middleName":"","lastName":"Grøntved","suffix":""},{"id":542831467,"identity":"b1523635-5e8e-4c74-9ed6-16342343a330","order_by":3,"name":"Peter Lund Kristensen","email":"","orcid":"","institution":"University of Southern Denmark","correspondingAuthor":false,"prefix":"","firstName":"Peter","middleName":"Lund","lastName":"Kristensen","suffix":""},{"id":542831468,"identity":"c88789c1-f8f1-4f0c-8d15-d3f35e49771e","order_by":4,"name":"Jesper Schmidt-Persson","email":"","orcid":"","institution":"University of Southern Denmark","correspondingAuthor":false,"prefix":"","firstName":"Jesper","middleName":"","lastName":"Schmidt-Persson","suffix":""},{"id":542831469,"identity":"c9ddec29-aa99-432f-88d9-bd7c39f4c812","order_by":5,"name":"Sarah Overgaard Slater","email":"","orcid":"","institution":"University of Southern Denmark","correspondingAuthor":false,"prefix":"","firstName":"Sarah","middleName":"Overgaard","lastName":"Slater","suffix":""},{"id":542831470,"identity":"a034ae45-1f05-4359-9549-7c68f856394e","order_by":6,"name":"Sofie Rath Mortensen","email":"","orcid":"","institution":"University of Southern Denmark","correspondingAuthor":false,"prefix":"","firstName":"Sofie","middleName":"Rath","lastName":"Mortensen","suffix":""},{"id":542831471,"identity":"07d31f32-5248-42eb-b087-c578cde44a27","order_by":7,"name":"Kim Meredith-Jones","email":"","orcid":"","institution":"University of Otago","correspondingAuthor":false,"prefix":"","firstName":"Kim","middleName":"","lastName":"Meredith-Jones","suffix":""},{"id":542831472,"identity":"c6b42586-1f81-43b2-822c-4e914f3ee872","order_by":8,"name":"Rachael Taylor","email":"","orcid":"","institution":"University of Otago","correspondingAuthor":false,"prefix":"","firstName":"Rachael","middleName":"","lastName":"Taylor","suffix":""},{"id":542831473,"identity":"2ed78f4e-aca8-4d13-980b-72f6433b6386","order_by":9,"name":"Jan Christian Brønd","email":"","orcid":"","institution":"University of Southern Denmark","correspondingAuthor":false,"prefix":"","firstName":"Jan","middleName":"Christian","lastName":"Brønd","suffix":""}],"badges":[],"createdAt":"2025-10-23 20:08:13","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7935022/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7935022/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s41606-026-00183-w","type":"published","date":"2026-04-24T15:58:01+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":96049216,"identity":"49b0ef7b-9174-486a-bdc6-e4c1fe68fc2f","added_by":"auto","created_at":"2025-11-17 06:31:03","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":133302,"visible":true,"origin":"","legend":"","description":"","filename":"Assessmentofchildrenssleep271025.docx","url":"https://assets-eu.researchsquare.com/files/rs-7935022/v1/196237b061c6aa5288d374b5.docx"},{"id":96049218,"identity":"23aece7f-b618-4421-94ee-206e7e7170f8","added_by":"auto","created_at":"2025-11-17 06:31:03","extension":"json","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11425,"visible":true,"origin":"","legend":"","description":"","filename":"74d1c06ca1c14b8c8bff951b7ea9486e.json","url":"https://assets-eu.researchsquare.com/files/rs-7935022/v1/64a1ce39c22b548fb1a70173.json"},{"id":96049221,"identity":"81464f87-d659-41a3-b6ea-a10f094293df","added_by":"auto","created_at":"2025-11-17 06:31:03","extension":"xml","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":102808,"visible":true,"origin":"","legend":"","description":"","filename":"74d1c06ca1c14b8c8bff951b7ea9486e1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7935022/v1/f77e3d1a5c8a2508c3cf3e29.xml"},{"id":96049222,"identity":"c9f26600-0c35-42de-887b-88d94f14e8da","added_by":"auto","created_at":"2025-11-17 06:31:03","extension":"pdf","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":37301,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7935022/v1/14f8a6c9f4b80243e445a005.pdf"},{"id":96049223,"identity":"82d11413-6e09-48a4-b3bf-5a6c48c8c706","added_by":"auto","created_at":"2025-11-17 06:31:03","extension":"pdf","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":24454,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7935022/v1/3a233f37f1950d4c1146b7ac.pdf"},{"id":96049227,"identity":"c04e8b67-8476-4897-808b-6c53a30e4b0e","added_by":"auto","created_at":"2025-11-17 06:31:03","extension":"pdf","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":26048,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7935022/v1/8fa7bc4bad3d727bd911809c.pdf"},{"id":96049226,"identity":"a990cae4-dc60-4f1c-bcaa-c052d35c36ac","added_by":"auto","created_at":"2025-11-17 06:31:03","extension":"xml","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":102452,"visible":true,"origin":"","legend":"","description":"","filename":"74d1c06ca1c14b8c8bff951b7ea9486e1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7935022/v1/f8b77b303d084313e1ce8db0.xml"},{"id":96049228,"identity":"167f2234-8433-4a95-8aa2-df2fc133135b","added_by":"auto","created_at":"2025-11-17 06:31:03","extension":"html","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":109577,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7935022/v1/58778790f4a9850df36a987d.html"},{"id":96247820,"identity":"9ea3e4f8-23f8-45ab-8b5f-e1541ea91d99","added_by":"auto","created_at":"2025-11-19 07:27:44","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":10702,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eImpact of 5- and 10-minute median filtering on raw ZM awakening predictions.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eFigure legend 1: The difference in number of awakenings between the raw ZM predictions vs. 5-minute, and 10-minute median filtered predictions for a random night (boy, 9 years). Grey line is the raw predictions; black line is the median filtered predictions. A: 5-minute median filter on raw ZM predictions, B: 10-minute median filter on raw ZM predictions.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"Figure1241.png","url":"https://assets-eu.researchsquare.com/files/rs-7935022/v1/eb86023a359bb969a4ae2081.png"},{"id":96049219,"identity":"f77ad409-9934-4c67-af7d-ee57bd3dd450","added_by":"auto","created_at":"2025-11-17 06:31:03","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":28015,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eThe two-model machine learning method.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eFigure legend 2: Graphical overview of the two-model machine learning method to classify bedtime, wake time, time in bed, sleep onset latency, wake after sleep onset, total sleep time, and sleep efficiency.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"Figure1242.png","url":"https://assets-eu.researchsquare.com/files/rs-7935022/v1/2ee009ed8447d07ab184be87.png"},{"id":96049220,"identity":"51621fb6-2b2b-4a3a-a2a7-59be8ce3277e","added_by":"auto","created_at":"2025-11-17 06:31:03","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":105148,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cem\u003eAgreement between PSG-measured and accelerometer-predicted time in bed and total sleep time.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eFigure legend 3: Bland Altman plots of the agreement between the PSG measured and accelerometry predicted time in bed and total sleep time including confidence intervals with mean difference (purple), upper limits of agreement (green) and lower limits of agreement (red) for the subjects included in the independent validation dataset.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"Figure1243.png","url":"https://assets-eu.researchsquare.com/files/rs-7935022/v1/cc48335b7e5ce0849806b17c.png"},{"id":107928185,"identity":"d70ce6c8-3c5a-447a-9a5c-7012eb57c653","added_by":"auto","created_at":"2026-04-27 16:09:06","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":569073,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7935022/v1/b8ab56d9-76d5-4607-a036-007ce758d44e.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Assessment of children’s sleep using thigh-worn Axivity AX3 accelerometers","fulltext":[{"header":"Background","content":"\u003cp\u003eIt is well-recognized that sleep, including duration, quality and timing, plays a critical role in physical and mental health and development of children and adolescents (1, 2). Consequently, accurate assessment of sleep is crucial for tracking sleep patterns and improving our understanding of the sleep-health relationship, with methods required to facilitate assessment in large-scale, longitudinal studies.\u003c/p\u003e\u003cp\u003eThe gold standard for sleep assessment, laboratory-based polysomnography (PSG), is impractical in large-scale observational and experimental studies due to the high cost, need for professional administration, participant burden and susceptibility to rater bias (3, 4). Alternatively, diaries and questionnaires have been used due to their cost-effectiveness and simplicity, although they are subject to recall bias (5). The Zmachine\u0026reg;ฏ Insight+ (ZM), is a wearable EEG-based sleep monitoring device validated against PSG (6, 7), offering an objective assessment of sleep, but can be implemented in a home-based environment without the need for professional monitoring. Crucially, the ZM facilitates multi-night analysis in free-living conditions due to its unobstructive and ease of use (8), capturing natural variation in sleep patterns. This makes it advantageous over single-night PSG, particularly as a data source in machine learning (ML), as it provides multiple nights of measurements without inter-rater bias. Despite these benefits, the ZM still poses a significant participant burden and cost.\u003c/p\u003e\u003cp\u003eAn innovative approach involves device-based measurement. These tools, which estimate several sleep metrics, are advantageous due to their reduced participant burden and elimination of potential recall bias. Prominent examples of such tools are body-worn accelerometers, which offer a practical and affordable means of assessing sleep patterns at home for extended periods. These devices also enable the capture of natural day-to-day variations in sleep that may be oversimplified by subjective methods. Their use in sleep and wake classification began with a wrist movement-based algorithm developed by Webster et al. in 1982 (9). This algorithm was refined in 1992 by Cole et al. (10), leading to the widely adopted Cole-Kripke model. With advancements in the field, a variety of techniques, including heuristic algorithms, ML models, regression, and deep learning, are now used to analyze data from hip and wrist-worn accelerometers (10\u0026ndash;15).\u003c/p\u003e\u003cp\u003eWhile wrist and hip-worn devices have benefited from extensive methodological development, thigh-worn accelerometers have not seen the same level of advancement. Existing studies mainly focus on distinguishing sleep from wakefulness, with emphasis on defining \u0026lsquo;waking time\u0026rsquo; and \u0026lsquo;bedtime\u0026rsquo; (16\u0026ndash;19). While recent advances in estimating sleep duration using thigh-worn devices include the promising algorithm by Johansson et al. (20), the application of ML in this area is still underexplored. Considering the potential of thigh-worn accelerometers for accurate physical behavior assessment and identification of postural transitions such as standing and sitting (21\u0026ndash;23), there is a significant research gap.\u003c/p\u003e\u003cp\u003eThus, this study aimed to develop a method using two machine learning models to estimate bedtime, wake time, and sleep metrics, utilizing raw data collected from thigh-worn Axivity AX3, a tri-axial accelerometer, and validate the method against PSG in an external dataset.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eDataset and Participants\u003c/h2\u003e\u003cp\u003eThis study used data from the SCREENS trial (24, 25), conducted from June 2019 to March 2021 in the Region of Southern Denmark, that evaluated the effect of limiting screen media usage within families. For our analysis, we focused on data from child participants aged 4 to 17 years. Our primary sources of data were accelerometer readings from Axivity AX3 devices (Axivity Ltd, Newcastle, United Kingdom), and EEG-derived sleep states and sleep metrics from the ZM device, from three consecutive days during baseline and follow-up.\u003c/p\u003e\u003cp\u003eThe Axivity AX3, an unobtrusive 3-axis accelerometer, was attached midway between the hip and knee on the right anterior thigh using elastic belts, and worn 24-hours a day for the full three days. Sleep state information was extracted using the ZM, a product of General Sleep Corporation. The ZM, which utilizes EEG hardware and signal processing algorithms, employs three self-adhesive, disposable sensors placed outside the hairline for EEG signal acquisition. The participants of the SCREENS trial were instructed to attach the ZM device when they went to bed and remove the device upon getting out of bed. The ZM uses two proprietary algorithms: Z-ALG and Z-PLUS. The Z-ALG is utilized for sleep detection, showcasing its suitability for in-home monitoring (6), while the Z-PLUS differentiates sleep stages, as evidenced by its alignment with expert evaluations using PSG data (7).\u003c/p\u003e\u003cp\u003eIn the current study, we treated all sleep stages (light sleep (N1 \u0026amp; N2), deep sleep (N3), and REM sleep) as a single category effectively deducing the output of the ZM to \u0026ldquo;awake\u0026rdquo; and \u0026ldquo;asleep\u0026rdquo;. Previous research has not been successful in accurately distinguishing between sleep stages when using accelerometry, and this approach was chosen because such differentiation is not necessary to derive the sleep metrics of interest and helps simplify the learning process of the ML algorithms. We chose to incorporate a 10-minute median-filtering of the raw predictions from the ZM device into our modeling process. Children typically experience around five to eight sleep cycles per night, with awakenings most likely at the end of each cycle (26). In examining the raw ZM predictions, we observed a notable overestimation in the number of awakenings for the children in our study, surpassing expected counts based on typical sleep cycle patterns illustrated in Fig.\u0026nbsp;1. The average sleep efficiency determined by the raw ZM predictions for our sample was 83%, which is slightly below the threshold of \u0026ge;85% indicating good sleep quality (1). In fact, prior research has indicated sleep efficiencies of over 90% in similar child cohorts (27, 28). This discrepancy suggests that the raw ZM predictions might be overestimating awake periods. Many of these brief awakenings could be considered as noise, which when present in the data, can potentially hinder the learning process of ML algorithms by obscuring the underlying patterns that the algorithms are trying to learn, leading to less accurate predictions.\u003c/p\u003e\u003cp\u003e\u003cem\u003eFigure caption 1: Impact of 5- and 10-minute median filtering on raw ZM awakening predictions.\u003c/em\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003e(Fig. 1)\u003c/h3\u003e\n\u003cp\u003e\u003cem\u003eFigure legend 1: The difference in number of awakenings between the raw ZM predictions vs. 5-minute, and 10-minute median filtered predictions for a random night (boy, 9 years). Grey line is the raw predictions; black line is the median filtered predictions. A: 5-minute median filter on raw ZM predictions, B: 10-minute median filter on raw ZM predictions.\u003c/em\u003e\u003c/p\u003e\u003cp\u003eOnly ZM recordings accompanied by complete accelerometer data and lasting between 7 and 14 hours were considered. These cutoffs were set to exclude very short or long periods of estimated sleep, as several children did not adhere to the protocol e.g. removing the ZM during the night or failing to turn off after removal. All nights were manually curated and nights during which the ZM reported sensor issues were excluded.\u003c/p\u003e\n\u003ch3\u003eData Preprocessing and Feature Extraction\u003c/h3\u003e\n\u003cp\u003eProcessing of the raw accelerometer data began with a low-pass filtration step using a 4th order Butterworth filter with a 5 Hz cut-off frequency to eliminate high-frequency noise similar to methods described by Skotte et al. (21). Any non-wear data was identified using previously described methods (29) and nights with non-wear were excluded in the development and validation of the method. The acceleration was processed in two second windows with one second overlap, providing a resolution of one second. For each window, the mean, standard deviation (SD), inclination, angle and turn angles were calculated. The data were then aggregated using a mean into 30-second epochs, so every sample corresponds to a 30-second epoch scored during the ZM recordings. For each of the inclination, angle and turn features a 30 second window SD was calculated and added. A minimum 5 step walking bout circadian rhythm feature (WalkCirc) was generated by using a wavelet step detection (30, 31) and subsequently enveloping from the last step bout at 11 PM and the first step bout after 7 AM. The feature defines the primary period of the day for which the subject is moving around (the first bout of walking until the last bout of walking). The feature is binary and pre-defines what is day and what is potentially night, thereby directing the algorithm on identifying the circadian rhythm. Together with a categorical weekday feature (weekday or weekend) the complete data set contains thirteen unique features describing the movement behavior and is presented in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e.\u003c/p\u003e\n\u003ch3\u003eModel training - Classification of sleep metrics\u003c/h3\u003e\n\u003cp\u003eThe sleep metrics included in this study are defined as follows:\u003c/p\u003e\u003cp\u003e\u003cem\u003eBedtime\u003c/em\u003e \u0026ndash; The start of the ZM recording, representing the timepoint when sleep is initiated.\u003c/p\u003e\u003cp\u003e\u003cem\u003eWake time\u003c/em\u003e \u0026ndash; The ending of the ZM recording, representing the end of sleep.\u003c/p\u003e\u003cp\u003e\u003cem\u003eTime In Bed\u003c/em\u003e (TIB) - The total duration of time in bed, which is defined as the time from the start to the end of the ZM recording.\u003c/p\u003e\u003cp\u003e\u003cem\u003eSleep Onset Latency\u003c/em\u003e (SOL) - The time it takes to transition from wakefulness to sustained sleep. It is calculated as the time from the beginning of the ZM recording until the first period when 10 out of 12 minutes are scored as sleep.\u003c/p\u003e\u003cp\u003e\u003cem\u003eWake After Sleep Onset\u003c/em\u003e (WASO) \u0026ndash; Time spent awake after initially falling asleep and before the final awakening. In our analysis, a period is counted as \u0026lsquo;awake\u0026rsquo; only if it consists of 3 or more contiguous 30-second epochs, which is also how the ZM summarizes WASO.\u003c/p\u003e\u003cp\u003e\u003cem\u003eTotal Sleep Time\u003c/em\u003e (TST) - The time spent asleep within the TIB, without sleep onset latency and wake after sleep onset.\u003c/p\u003e\u003cp\u003e\u003cem\u003eSleep Efficiency\u003c/em\u003e (SE) - The ratio between TST and TIB, representing the proportion of the time in bed that was actually spent asleep.\u003c/p\u003e\u003cp\u003eThe classification of TIB, and sleep/wake from acceleration measured at the thigh is a simple binary classification. We chose a simple classical decision tree for both classification tasks due to its transparency and interpretability \u0026ndash; particularly relevant in this context, as we wanted to trace how specific input features influence the output and avoid potential overfitting and the black-box issues often associated with highly flexible models like gradient boosting or neural networks. We employed a two-model strategy to assess TIB/time out of bed (TOB) and sleep/wake from the thigh-mounted accelerometer data. The algorithm approach is outlined in Fig.\u0026nbsp;2. The approach was used to simplify the prediction task by decomposing the multiclass problem of classifying out-of-bed-awake, in-bed-awake, and in-bed-asleep into two binary stages: first predicting TIB, then using the suggested TIB data to estimate sleep and wake time while in bed.\u003c/p\u003e\u003cp\u003e\u003cem\u003eFigure caption 2: The two-model machine learning method.\u003c/em\u003e\u003c/p\u003e\n\u003ch3\u003e(Fig. 2)\u003c/h3\u003e\n\u003cp\u003e\u003cem\u003eFigure legend 2: Graphical overview of the two-model machine learning method to classify bedtime, wake time, time in bed, sleep onset latency, wake after sleep onset, total sleep time, and sleep efficiency.\u003c/em\u003e\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eIndependent validation\u003c/h2\u003e\u003cp\u003eThere are numerous options for validating the generalizability of a supervised ML model and we utilized \u003cem\u003ek-\u003c/em\u003efold cross-validation during the development. However, a more realistic estimate of the external validity involves using an independent dataset collected in a different setting in children, using a similar device and placement, along with a reference measurement with sufficient accuracy like PSG or EEG. The independent validation was performed using data from a study conducted in New Zealand (NZ), in which simultaneous thigh-worn accelerometry (Axivity AX3) and PSG (Embletta MPRPG, ST\u0026thinsp;+\u0026thinsp;Proxy and TX Proxy, Natus, California, USA) data over one night from healthy children and adolescents aged 8\u0026ndash;16 years. Detailed procedures for the overnight PSG assessment, including device specifications, signal processing methods, and scoring of sleep stages according to American Academy of Sleep Medicine guidelines, have been described elsewhere (32). Participants in the NZ study were recruited via social media (i.e. Facebook), schools, and word of mouth. Children did not have any history of sleep disturbance (32).\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eStatistical\u003c/h3\u003e\n\u003cp\u003eAll raw accelerometery processing and model development were done in Matlab (Version 24.1.0 R2024a, Mathworks Inc., Natick, Massachusetts, US) and performance evaluation using R version 4.3.0 (2023-04-21) (R Core Team 2023) and the blandr package.\u003c/p\u003e\u003cp\u003eTo assess performance of each model on an epoch-to-epoch basis the following standard evaluation metrics are included:\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:accuracy=\\frac{TP+TN}{TP+TN+FP+FN}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:sensitivity=\\frac{TP}{TP+FN}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:specificity=\\frac{TN}{TN+FP}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equd\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:precision=\\frac{TP}{TP+FP}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Eque\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e\n$$\\:NPV=\\frac{TN}{TN+FN}$$\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Equf\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equf\" name=\"EquationSource\"\u003e\n$$\\:{F}_{1}=2\\cdot\\:\\frac{precision\\cdot\\:sensitivity}{precision+sensitivity}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:NPV\\)\u003c/span\u003e\u003c/span\u003e is negative predictive value, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{F}_{1}\\)\u003c/span\u003e\u003c/span\u003e is the F1 score, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:TP\\)\u003c/span\u003e\u003c/span\u003e is true positives, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:FP\\)\u003c/span\u003e\u003c/span\u003e is false positives, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:TN\\)\u003c/span\u003e\u003c/span\u003e is true negatives, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:FN\\)\u003c/span\u003e\u003c/span\u003e is false negatives.\u003c/p\u003e\u003cp\u003eTIB and sleep are defined as positive labels and TOB and wake as the negative labels in accordance with previous research (33, 34).\u003c/p\u003e\u003cp\u003eTo assess the performance of our models in deriving sleep metrics, we utilized Bland-Altman plots, Pearsons\u0026rsquo;s correlation (\u003cem\u003er\u003c/em\u003e) and Intra Class correlation (ICC). The Bland-Altman method was employed specifically to determine the level of agreement between the ZM and accelerometry derived predictions in the development and the agreement between the PSG and accelerometry predictions in the independent validation. We first calculated the mean difference with 95% CI between the measurement methods and then defined the limits of agreement (LOA) as the mean difference plus or minus 1.96 times the SD of these differences.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eBased on the curation of 906 nights, 597 nights were excluded for failing to meet the cut-off criteria, and 13 children were completely excluded due to no eligible nights, leaving 309 nights from 134 children with a mean of 2.3 nights per child (SD 0.98) eligible for model development. The ZM predictions encompassed 369,708 epochs, each 30 seconds long. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e presents the characteristics of the children in the development dataset (SCREENS), and children in the validation dataset (NZ). The development group had a mean age of 9.51 (SD 2.17) compared to a mean age of 11.66 (SD 2.25) in the independent validation group. The Danish sample was also leaner than the New Zealand sample, with 18.7% overweight compared with 36.0% respectively.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eParticipants characteristics.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSCREENS\u003c/p\u003e\u003cp\u003e\u003cem\u003eDevelopment data\u003c/em\u003e\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eNZ\u003c/p\u003e\u003cp\u003e\u003cem\u003eValidation data\u003c/em\u003e\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003e\u003cb\u003eSex, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eBoys\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e55 (41.04%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e69 (50.74%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eGirls\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e79 (58.96%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e67 (49.26%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAge (years)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e9.51 (2.17)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e11.66 (2.25)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003e\u003cb\u003eBMI-z*, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eThinness for age\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e5 (3.73%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e2 (1.47%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eNormal weight for age\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e104 (77.61%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e85 (62.50%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eOverweight for age\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e25 (18.66%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e49 (36.03%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003e\u003cb\u003eNights\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e309\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e136\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003e\u003cb\u003eDay, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eMonday\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e83 (26.86%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e38 (27.94%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eTuesday\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e102 (33.01%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e23 (16.91%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eWednesday\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e79 (25.57%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e17 (12.50%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eThursday\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e54 (17.48%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e35 (25.74%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eFriday\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e36 (11.65%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e13 (9.56%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eSaturday\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e52 (16.83%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e4 (2.94%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c2\" namest=\"c1\"\u003e\u003cp\u003eSunday\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e58 (18.77%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e6 (4.41%)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eCharacteristics of children participating in the SCREENS and NZ study. Age is presented as mean (SD).\u003c/p\u003e\u003cp\u003e*Based on WHOs interpretation of BMI z-score cut-offs.\u003c/p\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003eModel development\u003c/h2\u003e\u003cdiv id=\"Sec12\" class=\"Section3\"\u003e\u003ch2\u003ePermutation feature importance\u003c/h2\u003e\u003cp\u003eA total of 13 features were used in the development of ML I (TIB classification) whereas only 11 were used in ML II (sleep classification). ML II was applied only to data classified as TIB, which is why the features Steps and WalkCirc were not relevant in this context and were therefore excluded from this model. The importance of the individual features used in both models is presented in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. For TIB classification, the WalkCirc feature was the most important with temperature, inclination angle and SD of the inclination having minor contributions. The most important features in the sleep classification were the SD of the acceleration angle and maximum SD of the acceleration with minor contribution from SD of the acceleration, temperature and SD of the acceleration turn angle.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003ePermutation feature importance.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"3\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFeature\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eML I\u003c/p\u003e\u003cp\u003eTime In Bed\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eML II\u003c/p\u003e\u003cp\u003eSleep/wake\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTempure\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e2.1\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMacc\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eSDacc\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e3.7\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eSDmax\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e15.3\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eInclination\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.5\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAngle\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1.0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTurn\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eInclSD\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.4\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAnglSD\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e73.9\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTurnSD\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e2.4\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eSteps\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eWeekday\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.6\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eWalkCirc\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e98.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eOverview of the features included in the ML I for classifying time in bed and the ML II classifying sleep/wake, along with the percentage importance of each individual feature. Macc represents the mean acceleration, SDmax denotes the maximal standard deviation in acceleration in all three axes, while inclSD, anglSD and turnSD refer to the standard deviations of the inclination, angle, and turn, respectively.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec13\" class=\"Section2\"\u003e\u003ch2\u003eEpoch level performance\u003c/h2\u003e\u003cp\u003eThe performance of the epoch level classification of TIB and sleep with the model development is presented in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. All epoch level classification performance metrics for the identification of TIB were close to perfect and suggests that even though the model was simple it was possible to identify the time when children were in bed. By contrast, while accuracy and sensitivity were also high for detecting sleep, the ability to detect wake was low as indicated by a specificity of 0.45.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eEpoch-by-epoch performance on development data.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"6\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eSensitivity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSpecificity\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eF1\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eML I (time in bed)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.95\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.93\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.99\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.96\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.99\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eML II (sleep/wake)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.93\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.99\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0.45\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.59\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.85\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eModel accuracy, sensitivity, specificity, F1 and precision for the epoch level classification of ML I (classification of time in bed) and ML II (classification of sleep/wake while in bed).\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003eSleep outcomes on development\u003c/h2\u003e\u003cp\u003eMean difference and estimated LOA for the predicted TIB classification and sleep outcomes are presented in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. For the bedtime and wake time mean difference was close to zero and upper and lower LOA were less than 3.8 minutes. For TIB mean difference was \u0026lt;\u0026thinsp;9 minutes and the lower and upper LOA was \u0026minus;\u0026thinsp;67.8 and 85.7 minutes respectively. From the combined outputs of ML I and ML II, we derived TST, SE, SOL, and WASO. Mean difference for TST was zero with lower and upper LOA of -15.3 and 13.4 minutes respectively. The \u003cem\u003er\u003c/em\u003e suggested a strong correlation between the classification of bedtime, wake time, TIB and TST.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eAgreement on development data.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMean difference\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eLOA+\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eLOA-\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003er\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eBedtime (min)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-0.3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-3.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e3.2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.79\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eWake time (min)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-1.3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e2.6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.86\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTIB (min)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e8.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-67.8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e85.7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.72\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTST (min)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-15.3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e13.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.99\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eSE (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-1.5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-11.9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e9.03\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.69\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eSOL (min)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1.7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-51.0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e54.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.49\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eWASO (min)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e4.7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-16.1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e25.4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.94\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eAgreement between the ZM measured and accelerometry predicted bedtime, wake time, time in bed, total sleep time, sleep onset latency, sleep efficiency and wake after sleep onset for the subjects included in the development dataset.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003eIndependent validation\u003c/h2\u003e\u003cp\u003eMean difference, LOA, \u003cem\u003er\u003c/em\u003e and ICC of bedtime and wake time, TIB, TST, SE, WASO and SOL using the independent dataset is presented in Table\u0026nbsp;\u003cspan refid=\"Tab5\" class=\"InternalRef\"\u003e5\u003c/span\u003e and with the Bland Altman plot of both TIB and TST presented in Fig.\u0026nbsp;3. There was a weak \u003cem\u003er\u003c/em\u003e and ICC for SE, SOL and WASO. The \u003cem\u003er\u003c/em\u003e for bedtime, wake time, TIB and TST were moderate to strong, while the ICC for these metrics were moderate. The mean difference for estimating bedtime and wake time was 28.0 minutes (95%CI 19.5, 36.5) and 11.2 minutes (95%CI 1.2, 21.3), respectively. The mean difference for TIB and TST was 14.0 minutes (95%CI 3.3, 24.8) and 3.3 minutes (95%CI -7.3, 13.9) when validated against PSG.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab5\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eAgreement on validation data.\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"9\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c9\" colnum=\"9\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eMean difference (CI)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eLOA- (CI)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eLOA+ (CI)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003er\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eICC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003ep-value\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u003cp\u003eLower\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c9\"\u003e\u003cp\u003eUpper\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eBedtime (min)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e28.0\u003c/p\u003e\u003cp\u003e(19.5, 36.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-64.8\u003c/p\u003e\u003cp\u003e(-79.3, -50.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e120.7\u003c/p\u003e\u003cp\u003e(106.2, 135.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.66\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.59\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.32\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.74\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eWake time (min)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e11.2\u003c/p\u003e\u003cp\u003e(1.2, 21.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-98.7\u003c/p\u003e\u003cp\u003e(-115.9, -81.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e121.2\u003c/p\u003e\u003cp\u003e(104.0, 138.4)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.57\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.55\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.42\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.66\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTIB (min)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e14.0\u003c/p\u003e\u003cp\u003e(3.3, 24.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-103.6\u003c/p\u003e\u003cp\u003e(-122.1, -85.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e131.7\u003c/p\u003e\u003cp\u003e(113.3, 150.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.62\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.59\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.46\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.69\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTST (min)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e3.3\u003c/p\u003e\u003cp\u003e(-7.3, 13.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-112.2\u003c/p\u003e\u003cp\u003e(-130.3, -94.1)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e118.8\u003c/p\u003e\u003cp\u003e(100.7, 136.9)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.57\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.56\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.00\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.43\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.67\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eSE (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-1.6\u003c/p\u003e\u003cp\u003e(-2.6, -0.6)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-12.8\u003c/p\u003e\u003cp\u003e(-14.5, -11.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e9.5\u003c/p\u003e\u003cp\u003e(7.8, 11.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.23\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.21\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.01\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0.05\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.37\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eSOL (min)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e14.2\u003c/p\u003e\u003cp\u003e(10.1, 18.3)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-30.2\u003c/p\u003e\u003cp\u003e(-37.1, -23.2)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e58.6\u003c/p\u003e\u003cp\u003e(51.6, 65.5)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.08\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.01\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.04\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e-0.01\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.05\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eWASO (min)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e18.0\u003c/p\u003e\u003cp\u003e(14.5, 21.6)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-20.9\u003c/p\u003e\u003cp\u003e(-27.0, -14.8)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e56.9\u003c/p\u003e\u003cp\u003e(50.8, 63.0)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0.11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0.04\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0.29\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e-0.09\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c9\"\u003e\u003cp\u003e0.17\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eAgreement between the PSG measured and accelerometry predicted bedtime, wake time, time in bed, total sleep time, sleep onset latency, sleep efficiency and wake after sleep onset for the subjects included in the independent validation dataset.\u003c/p\u003e\u003cp\u003e\u003cem\u003eFigure caption 3: Agreement between PSG-measured and accelerometer-predicted time in bed and total sleep time.\u003c/em\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e(Fig.\u0026nbsp;3)\u003c/h2\u003e\u003cp\u003e\u003cem\u003eFigure legend 3: Bland Altman plots of the agreement between the PSG measured and accelerometry predicted time in bed and total sleep time including confidence intervals with mean difference (purple), upper limits of agreement (green) and lower limits of agreement (red) for the subjects included in the independent validation dataset.\u003c/em\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn this study, a method utilizing two ML models to predict TIB and sleep/wake was developed and validated based on features derived from acceleration and temperature data collected using thigh-worn Axivity AX3 devices. Based on the ML I model to predict TIB and the ML II to predict sleep/wake commonly derived sleep metrics including TST, SE, SOL and WASO were estimated and validated both during model development and using an independent dataset. Overall, thigh-worn estimates of bedtime, wake time, TIB and TST demonstrated moderate to strong validity compared to PSG reference estimates in the independent sample; however, the wide limits of agreement suggested considerable individual variability, and the method appeared more appropriate for analyzing group-level averages or for use in regression models where multiple days of data are available per participant, thereby reducing the impact of individual measurement error. The estimation of SE demonstrated low mean difference but poor correlation whereas for SOL and WASO both mean difference, LOA and correlation indicated that these outcomes were not valid.\u003c/p\u003e\u003cp\u003eCarlson and colleagues evaluated a third-party algorithm, \u0026ldquo;ProcessingPAL,\u0026rdquo; and a proprietary one, \u0026ldquo;CREA,\u0026rdquo; which achieved epoch-by-epoch accuracies of 89.4% and 89.7%, respectively, when classifying TIB and TOB using thigh-worn accelerometry. These algorithms, evaluated against self-reported measures among adolescents and adults, produced F1 scores of 87.2% and 86.6% (16). These scores were slightly lower than the performance of the proposed ML I model, which achieved F1 score and accuracy exceeding 95% in identifying TIB.\u003c/p\u003e\u003cp\u003eThe validation revealed a mean difference of 14.0 minutes in estimating TIB with our ML I model, reflecting trends observed in previous research by Winkler et al. conducted in young- middle-aged and older adults (19). They developed an algorithm that, despite a moderate correlation (\u003cem\u003er\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.67) between their algorithmic results and diary-recorded waking times, overestimated waking time by more than 30 minutes, resulting in an underestimation of TIB (19). This trend was further confirmed when Inan-Eroglu et al. examined Winkler et al.\u0026rsquo;s algorithm, revealing an underestimation of TIB with 9.8 minutes compared to self-reported measures in middle-aged adults (17). Similarly, a study by van der Berg et al. reported a slight underestimation of TIB in a sample of middle-aged and older adults. They employed a unique approach with their algorithm, which relied on quantifying the number and duration of sedentary periods to determine TIB, and active periods (standing or stepping) to identify wake times (18). Important to note is that predictive performance in determining TIB does not necessarily translate to accurate predictions of broader sleep quality metrics. The crucial task of detecting wake periods during TIB, a key factor in assessing further derived sleep quality metrics, is not effectively captured by TIB predictions alone. The distinction between actual sleep and time spent in bed awake, is critical for a comprehensive understanding of sleep quality.\u003c/p\u003e\u003cp\u003eJohansson and colleagues (20) presented in their study epoch-to-epoch performance metrics for sleep scoring using thigh-worn accelerometers in adults. They achieved a mean sensitivity of 0.84, specificity of 0.55, and accuracy of 0.80, using a single-night evaluation dataset of 71 adult subjects. Despite our ML II model achieved a sensitivity of 99% when identifying sleep, it, like Johansson\u0026rsquo;s et al.\u0026rsquo;s algorithm, struggled with detecting awake epochs. This was reflected in the low specificity score of 0.45, reported in our study. The challenge of low specificity is not unique to methods using data collected from thigh-worn accelerometers. Conley et al.\u0026rsquo;s meta-analysis (35) reported similar findings when estimating sleep using wrist-worn accelerometers among healthy adults, with a mean sensitivity, accuracy, and specificity of 0.89, 0.88, and 0.53, respectively. Furthermore, Patterson and colleagues (36) recently summarized the performance of various heuristic algorithms, ML, and deep learning models used to predict sleep. They found the mean sensitivity and specificity to be 93% (SD\u0026thinsp;=\u0026thinsp;2.8) and 60% (SD\u0026thinsp;=\u0026thinsp;11.1) respectively. These findings underscore the challenge of automating the detection of awake periods while in bed. Interestingly, despite low specificity on our ML II model, we observed an overestimation on SOL and WASO, contrasting with most previous research (11, 35). This overestimation of awake epochs indicated that only a small proportion of the awake predictions were correct.\u003c/p\u003e\u003cp\u003eIn a recent study by Meredith-Jones et al. (2024) the assessment of sleep using Axivity AX3 and ActiGraph GT3X devices worn on the wrist, back, hip, and thigh was compared to PSG among participants aged 8\u0026ndash;16 years (32). Eleven count-based algorithms were investigated and the results from the study suggested that for device placement the wrist provided the most accurate sleep estimates. For acceleration measured at the thigh the count-scaled algorithm had mean differences of 14, 2, 16, 4, and 26 minutes for sleep onset, sleep offset, sleep period time, TST, and WASO, respectively, and a mean difference of 3.5% for SE (32). However, no single algorithm consistently provided the most accurate assessment across all evaluated sleep outcomes and none of the algorithms were originally developed for use with thigh, which might explain the findings. Nevertheless, all eleven algorithms did demonstrate an impressive accuracy with epoch-by-epoch sleep-wake classification ranging from 71.9% to 87.3%, sensitivity ranging from 88.1% to 95.5% and specificity ranging from 52.8% to 87.2% (32). The epoch-by-epoch sleep/wake accuracy found in the present study was 93.2% with a sensitivity of 99.0% and specificity of 45.0%. The epoch-by-epoch performance for sleep classification seemed comparable between studies suggesting that algorithms developed for other placements could provide similar performance as an advanced method targeted to the specific device placement. However, using PSG in the assessment of sleep is an obtrusive method which enforces a protocol which facilitates a restricted evening and morning routine and may interfere with normal sleep or normal routines. The obtrusiveness of the PSG method makes it potentially easier to classify sleep-wake correctly and might not provide an accurate validity of the algorithms with measurements made in a natural environment.\u003c/p\u003e\u003cp\u003eA total of thirteen features were used in the ML I model when predicting TIB. The feature importance analysis showed the most important feature for classifying TIB was the WalkCirc. The WalkCirc feature was intended to describe the general walkable period during TOB. Other studies developing methods for identifying TIB have used circadian rhythm derived time features to improve the classification accuracy (37). However, time generated features might be problematic as these features are based on a priori knowledge that bedtime is expected under a strict statistical assumption not based on actual behavior. Including the WalkCirc feature allowed for the capture of behavior just prior to getting into bed and just after getting out of bed which often includes walking. In contrast, Skovgaard and colleagues (38) reported that a \u0026ldquo;sitting\u0026rdquo; feature provided the strongest signal when manually identifying TIB. Whereas physical activity intensity was not an accurate indicator of getting into or out of bed. Differences in the importance of features for identifying TIB might be explained by the physical activity intensity used to describe each feature. In the study by Skovgaard et al. (38) the features generated from acceleration at the thigh used activity type classifications developed by Skotte et al. (21) which also includes the identification of walking. This seems to suggest that the WalkCirc feature used in this study might be generated from the identification of walking. However, the identification of walking in the Skotte et al. method was based on intensity and not the identification of steps from heel strike and the requirement for at least 5 steps. Thus, using only intensity might be too sensitive and introduce misclassification.\u003c/p\u003e\u003cp\u003eSwalve et al. suggested that using temperature data could improve the validity of sleep assessment with activity monitors (39). In our ML II model, the contribution from temperature was only 2.1% of the overall feature importance. This was to some extent disappointing and might be explained by the Axivity AX3 device being filled with composite material. The reason to fill the device with a composite material is most likely to ensure the IP68 water resistance, which provides the ability for the device to resist water at 1.5-meter depth for up to 30 minutes. However, the water resistance comes at the price of a slow response to temperature changes. The slow response introduces a delay in the estimated temperature and thus diminishes the ability of the device to provide detail about short duration changes in sleep to wake. For future device development it would be interesting to place the temperature sensor closer to the casing surface and to use a casing material with higher conductivity for temperature. This solution was used with the discontinued device SenseWear Armband introduced by BodyMedia (40).\u003c/p\u003e\u003cp\u003eDuring TIB the volume of wake as compared to sleep is clearly imbalanced (41) suggesting that the classification of sleep/wake should take this into account. There are different methods available to over sample the wake data to ensure the optimal balance in the two classes. Synthetic Minority Over-sampling Technique (SMOTE) is a method to address class imbalance in datasets. It generates synthetic samples for the minority class by interpolating between existing samples and their nearest neighbors, improving model performance on imbalanced data. In this study we did attempt to use SMOTE but this did not improve performance as expected. This might be explained by the \u0026ldquo;wake\u0026rdquo; data per se and specifically that the interpolation simply does not add new information. This seemed to suggest that it was not the class imbalance which caused the poor performance of the wake identification but the lack of signal feature type or quality and thus the inherit challenge discriminating brain derived sleep from wake with acceleration and temperature measured at the thigh.\u003c/p\u003e\u003cp\u003eThe method developed in this study used an epoch length of 30 seconds which for physical activity assessment using accelerometry would typically underestimate high intensity activity (42). This could suggest that using a shorter epoch might add valuable information and thus improve the identification of wake during TIB. We conducted multiple sensitivity analyses (results not shown) in the development of the method and reducing the epoch length did not improve the model. This might be explained by the granularity of sleep data per se and that increasing the epoch length only adds more complexity to the data than adding valuable information due to the placement of the accelerometer. Leg movements during the night are typical during body relocation, which is quite different from arm movements during the night. The arms are commonly moved more during sleep as compared to the legs and full body which suggest that short epoch length with wrist worn devices might be valuable and improve performance. The increased complexity when using shorter epoch length and thigh acceleration might be solved by using ML methods, which are more flexible. However, it is important to acknowledge that linking PSG recorded sleep and wake with accelerometer-measured movements on the thigh is challenging and using too flexible ML methods with the classification of wake time during TIB is highly sensitive to overfitting which will cause poor generalizability with real world recordings (38).\u003c/p\u003e\u003cp\u003eIn this study we used an independent data set to validate the performance of the TIB and sleep/wake identification. Typically, the development of methods for TIB and sleep/wake classification are evaluated by splitting the dataset into 80% development and 20% validation. However, validation in an external independent dataset is better than a simple 80/20 split because it avoids potential data leakage or biases that may arise when training and testing data share the same source, ensuring a truly unbiased assessment of the methods performance in diverse real-world conditions. In the independent dataset, the sleep was determined using PSG, which is considered the gold standard for estimating sleep because it provides a comprehensive, objective assessment of sleep architecture and quality. It records multiple physiological signals simultaneously, including brain activity (EEG), eye movements (EOG), muscle activity (EMG), heart rate, and respiratory parameters, allowing precise identification of sleep stages and detection of sleep disorders. Its accuracy and breadth of data make it unmatched in sleep research and clinical diagnosis. However, the application of PSG is also quite complex and might interfere with the subject\u0026rsquo;s sleep and especially with the behavior occurring before and after getting into bed. In this proposed method the period before and after getting into bed was important for the accurate classification of bedtime and the validation performance was potentially not a true reflection of the performance with real world data. This further amplifies the importance of developing a method for the classification of TIB that avoids relying on a highly flexible ML method to prevent overfitting and ensure generalizability.\u003c/p\u003e\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e\u003ch2\u003eStrength and limitations\u003c/h2\u003e\u003cp\u003eA major strength of this study is the use of multiple nights for each subject in the development of the ML method for the classification of TIB and sleep/wake. The use of multiple nights allowed the algorithm to capture important intra-individual night-to-night variability, likely enhancing generalizability when applied to new individuals. Also, the validation of the method in an independent sample is a major strength of this study. However, EEG is considered a strong reference for developing sleep algorithms for accelerometry, EEG and especially PSG is an obtrusive method which require trained staff to conduct and the behavior before going to bed might not reflect the subject\u0026rsquo;s typical behavior during the evening. Conducting PSG recordings will impact the movement and thus the acceleration measured before and after entering bed. The ZM device used in the development of the proposed method was administered by the family and not as intrusive to the behavior before sleep as PSG and might be a better reflection of the actual behavior preceding and succeeding bedtime.\u003c/p\u003e\u003cp\u003eSubjects in the development dataset were asked to record three nights during baseline and follow-up using the ZM. This provided a large amount of training data for developing the method. However, there were numerous sensor problems observed with the recordings and from viewing the raw sleep staging it was clear that for some nights the instructed measurement protocol was not followed. To ensure sufficient data quality we needed to manually validate each night leading to the exclusion of 2/3 of the data, which was a limitation. Lastly, in some applications napping might be an important aspect to capture. Unfortunately, none of the available datasets contained napping, presenting a limitation to capture napping behavior reliably.\u003c/p\u003e\u003c/div\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn this study, a method utilizing two machine learning models for the classification of acceleration measured at the thigh into time in bed, and sleep/wake was developed. The method was validated in an independent sample and demonstrated a sufficient accuracy for assessing bedtime, wake time, time in bed and total sleep time at the group level; however, wide limits of agreement indicated that caution is warranted when interpreting sleep estimates for individual participants. The accuracy of estimating sleep efficiency was low and the estimation of sleep onset latency and wake after sleep onset was not valid with the proposed method.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eTIB Time in bed\u003c/p\u003e\u003cp\u003eTOB Time out of bed\u003c/p\u003e\u003cp\u003eTST Total sleep time\u003c/p\u003e\u003cp\u003eSE Sleep efficiency\u003c/p\u003e\u003cp\u003eSOL Sleep onset latency\u003c/p\u003e\u003cp\u003eWASO Wake after sleep onset\u003c/p\u003e\u003cp\u003ePSG Polysomnography\u003c/p\u003e\u003cp\u003eZM The Zmachine Insight+\u003c/p\u003e\u003cp\u003eML Machine learning\u003c/p\u003e\u003cp\u003eSCREENS The SCREENS trial\u003c/p\u003e\u003cp\u003eWalkCirc Walking bout circadian rhythm\u003c/p\u003e\u003cp\u003eNZ Independent validation study, New Zealand\u003c/p\u003e\u003cp\u003eML I Machine learning model I, time in bed classification\u003c/p\u003e\u003cp\u003eML II Machine learning model II, sleep classification\u003c/p\u003e\u003cp\u003eSMOTE Synthetic Minority Over-sampling Technique\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003cp\u003eThe SCREENS trial has been approved by the Ethical Committee of Southern Denmark (S- 20170213), and all data handling processes complied with the General Data Protection Regulation, ensuring the ethical and secure management of participant information.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eEthical approval\u003c/strong\u003e\u003cp\u003efor the New Zealand study was obtained from the University of Otago Human Ethics Committee (ref H18/073).\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eInformed consent\u003c/strong\u003e\u003cp\u003ewas obtained on behalf of all participants in both studies.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003cp\u003eNot applicable\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e\u003cp\u003eThis work was supported by funding from TrygFonden (grant number ID 130081 and 115606) and the European Research Council (grant number 716657).\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eAG acquired funding for the SCREENS trial, and conceptualization of the SCREENS trial was carried out by AG, PLK, JSP, SOS, and SRM. JSP, SOS, and SRM collected the data. Conceptualization, data curation, formal analysis, methodology development, and implementation of the two-model machine learning algorithm was done by JCB. RT and KMJ designed and conceptualized the NZ study used for independent validation. MSJ and JCB drafted the original manuscript. Visualization of the manuscript was performed by MSJ. EHL, AG, PLK, JSP, SOS, SRM, KMJ, and RT, conducted a critical review and provided substantial feedback on the manuscript. All authors read and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe data underlying this study are not publicly available due to legal and ethical restrictions under the General Data Protection Regulation (GDPR) and Danish data protection laws. Researchers who wish to access the data must submit a formal application to the relevant steering committee(s) of the respective projects. Access will be granted only if the proposed use complies with applicable legal and ethical requirements.The computer code and models are shared in a GitHub repository: jbrond/ThighSleepIndentification (https://github.com/jbrond/ThighSleepIndentification.git)\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003e1. Ohayon M, Wickwire EM, Hirshkowitz M, Albert SM, Avidan A, Daly FJ, et al. National Sleep Foundation's sleep quality recommendations: first report. Sleep Health. 2017;3(1):6\u0026ndash;19; doi:10.1016/j.sleh.2016.11.006\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e2. Hirshkowitz M, Whiton K, Albert SM, Alessi C, Bruni O, DonCarlos L, et al. National Sleep Foundation's updated sleep duration recommendations: final report. Sleep Health. 2015;1(4):233\u0026thinsp;\u0026minus;\u0026thinsp;43; doi:10.1016/j.sleh.2015.10.004\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e3. Van de Water AT, Holmes A, Hurley DA. Objective measurements of sleep for non-laboratory settings as alternatives to polysomnography\u0026ndash;a systematic review. J Sleep Res. 2011;20(1 Pt 2):183\u0026ndash;200; doi:10.1111/j.1365-2869.2009.00814.x\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e4. Lee YJ, Lee JY, Cho JH, Choi JH. Interrater reliability of sleep stage scoring: a meta-analysis. J Clin Sleep Med. 2022;18(1):193\u0026ndash;202; doi:10.5664/jcsm.9538\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e5. Arnardottir ES, Islind AS, \u0026Oacute;skarsd\u0026oacute;ttir M. The Future of Sleep Measurements: A Review and Perspective. Sleep Med Clin. 2021;16(3):447\u0026thinsp;\u0026minus;\u0026thinsp;64; doi:10.1016/j.jsmc.2021.05.004\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e6. Kaplan RF, Wang Y, Loparo KA, Kelly MR, Bootzin RR. Performance evaluation of an automated single-channel sleep-wake detection algorithm. Nat Sci Sleep. 2014;6:113\u0026thinsp;\u0026minus;\u0026thinsp;22; doi:10.2147/nss.S71159\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e7. Wang Y, Loparo KA, Kelly MR, Kaplan RF. Evaluation of an automated single-channel sleep staging algorithm. Nat Sci Sleep. 2015;7:101\u0026thinsp;\u0026minus;\u0026thinsp;11; doi:10.2147/nss.S77888\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e8. Pedersen J, Rasmussen MGB, Olesen LG, Kristensen PL, Gr\u0026oslash;ntved A. Self-administered electroencephalography-based sleep assessment: compliance and perceived feasibility in children and adults. Sleep Science and Practice. 2021;5(1):8; doi:10.1186/s41606-021-00059-1\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e9. Webster JB, Kripke DF, Messin S, Mullaney DJ, Wyborney G. An activity-based sleep monitor system for ambulatory use. Sleep. 1982;5(4):389\u0026thinsp;\u0026minus;\u0026thinsp;99; doi:10.1093/sleep/5.4.389\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e10. Cole RJ, Kripke DF, Gruen W, Mullaney DJ, Gillin JC. Automatic sleep/wake identification from wrist activity. Sleep. 1992;15(5):461-9; doi:10.1093/sleep/15.5.461\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e11. Palotti J, Mall R, Aupetit M, Rueschman M, Singh M, Sathyanarayana A, et al. Benchmark on a large cohort for sleep-wake classification with machine learning techniques. NPJ Digit Med. 2019;2:50; doi:10.1038/s41746-019-0126-9\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e12. Sazonov E, Sazonova N, Schuckers S, Neuman M. Activity-based sleep-wake identification in infants. Physiol Meas. 2004;25(5):1291\u0026thinsp;\u0026minus;\u0026thinsp;304; doi:10.1088/0967-3334/25/5/018\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e13. Sadeh A, Sharkey KM, Carskadon MA. Activity-based sleep-wake identification: an empirical test of methodological issues. Sleep. 1994;17(3):201-7; doi:10.1093/sleep/17.3.201\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e14. van Hees VT, Sabia S, Anderson KN, Denton SJ, Oliver J, Catt M, et al. A Novel, Open Access Method to Assess Sleep Duration Using a Wrist-Worn Accelerometer. PLoS One. 2015;10(11):e0142533; doi:10.1371/journal.pone.0142533\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e15. Sundararajan K, Georgievska S, Te Lindert BHW, Gehrman PR, Ramautar J, Mazzotti DR, et al. Sleep classification from wrist-worn accelerometer data using random forests. Sci Rep. 2021;11(1):24; doi:10.1038/s41598-020-79217-x\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e16. Carlson JA, Tuz-Zahra F, Bellettiere J, Ridgers ND, Steel C, Bejarano C, et al. Validity of Two Awake Wear-Time Classification Algorithms for activPAL in Youth, Adults, and Older Adults. J Meas Phys Behav. 2021;4(2):151\u0026thinsp;\u0026minus;\u0026thinsp;62; doi:10.1123/jmpb.2020-0045\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e17. Inan-Eroglu E, Huang B-H, Shepherd L, Pearson N, Koster A, Palm P, et al. Comparison of a Thigh-Worn Accelerometer Algorithm With Diary Estimates of Time in Bed and Time Asleep: The 1970 British Cohort Study. Journal for the Measurement of Physical Behaviour. 2021;4(1):60\u0026thinsp;\u0026minus;\u0026thinsp;7; doi:10.1123/jmpb.2020-0033\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e18. van der Berg JD, Willems PJ, van der Velde JH, Savelberg HH, Schaper NC, Schram MT, et al. Identifying waking time in 24-h accelerometry data in adults using an automated algorithm. J Sports Sci. 2016;34(19):1867-73; doi:10.1080/02640414.2016.1140908\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e19. Winkler EA, Bodicoat DH, Healy GN, Bakrania K, Yates T, Owen N, et al. Identifying adults' valid waking wear time by automated estimation in activPAL data collected with a 24 h wear protocol. Physiol Meas. 2016;37(10):1653-68; doi:10.1088/0967-3334/37/10/1653\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e20. Johansson PJ, Crowley P, Axelsson J, Franklin K, Garde AH, Hettiarachchi P, et al. Development and performance of a sleep estimation algorithm using a single accelerometer placed on the thigh: an evaluation against polysomnography. J Sleep Res. 2023;32(2):e13725; doi:10.1111/jsr.13725\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e21. Skotte J, Korsh\u0026oslash;j M, Kristiansen J, Hanisch C, Holtermann A. Detection of physical activity types using triaxial accelerometers. J Phys Act Health. 2014;11(1):76\u0026ndash;84; doi:10.1123/jpah.2011-0347\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e22. Arvidsson D, Fridolfsson J, B\u0026ouml;rjesson M, Andersen LB, Ekblom \u0026Ouml;, Dencker M, et al. Re-examination of accelerometer data processing and calibration for the assessment of physical activity intensity. Scand J Med Sci Sports. 2019;29(10):1442-52; doi:10.1111/sms.13470\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e23. Br\u0026oslash;nd JC, Gr\u0026oslash;ntved A, Andersen LB, Arvidsson D, Olesen LG. Simple Method for the Objective Activity Type Assessment with Preschoolers, Children and Adolescents. Children (Basel). 2020;7(7); doi:10.3390/children7070072\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e24. Rasmussen MGB, Pedersen J, Olesen LG, Brage S, Klakk H, Kristensen PL, et al. Short-term efficacy of reducing screen media use on physical activity, sleep, and physiological stress in families with children aged 4\u0026ndash;14: study protocol for the SCREENS randomized controlled trial. BMC Public Health. 2020;20(1):380; doi:10.1186/s12889-020-8458-6\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e25. Pedersen J, Rasmussen MGB, S\u0026oslash;rensen SO, Mortensen SR, Olesen LG, Br\u0026oslash;nd JC, et al. Effects of Limiting Recreational Screen Media Use on Physical Activity and Sleep in Families With Children: A Cluster Randomized Clinical Trial. JAMA Pediatr. 2022;176(8):741-9; doi:10.1001/jamapediatrics.2022.1519\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e26. Galland BC, Taylor BJ, Elder DE, Herbison P. Normal sleep patterns in infants and children: a systematic review of observational studies. Sleep Med Rev. 2012;16(3):213\u0026thinsp;\u0026minus;\u0026thinsp;22; doi:10.1016/j.smrv.2011.06.001\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e27. Pedersen MJ, Leonthin H, Mahler B, Rittig S, Jennum PJ, Kamperis K. Two nights of home polysomnography in healthy 7-14-year-old children - Feasibility and intraindividual variability. Sleep Med. 2023;101:87\u0026ndash;92; doi:10.1016/j.sleep.2022.10.027\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e28. Palm L, Persson E, Elmqvist D, Blennow G. Sleep and wakefulness in normal preadolescent children. Sleep. 1989;12(4):299\u0026ndash;308; doi:10.1093/sleep/12.4.299\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e29. Skovgaard EL, Roswall MA, Pedersen NH, Larsen KT, Gr\u0026oslash;ntved A, Br\u0026oslash;nd JC. Generalizability and performance of methods to detect non-wear with free-living accelerometer recordings. Sci Rep. 2023;13(1):2496; doi:10.1038/s41598-023-29666-x\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e30. Godfrey A, Del Din S, Barry G, Mathers JC, Rochester L. Instrumenting gait with an accelerometer: a system and algorithm examination. Med Eng Phys. 2015;37(4):400-7; doi:10.1016/j.medengphy.2015.02.003\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e31. Gustafsson M, Schi\u0026oslash;ttz-Christensen B, Wedderkopp N, Br\u0026oslash;nd JC. Step Count in Patients With Lumbar Spinal Stenosis: Accuracy During Walking and Nonwalking Activities. Spine (Phila Pa 1976). 2022;47(17):1203-11; doi:10.1097/brs.0000000000004385\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e32. Meredith-Jones KA, Haszard JJ, Graham-DeMello A, Campbell A, Stewart T, Galland BC, et al. Validation of actigraphy sleep metrics in children aged 8 to 16 years: considerations for device type, placement and algorithms. Int J Behav Nutr Phys Act. 2024;21(1):40; doi:10.1186/s12966-024-01590-x\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e33. HJORTH MF, CHAPUT J-P, DAMSGAARD CT, DALSKOV S-M, MICHAELSEN KF, TETENS I, et al. Measure of sleep and physical activity by a single accelerometer: Can a waist-worn Actigraph adequately measure sleep in children? Sleep and Biological Rhythms. 2012;10(4):328\u0026thinsp;\u0026minus;\u0026thinsp;35; doi: 10.1111/j.1479-8425.2012.00578.x\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e34. Kushida CA, Chang A, Gadkary C, Guilleminault C, Carrillo O, Dement WC. Comparison of actigraphic, polysomnographic, and subjective assessment of sleep parameters in sleep-disordered patients. Sleep Med. 2001;2(5):389\u0026thinsp;\u0026minus;\u0026thinsp;96; doi:10.1016/s1389-9457(00)00098-8\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e35. Conley S, Knies A, Batten J, Ash G, Miner B, Hwang Y, et al. Agreement between actigraphic and polysomnographic measures of sleep in adults with and without chronic conditions: A systematic review and meta-analysis. Sleep Med Rev. 2019;46:151\u0026thinsp;\u0026minus;\u0026thinsp;60; doi:10.1016/j.smrv.2019.05.001\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e36. Patterson MR, Nunes AAS, Gerstel D, Pilkar R, Guthrie T, Neishabouri A, et al. 40 years of actigraphy in sleep medicine and current state of the art algorithms. NPJ Digit Med. 2023;6(1):51; doi:10.1038/s41746-023-00802-1\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e37. Walch O, Huang Y, Forger D, Goldstein C. Sleep stage prediction with raw acceleration and photoplethysmography heart rate data derived from a consumer wearable device. Sleep. 2019;42(12); doi:10.1093/sleep/zsz180\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e38. Skovgaard EL, Pedersen J, M\u0026oslash;ller NC, Gr\u0026oslash;ntved A, Br\u0026oslash;nd JC. Manual Annotation of Time in Bed Using Free-Living Recordings of Accelerometry Data. Sensors (Basel). 2021;21(24); doi:10.3390/s21248442\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e39. Swalve N, Harfmann B, Mitrzyk J, Montoye AHK. Utility of Activity Monitors and Thermometry in Assessing Sleep Stages and Sleep Quality. Journal for the Measurement of Physical Behaviour. 2018;1(3):108\u0026thinsp;\u0026minus;\u0026thinsp;21; doi:10.1123/jmpb.2018-0012\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e40. Lopez GA, Br\u0026oslash;nd JC, Andersen LB, Dencker M, Arvidsson D. Validation of SenseWear Armband in children, adolescents, and adults. Scand J Med Sci Sports. 2018;28(2):487\u0026thinsp;\u0026minus;\u0026thinsp;95; doi:10.1111/sms.12920\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e41. Galland BC, Short MA, Terrill P, Rigney G, Haszard JJ, Coussens S, et al. Establishing normal values for pediatric nighttime sleep measured by actigraphy: a systematic review and meta-analysis. Sleep. 2018;41(4); doi:10.1093/sleep/zsy017\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e42. Edwardson CL, Gorely T. Epoch length and its effect on physical activity intensity. Med Sci Sports Exerc. 2010;42(5):928\u0026thinsp;\u0026minus;\u0026thinsp;34; doi:10.1249/MSS.0b013e3181c301f5\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"sleep-science-and-practice","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ssap","sideBox":"Learn more about [Sleep Science and Practice](http://sleep.biomedcentral.com)","snPcode":"41606","submissionUrl":"https://submission.nature.com/new-submission/41606/3","title":"Sleep Science and Practice","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"accelerometry, actigraphy, validation, polysomnography, sleep assessment, sleep/wake cycles, children, adolescents, machine learning, wearable sensors","lastPublishedDoi":"10.21203/rs.3.rs-7935022/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7935022/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eAccurate assessment of sleep is vital, but the gold standard, polysomnography, is costly and impractical for large-scale studies. An alternative is wearable accelerometers, which reduce participant burden and eliminate potential recall biases. This study aimed to develop and validate a method for estimating time in bed (TIB), total sleep time (TST), sleep efficiency (SE), sleep onset latency (SOL), and wake after sleep onset (WASO) utilizing machine learning applied to accelerometry data.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eData on 309 nights from 134 children aged 4\u0026ndash;17 years was used to develop a method utilizing two machine learning models applied to data from thigh-worn accelerometers to estimate sleep metrics. Inputs were collected simultaneously from the Zmachine Insight\u0026thinsp;+\u0026thinsp;and raw data from thigh-worn accelerometers, and validated using k-fold cross-validation. The method was then cross-validated against polysomnography in an independent sample of 136 children aged 8\u0026ndash;16 years.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eThe independent validation showed overestimations of 28.0 minutes for bedtime and 11.2 minutes for wake time, with ICC of 0.59 and 0.55. TIB and TST were overestimated by 13.8 and 3.4 minutes with ICC of 0.59 and 0.56, respectively. The correlation for estimating SE, SOL and WASO was weak with ICC of 0.21, 0.01 and 0.04, respectively.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eThis method demonstrated sufficient accuracy for assessing bedtime, wake time, TIB and TST at the group level when validated in an independent sample against polysomnography, although wide limits of agreement suggest limited precision for individual-level assessments. Low agreement for SE, SOL and WASO indicated insufficient accuracy of the method for these metrics.\u003c/p\u003e","manuscriptTitle":"Assessment of children’s sleep using thigh-worn Axivity AX3 accelerometers","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-17 06:30:58","doi":"10.21203/rs.3.rs-7935022/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-11-26T21:59:20+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-21T07:17:03+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-14T14:02:37+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-11-10T16:34:52+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"64474825023256677165523560544464256310","date":"2025-11-07T08:04:36+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"78431827044921143836058877765830173504","date":"2025-11-06T22:29:50+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"166802872849215451325336485057609929861","date":"2025-11-05T16:13:52+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"191219878377066290692040266125802721959","date":"2025-11-05T16:04:35+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-11-04T22:21:48+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-10-30T08:44:22+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-10-30T08:43:46+00:00","index":"","fulltext":""},{"type":"submitted","content":"Sleep Science and Practice","date":"2025-10-23T19:58:01+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"sleep-science-and-practice","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"ssap","sideBox":"Learn more about [Sleep Science and Practice](http://sleep.biomedcentral.com)","snPcode":"41606","submissionUrl":"https://submission.nature.com/new-submission/41606/3","title":"Sleep Science and Practice","twitterHandle":"@BioMedCentral","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"BMC/SO AJ","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"81ce3569-92d5-44a9-8db5-b940a717776f","owner":[],"postedDate":"November 17th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2026-04-27T16:06:58+00:00","versionOfRecord":{"articleIdentity":"rs-7935022","link":"https://doi.org/10.1186/s41606-026-00183-w","journal":{"identity":"sleep-science-and-practice","isVorOnly":false,"title":"Sleep Science and Practice"},"publishedOn":"2026-04-24 15:58:01","publishedOnDateReadable":"April 24th, 2026"},"versionCreatedAt":"2025-11-17 06:30:58","video":"","vorDoi":"10.1186/s41606-026-00183-w","vorDoiUrl":"https://doi.org/10.1186/s41606-026-00183-w","workflowStages":[]},"version":"v1","identity":"rs-7935022","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7935022","identity":"rs-7935022","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}