Raw Accelerometry Data Research Articles

Impacts on study design when implementing digital measures in Parkinson's disease-modifying therapy trials.

Parkinson's Disease affects over 8.5 million people and there are currently no medications approved to treat underlying disease. Clinical trials for disease modifying therapies (DMT) are hampered by a lack of sufficiently sensitive measures to detect treatment effect. Reliable digital assessments of motor function allow for frequent at-home measurements that may be able to sensitively detect disease progression. Here, we estimate the test-retest reliability of a suite of at-home motor measures derived from raw triaxial accelerometry data collected from 44 participants (21 with confirmed PD) and use the estimates to simulate digital measures in DMT trials. We consider three schedules of assessments and fit linear mixed models to the simulated data to determine whether a treatment effect can be detected. We find at-home measures vary in reliability; many have ICCs as high as or higher than MDS-UPDRS part III total score. Compared with quarterly in-clinic assessments, frequent at-home measures reduce the sample size needed to detect a 30% reduction in disease progression from over 300 per study arm to 150 or less than 100 for bursts and evenly spaced at-home assessments, respectively. The results regarding superiority of at-home assessments for detecting change over time are robust to relaxing assumptions regarding the responsiveness to disease progression and variability in progression rates. Overall, at-home measures have a favorable reliability profile for sensitive detection of treatment effects in DMT trials. Future work is needed to better understand the causes of variability in PD progression and identify the most appropriate statistical methods for effect detection.

Extracting actigraphy-based walking features with structured functional principal components

Objective. We extract walking features from raw accelerometry data while accounting for varying cadence and commonality of features among subjects. Walking is the most performed type of physical activity. Thus, we explore if an individual’s physical health is related to these walking features. Approach. We use data collected using ActiGraph GT3X+ devices (sampling rate = 80 Hz) as part of the developmental epidemiologic cohort study, I = 48, age = 78.7±5.7 years, 45.8% women. We apply structured functional principal component analysis (SFPCA) to extract features from walking signals on both, the subject-specific and the subject-spectrum-specific level of a fast-paced 400 m walk, an indicator of aerobic fitness in older adults. We also use the subject-specific level feature scores to study their associations with age and physical performance measures. Specifically, we transform the raw data into the frequency domain by applying local Fast Fourier Transform to obtain the walking spectra. SFPCA decomposes these spectra into easily interpretable walking features expressed as cadence and acceleration, which can be related to physical performance. Main results. We found that five subject-specific and 19 subject-spectrum-specific level features explained more than 85% of their respective level variation, thus significantly reducing the complexity of the data. Our results show that 54% of the total data variation arises at the subject-specific and 46% at the subject-spectrum-specific level. Moreover, we found that higher acceleration magnitude at the cadence was associated with younger age, lower BMI, faster average cadence and higher short physical performance battery scores. Lower acceleration magnitude at the cadence and higher acceleration magnitude at cadence multiples 2.5 and 3.5 are related to older age and higher blood pressure. Significance. SFPCA extracted subject-specific level empirical walking features which were meaningfully associated with several health indicators and younger age. Thus, an individual’s walking pattern could shed light on subclinical stages of somatic diseases.

Jerks are Useful: Extracting pulse rate from wrist-placed accelerometry jerk during sleep in children.

Evaluate wrist-placed accelerometry predicted heartrate compared to electrocardiogram (ECG) heartrate in children during sleep. Children (n=82, 61% male, 43.9% Black) wore a wrist-placed Apple Watch Series 7 (AWS7) and ActiGraph GT9X during a polysomnogram. 3-Axis accelerometry data was extracted from AWS7 and the GT9X. Accelerometry heartrate estimates were derived from jerk (the rate of acceleration change), computed using the peak magnitude frequency in short time Fourier Transforms of Hilbert transformed jerk computed from acceleration magnitude. Heartrates from ECG traces were estimated from R-R intervals using R-pulse detection. Lin's Concordance Correlation Coefficient (CCC), mean absolute error (MAE) and mean absolute percent error (MAPE) assessed agreement with ECG estimated heartrate. Secondary analyses explored agreement by polysomnography sleep stage and a signal quality metric. The developed scripts are available on Github. For the GT9X, CCC was poor at -0.11 and MAE and MAPE were high at 16.8 (SD=14.2) beats/minute and 20.4% (SD=18.5%). For AWS7, CCC was moderate at 0.61 while MAE and MAPE were lower at 6.4 (SD=9.9) beats/minute and 7.3% (SD=10.3%). Accelerometry estimated heartrate for AWS7 was more closely related to ECG heartrate during N2, N3 and REM sleep than lights on, wake, and N1 and when signal quality was high. These patterns were not evident for the GT9X. Raw accelerometry data extracted from AWS7, but not the GT9X, can be used to estimate heartrate in children while they sleep. Future work is needed to explore the sources (i.e., hardware, software, etc.) of the GT9X's poor performance.

Comparison of raw accelerometry data from ActiGraph, Apple Watch, Garmin, and Fitbit using a mechanical shaker table.

The purpose of this study was to evaluate the reliability and validity of the raw accelerometry output from research-grade and consumer wearable devices compared to accelerations produced by a mechanical shaker table. Raw accelerometry data from a total of 40 devices (i.e., n = 10 ActiGraph wGT3X-BT, n = 10 Apple Watch Series 7, n = 10 Garmin Vivoactive 4S, and n = 10 Fitbit Sense) were compared to reference accelerations produced by an orbital shaker table at speeds ranging from 0.6 Hz (4.4 milligravity-mg) to 3.2 Hz (124.7mg). Two-way random effects absolute intraclass correlation coefficients (ICC) tested inter-device reliability. Pearson product moment, Lin's concordance correlation coefficient (CCC), absolute error, mean bias, and equivalence testing were calculated to assess the validity between the raw estimates from the devices and the reference metric. Estimates from Apple, ActiGraph, Garmin, and Fitbit were reliable, with ICCs = 0.99, 0.97, 0.88, and 0.88, respectively. Estimates from ActiGraph, Apple, and Fitbit devices exhibited excellent concordance with the reference CCCs = 0.88, 0.83, and 0.85, respectively, while estimates from Garmin exhibited moderate concordance CCC = 0.59 based on the mean aggregation method. ActiGraph, Apple, and Fitbit produced similar absolute errors = 16.9mg, 21.6mg, and 22.0mg, respectively, while Garmin produced higher absolute error = 32.5mg compared to the reference. ActiGraph produced the lowest mean bias 0.0mg (95%CI = -40.0, 41.0). Equivalence testing revealed raw accelerometry data from all devices were not statistically significantly within the equivalence bounds of the shaker speed. Findings from this study provide evidence that raw accelerometry data from Apple, Garmin, and Fitbit devices can be used to reliably estimate movement; however, no estimates were statistically significantly equivalent to the reference. Future studies could explore device-agnostic and harmonization methods for estimating physical activity using the raw accelerometry signals from the consumer wearables studied herein.

Abstract P464: Performance of an Automated Sleep Scoring Approach for Actigraphy Data in Children and Adolescents

Background: Actigraphy has advanced the measurement of sleep for cardiometabolic research in childhood. However, actigraphy methods lack standardization, limiting rigor and reproducibility. GGIR, an open-source R package, facilitates raw accelerometer data processing to generate sleep data in a standardized manner. This tool has been validated and applied to adult data, but less is known about the suitability of this tool for generating sleep estimates among children. We aimed to (1) determine the validity of multiple sleep scoring settings available within GGIR in children, and (2) evaluate how this tool automatically generates sleep data in community-based pediatric studies. Method: The validation study included 30 children (14 female, 8-16y) referred for in-lab overnight polysomnography at Children’s Hospital of Philadelphia. Participants wore a GENEActiv during polysomnography that captured data at 50 Hz, which were aligned into 30-second epochs against the sleep stages from polysomnography scored by a sleep medicine physician. Balanced accuracy was the primary validation metric, which is the average of sensitivity and specificity. For community-based testing, three pediatric cohorts were used: Study 1 (Philadelphia; 12-16y; GENEActiv, N=251), Study 2 (Mexico City, 9-18y; ActiGraph GT3X-BT; N=543), and Study 3 (Houston; 5-8y; ActiGraph GT3X+, N=226). We processed raw accelerometry data with the Cole-Kripke (CK), Sadeh (S), and Van Hees (VH) algorithm settings. Result: For in-lab validation, 271.5 hours of polysomnography data were collected with 80% of the epochs scored as sleep; sleep duration averaged 7.3 hours (SD = 1.3). In rank order, average balanced accuracies were 0.81 (SD = 0.08) for GGIR-CK, 0.77 (SD = 0.09) for GGIR-VH, and 0.68 (SD = 0.11) for GGIR-S. The GGIR-S algorithm had the poorest detection of night awakenings (specificity: 0.38 [SD = 0.21]). For community-based data, without manually altering the data, the average sleep duration was highest when using the Sadeh algorithm in each study (8.3 [SD = 1.6], 8.0 [SD = 1.9], and 5.2 [SD = 4.4] hours for studies 1-3) and lower when using the GGIR-CK algorithm (7.1 [SD = 1.4], 6.9 [SD = 1.7], 4.3 [SD = 3.7] hours for studies 1-3). Approximately 2% of the data were affected by obvious errors (e.g., negative wake after sleep onset), with almost all such errors detected with the Sadeh algorithm. Conclusion: In a controlled setting, GGIR-CK had the optimal balanced accuracy performance for estimating sleep duration in children. In keeping with the lab-based finding of poorer detection of night awakenings, sleep duration was highest in the community samples when using GGIR-S. When applied to community-based data the automated GGIR approach generated sleep data that included errors, necessitating additional data cleaning. Incorporating standardized data cleaning steps within GGIR would improve rigor and reproducibility.

Gender Gap for Accelerometry-Based Physical Activity Across Different Age Groups in 5 Brazilian Cohort Studies.

This study aims to evaluate the gender inequalities in accelerometer-based physical activity (PA) across different age groups using data from 5 Pelotas (Brazil) cohorts. The data come from 4 birth cohort studies, covering all live births in the urban area of Pelotas for each respective year (1982, 1993, 2004, and 2015), and the Como vai? cohort study focusing on 60years and above. Raw accelerometry data were collected on the nondominant wrist using GENEActive/ActiGraph devices and processed with the GGIR package. Overall PA was calculated at ages 1, 2, 4, 6, 11, 15, 18, 23, 30, and 60+ years, while moderate to vigorous PA was calculated from 6years onward. Absolute (difference) and relative (ratio) gender inequalities were calculated and intersectionality between gender and wealth was also evaluated. The sample sizes per cohort ranged from 965 to 3462 participants. The mean absolute gender gap was 19.3minutes (95% confidence interval, 12.7-25.9), with the widest gap at 18years (32.9min; 95% confidence interval, 30.1-35.7) for moderate to vigorous PA. The highest relative inequality was found in older adults (ratio 2.0; 95% confidence interval, 1.92-2.08). Our intersectionality results showed that the poorest men being the most active group, accumulating around 60minutes more moderate to vigorous PA per day compared with the wealthiest women at age 18. Men were more physically active than women in all ages evaluated. PA gender inequalities start at an early age and intensify in transition periods of life. Relative inequalities were marked among older adults.

A Device Agnostic Approach to Predict Children's Activity from Consumer Wearable Accelerometer Data: A Proof-of-Concept Study.

This study examined the potential of a device agnostic approach for predicting physical activity from consumer wearable accelerometry compared with a research-grade accelerometry. Seventy-five 5- to 12-year-olds (58% male, 63% White) participated in a 60-min protocol. Children wore wrist-placed consumer wearables (Apple Watch Series 7 and Garmin Vivoactive 4) and a research-grade device (ActiGraph GT9X) concurrently with an indirect calorimeter (COSMED K5). Activity intensities (i.e., inactive, light, moderate-to-vigorous physical activity) were estimated via indirect calorimetry (criterion), and the Hildebrand thresholds were applied to the raw accelerometer data from the consumer wearables and research-grade device. Epoch-by-epoch (e.g., weighted sensitivity, specificity) and discrepancy (e.g., mean bias, absolute error) analyses evaluated agreement between accelerometry-derived and criterion estimates. Equivalence testing evaluated the equivalence of estimates produced by the consumer wearables and ActiGraph. Estimates produced by the raw accelerometry data from ActiGraph, Apple, and Garmin produced similar criterion agreement with weighted sensitivity = 68.2% (95% confidence interval (CI), 67.1%-69.3%), 73.0% (95% CI, 71.8%-74.3%), and 66.6% (95% CI, 65.7%-67.5%), respectively, and weighted specificity = 84.4% (95% CI, 83.6%-85.2%), 82.0% (95% CI, 80.6%-83.4%), and 75.3% (95% CI, 74.7%-75.9%), respectively. Apple Watch produced the lowest mean bias (inactive, -4.0 ± 4.5; light activity, 2.1 ± 4.0) and absolute error (inactive, 4.9 ± 3.4; light activity, 3.6 ± 2.7) for inactive and light physical activity minutes. For moderate-to-vigorous physical activity, ActiGraph produced the lowest mean bias (1.0 ± 2.9) and absolute error (2.8 ± 2.4). No ActiGraph and consumer wearable device estimates were statistically significantly equivalent. Raw accelerometry estimated inactive and light activity from wrist-placed consumer wearables performed similarly to, if not better than, a research-grade device, when compared with indirect calorimetry. This proof-of-concept study highlights the potential of device-agnostic methods for quantifying physical activity intensity via consumer wearables.

Evaluation of a device-agnostic approach to predict sleep from raw accelerometry data collected by Apple Watch Series 7, Garmin Vivoactive 4, and ActiGraph GT9X Link in children with sleep disruptions

Goal and aimsEvaluate the performance of a sleep scoring algorithm applied to raw accelerometry data collected from research-grade and consumer wearable actigraphy devices against polysomnography. Focus method/technologyAutomatic sleep/wake classification using the Sadeh algorithm applied to raw accelerometry data from ActiGraph GT9X Link, Apple Watch Series 7, and Garmin Vivoactive 4. Reference method/technologyStandard manual PSG sleep scoring. SampleFifty children with disrupted sleep (M = 8.5 years, range = 5-12 years, 42% Black, 64% male). DesignParticipants underwent to single night lab polysomnography while wearing ActiGraph, Apple, and Garmin devices. Core analyticsDiscrepancy and epoch-by-epoch analyses for sleep/wake classification (devices vs. polysomnography). Additional analytics and exploratory analysesEquivalence testing for sleep/wake classification (research-grade actigraphy vs. commercial devices). Core outcomesCompared to polysomnography, accuracy, sensitivity, and specificity were 85.5, 87.4, and 76.8, respectively, for Actigraph; 83.7, 85.2, and 75.8, respectively, for Garmin; and 84.6, 86.2, and 77.2, respectively, for Apple. The magnitude and trend of bias for total sleep time, sleep efficiency, sleep onset latency, and wake after sleep were similar between the research and consumer wearable devices. Important additional outcomesEquivalence testing indicated that total sleep time and sleep efficiency estimates from the research and consumer wearable devices were statistically significantly equivalent. Core conclusionThis study demonstrates that raw acceleration data from consumer wearable devices has the potential to be harnessed to predict sleep in children. While further work is needed, this strategy could overcome current limitations related to proprietary algorithms for predicting sleep in consumer wearable devices.

0280 Unlocking the Potential of Consumer Wearables for Predicting Sleep in Children: A Device-Agnostic Machine Learning Approach

Abstract Introduction Consumer wearables use accelerometry and heart rate to predict sleep, the same data signals used by research-grade devices to classify sleep/wake patterns. However, metrics predicted by consumer wearables are based on black box algorithms limiting their use in research. The objective of this study was to develop algorithms for predicting sleep in children based on the raw accelerometry and heart rate data from a popular consumer wearable device, therefore bypassing the onboard black box algorithms. Methods 38 children (M=8.5 years, SD=2.4, 42% black, 61% male) underwent overnight laboratory-based polysomnography (PSG) while wearing an Apple Watch Series 7. Heart rate and accelerometry data were collected via the Apple Watch application program interface. Features extracted included (1) age, (2) heart rate, (3) y-axis offset angle, (4) y-axis angle relative to x-axis (y-angle), (5) x-axis fast Fourier transformation (FFT) 4Hz, (6) x-axis FFT 9Hz, (7) vector magnitude (VM) FFT 9Hz, (8) VM FFT 14Hz, (9) mean power dispersion (MPD), (10) bandpass filter followed by Euclidean norm/vector magnitude (BFEN), (11) dominant signal power at 0.6–2.5Hz (PMAXBAND), and (12) activity counts. Four machine learning models: logistic regression (LR), K-nearest neighbor (KNN), random forest (RF), and neural network (NN) predicted sleep or wake. Model performance was evaluated by F1 score, precision, sensitivity, and specificity. Feature importance was evaluated using the RF model. Results LR achieved the best performance according to F1 score (RF: 85.04, LR: 86.50, KNN: 82.52, NN: 85.13). LR also had the highest precision (RF:88.33, LR:88.94, KNN: 86.77, NN:87.95), and nearly the highest sensitivity (RF:84.52, LR:86.11, KNN:84.73, NN:86.55) and specificity (RF:75.72, LR:75.00, KNN:54.01, NN:60.57). BFEN was the most important feature, followed by MPD, age, PMAXBAND, VM FFT 14Hz, x-axis FFT 9Hz, heart rate, y-axis offset angle, y-angle, activity count, VM FFT 9Hz, x-axis FFT 4Hz. Conclusion Raw accelerometry data combined with heart rate data from the Apple Watch Series 7 was able to detect sleep and wake in children. Leveraging the raw accelerometry data is a viable alternative to relying on black-box algorithms for sleep classification in consumer wearable devices. Support (if any)

Validity of device-measured activity in patients with an implantable cardioverter-defibrillator

Abstract Funding Acknowledgements Type of funding sources: Public grant(s) – EU funding. Main funding source(s): Horizon Introduction There is a growing interest in the use of accelerometers and sensors embedded in implantable cardioverter-defibrillators (ICDs) for monitoring patient activity. Despite evidence regarding the potential clinical value of device-measured activity (D-PA), the validity of these measurements has not yet been established. Objective To assess the validity of device-measured activity against a research-grade, widely validated wearable accelerometer. Methods This is a subanalysis of the ongoing multicenter, prospective SafeHeart study. Raw accelerometry data was continuously sampled at 50Hz from a wrist-worn accelerometer (GENEActiv) during 12 months. Days with at least 22 hours of wear time were used to create summary measures of time in activity, daily active volume and total slow walking steps. These measures were compared to D-PA harmonised as percentage of active time per day, from four different ICD vendors’ remote transmission data, using linear mixed effect models. Results Wearable and device-measured activity data in 51 ICD patients rendered 1228 days (mean 24 days ±19) with both wearable and device-measured activity data. There were significant differences between wearable and device-measured accelerometery in the average time active per day (Table 1). For two vendors significant associations between D-PA, daily active volume, and total slow walking steps were observed. Also, associations between D-PA and daily active time and moderate vigorous physical activity were found in a third vendor. For the fourth vendor no association between any wearable activity metric and D-PA was found. Inter-patient differences accounted for 73.1% of the total variance in D-PA. Conclusion Results demonstrate substantial differences in device-measured activity measurements compared to research-grade activity data. This has implications for the utility and generalizability of D-PA as clinical parameter.

Application of de-shape synchrosqueezing to estimate gait cadence from a single-sensor accelerometer placed in different body locations

Objective. Commercial and research-grade wearable devices have become increasingly popular over the past decade. Information extracted from devices using accelerometers is frequently summarized as ‘number of steps’ (commercial devices) or ‘activity counts’ (research-grade devices). Raw accelerometry data that can be easily extracted from accelerometers used in research, for instance ActiGraph GT3X+, are frequently discarded. Approach. Our primary goal is proposing an innovative use of the de-shape synchrosqueezing transform to analyze the raw accelerometry data recorded from a single sensor installed in different body locations, particularly the wrist, to extract gait cadence when a subject is walking. The proposed methodology is tested on data collected in a semi-controlled experiment with 32 participants walking on a one-kilometer predefined course. Walking was executed on a flat surface as well as on the stairs (up and down). Main results. The cadences of walking on a flat surface, ascending stairs, and descending stairs, determined from the wrist sensor, are 1.98 ± 0.15 Hz, 1.99 ± 0.26 Hz, and 2.03 ± 0.26 Hz respectively. The cadences are 1.98 ± 0.14 Hz, 1.97 ± 0.25 Hz, and 2.02 ± 0.23 Hz, respectively if determined from the hip sensor, 1.98 ± 0.14 Hz, 1.93 ± 0.22 Hz and 2.06 ± 0.24 Hz, respectively if determined from the left ankle sensor, and 1.98 ± 0.14 Hz, 1.97 ± 0.22 Hz, and 2.04 ± 0.24 Hz, respectively if determined from the right ankle sensor. The difference is statistically significant indicating that the cadence is fastest while descending stairs and slowest when ascending stairs. Also, the standard deviation when the sensor is on the wrist is larger. These findings are in line with our expectations. Conclusion. We show that our proposed algorithm can extract the cadence with high accuracy, even when the sensor is placed on the wrist.

Processing of Accelerometry Data with GGIR in Motor Activity Research Consortium for Health

The Mobile Motor Activity Research Consortium for Health (mMARCH) is a collaborative network of clinical and community studies that employ common digital mobile protocols and collect common clinical and biological measures across participating studies. At a high level, a key scientific goal which spans mMARCH studies is to develop a better understanding of the interrelationships between physical activity (PA), sleep (SL), and circadian rhythmicity (CR) and mental and physical health in children, adolescents, and adults. mMARCH studies employ wrist-worn accelerometry to obtain objective measures of PA/SL/CR. However, there is currently no consensus on a standard data processing pipeline for raw accelerometry data and few open-source tools which facilitate their development. The R package GGIR is the most prominent open-source software package for processing raw accelerometry data, offering great functionality and substantial user flexibility. However, even with GGIR, processing done in a harmonized and reproducible fashion across multiple analytical centers requires a nontrivial amount of expertise combined with a careful implementation. In addition, there are many statistical methods useful for analyzing PA/SL/CR patterns using accelerometry data which are implemented in non-GGIR R packages, including methods from multivariate statistics, functional data analysis, distributional data analysis, and time series analyses. To address the issues of multisite harmonization and additional feature creation, mMARCH developed a streamlined harmonized and reproducible pipeline for loading and cleaning raw accelerometry data via GGIR, merging GGIR, and non-GGIR features of PA/SL/CR together, implementing several additional data and feature quality checks, and performing multiple analyses including Joint and Individual Variation Explained, an unsupervised machine learning dimension reduction technique that identifies latent factors capturing joint across and individual to each of three domains of PA/SL/CR. The pipeline is easily modified to calculate additional features of interest, and allows for studies not affiliated with mMARCH to apply a pipeline which facilitates direct comparisons of scientific results in published work by mMARCH studies. This manuscript describes the pipeline and illustrates the use of combined GGIR and non-GGIR features by applying Joint and Individual Variation Explained to the accelerometry component of CoLaus|PsyCoLaus, one of mMARCH sites. The pipeline is publicly available via open-source R package mMARCH.AC.

Estimating Running Speed From Wrist- or Waist-Worn Wearable Accelerometer Data: A Machine Learning Approach

Background: Running is a popular form of exercise, and its physiological effects are strongly modulated by speed. Accelerometry-based activity monitors are commonly used to measure physical activity in research, but no method exists to estimate running speed from only accelerometer data. Methods: Using three cohorts totaling 72 subjects performing treadmill and outdoor running, we developed linear, ridge, and gradient-boosted tree regression models to estimate running speed from raw accelerometer data from waist- or wrist-worn devices. To assess model performance in a real-world scenario, we deployed the best-performing model to data from 16 additional runners completing a 13-week training program while equipped with waist-worn accelerometers and commercially available foot pods. Results: Linear, ridge, and boosted tree models estimated speed with 12.0%, 11.6%, and 11.2% mean absolute percentage error, respectively, using waist-worn accelerometer data. Errors were greater using wrist-worn data, with linear, ridge, and boosted tree models achieving 13.8%, 14.0%, and 12.8% error. Across 663 free-living runs, speed was significantly associated with run duration (p = .009) and perceived run intensity (p = .008). Speed was nonsignificantly associated with fatigue (p = .07). Estimated speeds differed from foot pod measurements by 7.25%; associations and statistical significance were similar when speed was assessed via accelerometry versus via foot pod. Conclusion: Raw accelerometry data can be used to estimate running speed in free-living data with sufficient accuracy to detect associations with important measures of health and performance. Our approach is most useful in studies where research grade accelerometry is preferable to traditional global positioning system or foot pod-based measurements, such as in large-scale observational studies on physical activity.

AccNet24: A deep learning framework for classifying 24-hour activity behaviours from wrist-worn accelerometer data under free-living environments

ObjectiveAlthough machine learning techniques have been repeatedly used for activity prediction from wearable devices, accurate classification of 24-hour activity behaviour categories from accelerometry data remains a challenge. We developed and validated a deep learning-based framework for classifying 24-hour activity behaviours from wrist-worn accelerometers. MethodsUsing an openly available dataset with free-living wrist-based raw accelerometry data from 151 participants (aged 18–91 years), we developed a deep learning framework named AccNet24 to classify 24-hour activity behaviours. First, the acceleration signal (x, y, and z-axes) was segmented into 30-second nonoverlapping windows, and signal-to-image conversion was performed for each segment. Deep features were automatically extracted from the signal images using transfer learning and transformed into a lower-dimensional feature space. These transformed features were then employed to classify the activity behaviours as sleep, sedentary behaviour, and light-intensity (LPA) and moderate-to-vigorous physical activity (MVPA) using a bidirectional long short-term memory (BiLSTM) recurrent neural network. AccNet24 was trained and validated with data from 101 and 25 randomly selected participants and tested with the remaining unseen 25 participants. We also extracted 112 hand-crafted time and frequency domain features from 30-second windows and used them as inputs to five commonly used machine learning classifiers, including random forest, support vector machines, artificial neural networks, decision tree, and naïve Bayes to classify the 24-hour activity behaviour categories. ResultsUsing the same training, validation, and test data and window size, the classification accuracy of AccNet24 outperformed the accuracy of the other five machine learning classification algorithms by 16%–30% on unseen data. ConclusionAccNet24, relying on signal-to-image conversion, deep feature extraction, and BiLSTM achieved consistently high accuracy (>95 %) in classifying the 24-hour activity behaviour categories as sleep, sedentary, LPA, and MVPA. The next generation accelerometry analytics may rely on deep learning techniques for activity prediction.

Mapping features and patterns of accelerometry data on human movement in different age groups and associated health problems: A cross-sectional study.

Human movement is considered one of the important factors for maintaining an independent life. Individuals in different age groups have different characteristics of locomotion patterns and some health conditions can affect or be affected by mobility changes. Few studies clarify or present data about the influence of different ages and biopsychosocial factors on accelerometry features. The aim of this study was to identify characteristics and variables in the frequency signals for different age groups and their relationship with associated health conditions in raw accelerometry data obtained from the use of a triaxial accelerometer during 7days of activities of daily living. A cross-sectional study was conducted based on the database of the first evaluations of the Epidemiological Study of Movement (EPIMOV) cohort. Frequency, signal amplitude, and entropy accelerometry features of EPIMOV participants who used a triaxial accelerometer for 7days were extracted. Sociodemographic, clinical, anthropometric and physical activity assessments were also performed. Two-way ANOVA was performed to compare accelerometry features within different age groups. A series of stepwise multiple regressions were performed on accelerometry variables to analyze their relationships with demographic, anthropometric and cardiovascular risk variables. The sample consisted mostly of female, white, and high school graduates. The most prevalent cardiovascular risk factors were sedentary behavior and obesity. When analyzing the accelerometry variables, it was possible to observe that the entropy feature, and the counts, decrease in the group of older adults, while the feature of harmonic components of gait (frequency×amplitude) increases in the group of older adults. Regarding the amplitude feature, there were no significant differences between the groups. Through stepwise multiple linear regression, it was possible to observe that demographic, anthropometric and cardiovascular risk factors are associated with most accelerometry variables. The results confirm that human movement can be influenced by different ages, sex, demographic, anthropometric and cardiovascular risk factors. Future studies and clinical analyzes can use the methods proposed in this research to adjust movement patterns for sex and different age groups, thus obtaining new interpretations about human movement.

Methods for estimating physical activity and energy expenditure using raw accelerometry data or novel analytical approaches: a repository, framework, and reporting guidelines

The proliferation of approaches for analyzing accelerometer data using raw acceleration or novel analytic approaches like machine learning (‘novel methods’) outpaces their implementation in practice. This may be due to lack of accessibility, either because authors do not provide their developed models or because these models are difficult to find when included as supplementary material. Additionally, when access to a model is provided, authors may not include example data or instructions on how to use the model. This further hinders use by other researchers, particularly those who are not experts in statistics or writing computer code. Objective: We created a repository of novel methods of analyzing accelerometer data for the estimation of energy expenditure and/or physical activity intensity and a framework and reporting guidelines to guide future work. Approach: Methods were identified from a recent scoping review. Available code, models, sample data, and instructions were compiled or created. Main Results: Sixty-three methods are hosted in the repository, in preschoolers (n = 6), children/adolescents (n = 20), and adults (n = 42), using hip (n = 45), wrist (n = 25), thigh (n = 4), chest (n = 4), ankle (n = 6), other (n = 4), or a combination of monitor wear locations (n = 9). Fifteen models are implemented in R, while 48 are provided as cut-points, equations, or decision trees. Significance: The developed tools should facilitate the use and development of novel methods for analyzing accelerometer data, thus improving data harmonization and consistency across studies. Future advances may involve including models that authors did not link to the original published article or those which identify activity type.

A Deep Learning Approach to Classify Sitting and Sleep History from Raw Accelerometry Data during Simulated Driving.

Prolonged sitting and inadequate sleep can impact driving performance. Therefore, objective knowledge of a driver’s recent sitting and sleep history could help reduce safety risks. This study aimed to apply deep learning to raw accelerometry data collected during a simulated driving task to classify recent sitting and sleep history. Participants (n = 84, Mean ± SD age = 23.5 ± 4.8, 49% Female) completed a seven-day laboratory study. Raw accelerometry data were collected from a thigh-worn accelerometer during a 20-min simulated drive (8:10 h and 17:30 h each day). Two convolutional neural networks (CNNs; ResNet-18 and DixonNet) were trained to classify accelerometry data into four classes (sitting or breaking up sitting and 9-h or 5-h sleep). Accuracy was determined using five-fold cross-validation. ResNet-18 produced higher accuracy scores: 88.6 ± 1.3% for activity (compared to 77.2 ± 2.6% from DixonNet) and 88.6 ± 1.1% for sleep history (compared to 75.2 ± 2.6% from DixonNet). Class activation mapping revealed distinct patterns of movement and postural changes between classes. Findings demonstrate the suitability of CNNs in classifying sitting and sleep history using thigh-worn accelerometer data collected during a simulated drive. This approach has implications for the identification of drivers at risk of fatigue-related impairment.

Calibration of an Accelerometer Activity Index among Older Women and Its Association with Cardiometabolic Risk Factors.

Traditional summary metrics provided by accelerometer device manufacturers, known as counts, are proprietary and manufacturer specific, making them difficult to compare studies using different devices. Alternative summary metrics based on raw accelerometry data have been introduced in recent years. However, they were often not calibrated on ground truth measures of activity-related energy expenditure for direct translation into continuous activity intensity levels. Our purpose is to calibrate, derive, and validate thresholds among women 60 years and older based on a recently proposed transparent raw data based accelerometer activity index (AAI), and to demonstrate its application in association with cardiometabolic risk factors. We first built calibration equations for estimating metabolic equivalents (METs) continuously using AAI and personal characteristics using internal calibration data (n=199). We then derived AAI cutpoints to classify epochs into sedentary behavior and intensity categories. The AAI cutpoints were applied to 4,655 data units in the main study. We then utilized linear models to investigate associations of AAI sedentary behavior and physical activity intensity with cardiometabolic risk factors. We found that AAI demonstrated great predictive accuracy for METs (R2=0.74). AAI-based physical activity measures were associated in the expected directions with body mass index (BMI), blood glucose, and high density lipoprotein (HDL) cholesterol. The calibration framework for AAI and the cutpoints derived for women older than 60 years can be applied to ongoing epidemiologic studies to more accurately define sedentary behavior and physical activity intensity exposures which could improve accuracy of estimated associations with health outcomes.

Investigating Wrist-Based Acceleration Summary Measures across Different Sample Rates towards 24-Hour Physical Activity and Sleep Profile Assessment.

Wrist-worn wearable sensors have attracted considerable research interest because of their potential in providing continuous, longitudinal, non-invasive measurements, leading to insights into Physical Activity (PA), sleep, and circadian variability. Three key practical considerations for research-grade wearables are as follows: (a) choosing an appropriate sample rate, (b) summarizing raw three-dimensional accelerometry data for further processing (accelerometry summary measures), and (c) accurately estimating PA levels and sleep towards understanding participants’ 24-hour profiles. We used the CAPTURE-24 dataset, where 148 participants concurrently wore a wrist-worn three-dimensional accelerometer and a wearable camera over approximately 24 h to obtain minute-by-minute labels: sleep; and sedentary light, moderate, and vigorous PA. We propose a new acceleration summary measure, the Rate of Change Acceleration Movement (ROCAM), and compare its performance against three established approaches summarizing three-dimensional acceleration data towards replicating the minute-by-minute labels. Moreover, we compare findings where the acceleration data was sampled at 10, 25, 50, and 100 Hz. We demonstrate the competitive advantage of ROCAM towards estimating the five labels (80.2% accuracy) and building 24-hour profiles where the sample rate of 10 Hz is fully sufficient. Collectively, these findings provide insights facilitating the deployment of large-scale longitudinal actigraphy data processing towards 24-hour PA and sleep-profile assessment.

Mechanical loading prediction through accelerometry data during walking and running

ABSTRACT Currently, there is no way to assess mechanical loading variables such as peak ground reaction forces (pGRF) and peak loading rate (pLR) in clinical settings. The purpose of this study was to develop accelerometry-based equations to predict both pGRF and pLR during walking and running. One hundred and thirty one subjects (79 females; 76.9 ± 19.6 kg) walked and ran at different speeds (2–14 km·h−1) on a force plate–instrumented treadmill while wearing accelerometers at their ankle, lower back and hip. Regression equations were developed to predict pGRF and pLR from accelerometry data. Leave-one-out cross-validation was used to calculate prediction accuracy and Bland–Altman plots. Our pGRF prediction equation was compared with a reference equation previously published. Body mass and peak acceleration were included for pGRF prediction and body mass and peak acceleration rate for pLR prediction. All pGRF equation coefficients of determination were above 0.96, and a good agreement between actual and predicted pGRF was observed, with a mean absolute percent error (MAPE) below 7.3%. Accuracy indices from our equations were better than previously developed equations. All pLR prediction equations presented a lower accuracy compared to those developed to predict pGRF. Walking and running pGRF can be predicted with high accuracy by accelerometry-based equations, representing an easy way to determine mechanical loading in free-living conditions. The pLR prediction equations yielded a somewhat lower prediction accuracy compared with the pGRF equations.