Wrist-worn Accelerometer Data Research Articles

Hip and Wrist-Worn Accelerometer Data Analysis for Toddler Activities.

Although accelerometry data are widely utilized to estimate physical activity and sedentary behavior among children age 3 years or older, for toddlers age 1 and 2 year(s), accelerometry data recorded during such behaviors have been far less examined. In particular, toddler’s unique behaviors, such as riding in a stroller or being carried by an adult, have not yet been examined. The objective of this study was to describe accelerometry signal outputs recorded during participation in nine types of behaviors (i.e., running, walking, climbing up/down, crawling, riding a ride-on toy, standing, sitting, riding in a stroller/wagon, and being carried by an adult) among toddlers. Twenty-four toddlers aged 13 to 35 months (50% girls) performed various prescribed behaviors during free play in a commercial indoor playroom while wearing ActiGraph wGT3X-BT accelerometers on a hip and a wrist. Participants’ performances were video-recorded. Based on the video data, accelerometer data were annotated with behavior labels to examine accelerometry signal outputs while performing the nine types of behaviors. Accelerometer data collected during 664 behavior assessments from the 21 participants were used for analysis. Hip vertical axis counts for walking were low (median = 49 counts/5 s). They were significantly lower than those recorded while a toddler was “carried” by an adult (median = 144 counts/5 s; p < 0.01). While standing, sitting, and riding in a stroller, very low hip vertical axis counts were registered (median ≤ 5 counts/5 s). Although wrist vertical axis and vector magnitude counts for “carried” were not higher than those for walking, they were higher than the cut-points for sedentary behaviors. Using various accelerometry signal features, machine learning techniques showed 89% accuracy to differentiate the “carried” behavior from ambulatory movements such as running, walking, crawling, and climbing. In conclusion, hip vertical axis counts alone may be unable to capture walking as physical activity and “carried” as sedentary behavior among toddlers. Machine learning techniques that utilize additional accelerometry signal features could help to recognize behavior types, especially to differentiate being “carried” from ambulatory movements.

Identifying Free-Living Physical Activities Using Lab-Based Models with Wearable Accelerometers

The purpose of this study was to classify, and model various physical activities performed by a diverse group of participants in a supervised lab-based protocol and utilize the model to identify physical activity in a free-living setting. Wrist-worn accelerometer data were collected from () adult participants; age 18–64 years, and processed the data to identify and model unique physical activities performed by the participants in controlled settings. The Gaussian mixture model (GMM) and the hidden Markov model (HMM) algorithms were used to model the physical activities with time and frequency-based accelerometer features. An overall model accuracy of 92.7% and 94.7% were achieved to classify 24 physical activities using GMM and HMM, respectively. The most accurate model was then used to identify physical activities performed by 20 participants, each recorded for two free-living sessions of approximately six hours each. The free-living activity intensities were estimated with 80% accuracy and showed the dominance of stationary and light intensity activities in 36 out of 40 recorded sessions. This work proposes a novel activity recognition process to identify unsupervised free-living activities using lab-based classification models. In summary, this study contributes to the use of wearable sensors to identify physical activities and estimate energy expenditure in free-living settings.

Classifier Personalization for Activity Recognition Using Wrist Accelerometers.

Intersubject variability in accelerometer-based activity recognition may significantly affect classification accuracy, limiting a reliable extension of methods to new users. In this paper, we propose an approach for personalizing classification rules to a single person. We demonstrate that the method improves activity detection from wrist-worn accelerometer data on a four-class recognition problem of interest to the exercise science community, where classes are ambulation, cycling, sedentary, and other. We extend a previously published activity classification method based on support vector machines so that it estimates classification uncertainty. Uncertainty is used to drive data label requests from the user, and the resulting label information is used to update the classifier. Two different datasets-one from 33 adults with 26 activity types, and another from 20 youth with 23 activity types-were used to evaluate the method using leave-one-subject-out and leave-one-group-out cross validation. The new method improved overall recognition accuracy up to 11% on average, with some large person-specific improvements (ranging from -2% to +36%). The proposed method is suitable for online implementation supporting real-time recognition systems.

Wrist-worn Accelerometry for Runners: Objective Quantification of Training Load.

ABSTRACTPurposeThis study aimed to apply open-source analysis code to raw habitual physical activity data from wrist-worn monitors to: 1) objectively, unobtrusively, and accurately discriminate between “running” and “nonrunning” days; and 2) develop and compare simple accelerometer-derived metrics of external training load with existing self-report measures.MethodsSeven-day wrist-worn accelerometer (GENEActiv; Activinsights Ltd, Kimbolton, UK) data obtained from 35 experienced runners (age, 41.9 ± 11.4 yr; height, 1.72 ± 0.08 m; mass, 68.5 ± 9.7 kg; body mass index, 23.2 ± 2.2 kg·m−2; 19 [54%] women) every other week over 9 to 18 wk were date-matched with self-reported training log data. Receiver operating characteristic analyses were applied to accelerometer metrics (“Average Acceleration,” “Most Active-30mins,” “Mins≥400 mg”) to discriminate between “running” and “nonrunning” days and cross-validated (leave one out cross-validation). Variance explained in training log criterion metrics (miles, duration, training load) by accelerometer metrics (Mins≥400 mg, “workload (WL) 400-4000 mg”) was examined using linear regression with leave one out cross-validation.ResultsMost Active-30mins and Mins≥400 mg had >94% accuracy for correctly classifying “running” and “nonrunning” days, with validation indicating robustness. Variance explained in miles, duration, and training load by Mins≥400 mg (67%–76%) and WL400–4000 mg (55%–69%) was high, with validation indicating robustness.ConclusionsWrist-worn accelerometer metrics can be used to objectively, unobtrusively, and accurately identify running training days in runners, reducing the need for training logs or user input in future prospective research or commercial activity tracking. The high percentage of variance explained in existing self-reported measures of training load by simple, accelerometer-derived metrics of external training load supports the future use of accelerometry for prospective, preventative, and prescriptive monitoring purposes in runners.

Cross-validation and out-of-sample testing of physical activity intensity predictions with a wrist-worn accelerometer.

Wrist-worn accelerometers are gaining popularity for measurement of physical activity. However, few methods for predicting physical activity intensity from wrist-worn accelerometer data have been tested on data not used to create the methods (out-of-sample data). This study utilized two previously collected data sets [Ball State University (BSU) and Michigan State University (MSU)] in which participants wore a GENEActiv accelerometer on the left wrist while performing sedentary, lifestyle, ambulatory, and exercise activities in simulated free-living settings. Activity intensity was determined via direct observation. Four machine learning models (plus 2 combination methods) and six feature sets were used to predict activity intensity (30-s intervals) with the accelerometer data. Leave-one-out cross-validation and out-of-sample testing were performed to evaluate accuracy in activity intensity prediction, and classification accuracies were used to determine differences among feature sets and machine learning models. In out-of-sample testing, the random forest model (77.3-78.5%) had higher accuracy than other machine learning models (70.9-76.4%) and accuracy similar to combination methods (77.0-77.9%). Feature sets utilizing frequency-domain features had improved accuracy over other feature sets in leave-one-out cross-validation (92.6-92.8% vs. 87.8-91.9% in MSU data set; 79.3-80.2% vs. 76.7-78.4% in BSU data set) but similar or worse accuracy in out-of-sample testing (74.0-77.4% vs. 74.1-79.1% in MSU data set; 76.1-77.0% vs. 75.5-77.3% in BSU data set). All machine learning models outperformed the euclidean norm minus one/GGIR method in out-of-sample testing (69.5-78.5% vs. 53.6-70.6%). From these results, we recommend out-of-sample testing to confirm generalizability of machine learning models. Additionally, random forest models and feature sets with only time-domain features provided the best accuracy for activity intensity prediction from a wrist-worn accelerometer. NEW & NOTEWORTHY This study includes in-sample and out-of-sample cross-validation of an alternate method for deriving meaningful physical activity outcomes from accelerometer data collected with a wrist-worn accelerometer. This method uses machine learning to directly predict activity intensity. By so doing, this study provides a classification model that may avoid high errors present with energy expenditure prediction while still allowing researchers to assess adherence to physical activity guidelines.

Ensemble Methods for Classification of Physical Activities from Wrist Accelerometry.

To investigate whether the use of ensemble learning algorithms improve physical activity recognition accuracy compared to the single classifier algorithms, and to compare the classification accuracy achieved by three conventional ensemble machine learning methods (bagging, boosting, random forest) and a custom ensemble model comprising four algorithms commonly used for activity recognition (binary decision tree, k nearest neighbor, support vector machine, and neural network). The study used three independent data sets that included wrist-worn accelerometer data. For each data set, a four-step classification framework consisting of data preprocessing, feature extraction, normalization and feature selection, and classifier training and testing was implemented. For the custom ensemble, decisions from the single classifiers were aggregated using three decision fusion methods: weighted majority vote, naïve Bayes combination, and behavior knowledge space combination. Classifiers were cross-validated using leave-one subject out cross-validation and compared on the basis of average F1 scores. In all three data sets, ensemble learning methods consistently outperformed the individual classifiers. Among the conventional ensemble methods, random forest models provided consistently high activity recognition; however, the custom ensemble model using weighted majority voting demonstrated the highest classification accuracy in two of the three data sets. Combining multiple individual classifiers using conventional or custom ensemble learning methods can improve activity recognition accuracy from wrist-worn accelerometer data.

Surveillance of Youth Physical Activity and Sedentary Behavior With Wrist Accelerometry.

Accurate tracking of physical activity (PA) and sedentary behavior (SB) is important to advance public health, but little is known about how to interpret wrist-worn accelerometer data. This study compares youth estimates of SB and moderate to vigorous PA (MVPA) obtained using raw and count-based processing methods. Data were collected between April and October 2014 for the National Cancer Institute's Family Life, Activity, Sun, Health, and Eating Study: a cross-sectional Internet-based study of youth/family cancer prevention behaviors. A subsample of 628 adolescents (aged 12-17 years) wore the ActiGraph GT3X+ on the wrist for 7 days. In 2015-2016, SB and MVPA time were calculated from raw data using R-package GGIR and from activity counts data using published cutpoints (Crouter and Chandler). Estimates were compared across age, sex, and weight status to examine the impact of processing methods on behavioral outcomes. ActiGraph data were available for 408 participants. Large differences in SB and MVPA time were observed between processing methods, but age and gender patterns were similar. Younger children (aged 12-14 years) had lower sedentary time and greater MVPA time (p-values <0.05) than older children (aged 15-17 years), consistent across methods. The proportion of youth with ≥60 minutes of MVPA/day was highest with the Crouter methods (~50%) and lowest with GGIR (~0%). Conclusions about youth PA and SB are influenced by the wrist-worn accelerometer data processing method. Efforts to harmonize processing methods are needed to promote standardization and facilitate reporting of monitor-based PA data.

Lab-based validation of different data processing methods for wrist-worn ActiGraph accelerometers in young adults

The wrist is increasingly being used as the preferred site for objectively assessing physical activity but the relative accuracy of processing methods for wrist data has not been determined. Objective: This study evaluates the validity of four processing methods for wrist-worn ActiGraph (AG) data against energy expenditure (EE) measured using a portable metabolic analyzer (OM; Oxycon mobile) and the Compendium of physical activity. Approach: Fifty-one adults (ages 18–40) completed 15 activities ranging from sedentary to vigorous in a laboratory setting while wearing an AG and the OM. Estimates of EE and categorization of activity intensity were obtained from the AG using a linear method based on Hildebrand cutpoints (HLM), a non-linear modification of this method (HNLM), and two methods developed by Staudenmayer based on a Linear Model (SLM) and using random forest (SRF). Estimated EE and classification accuracy were compared to the OM and Compendium using Bland–Altman plots, equivalence testing, mean absolute percent error (MAPE), and Kappa statistics. Main results: Overall, classification agreement with the Compendium was similar across methods ranging from a Kappa of 0.46 (HLM) to 0.54 (HNLM). However, specificity and sensitivity varied by method and intensity, ranging from a sensitivity of 0% (HLM for sedentary) to a specificity of ~99% for all methods for vigorous. None of the methods was significantly equivalent to the OM (p > 0.05). Significance: Across activities, none of the methods evaluated had a high level of agreement with criterion measures. Additional research is needed to further refine the accuracy of processing wrist-worn accelerometer data.

Field evaluation of a random forest activity classifier for wrist-worn accelerometer data.

Wrist-worn accelerometers are convenient to wear and associated with greater wear-time compliance. Previous work has generally relied on choreographed activity trials to train and test classification models. However, validity in free-living contexts is starting to emerge. Study aims were: (1) train and test a random forest activity classifier for wrist accelerometer data; and (2) determine if models trained on laboratory data perform well under free-living conditions. Twenty-one participants (mean age=27.6±6.2) completed seven lab-based activity trials and a 24h free-living trial (N=16). Participants wore a GENEActiv monitor on the non-dominant wrist. Classification models recognising four activity classes (sedentary, stationary+, walking, and running) were trained using time and frequency domain features extracted from 10-s non-overlapping windows. Model performance was evaluated using leave-one-out-cross-validation. Models were implemented using the randomForest package within R. Classifier accuracy during the 24h free living trial was evaluated by calculating agreement with concurrently worn activPAL monitors. Overall classification accuracy for the random forest algorithm was 92.7%. Recognition accuracy for sedentary, stationary+, walking, and running was 80.1%, 95.7%, 91.7%, and 93.7%, respectively for the laboratory protocol. Agreement with the activPAL data (stepping vs. non-stepping) during the 24h free-living trial was excellent and, on average, exceeded 90%. The ICC for stepping time was 0.92 (95% CI=0.75-0.97). However, sensitivity and positive predictive values were modest. Mean bias was 10.3min/d (95% LOA=-46.0 to 25.4min/d). The random forest classifier for wrist accelerometer data yielded accurate group-level predictions under controlled conditions, but was less accurate at identifying stepping verse non-stepping behaviour in free living conditions Future studies should conduct more rigorous field-based evaluations using observation as a criterion measure.

Machine learning for activity recognition: hip versus wrist data

Problem addressed: Wrist-worn accelerometers are associated with greater compliance. However, validated algorithms for predicting activity type from wrist-worn accelerometer data are lacking. This study compared the activity recognition rates of an activity classifier trained on acceleration signal collected on the wrist and hip.Methodology: 52 children and adolescents (mean age 13.7 ± 3.1 year) completed 12 activity trials that were categorized into 7 activity classes: lying down, sitting, standing, walking, running, basketball, and dancing. During each trial, participants wore an ActiGraph GT3X+ tri-axial accelerometer on the right hip and the non-dominant wrist. Features were extracted from 10-s windows and inputted into a regularized logistic regression model using R (Glmnet + L1).Results: Classification accuracy for the hip and wrist was 91.0% ± 3.1% and 88.4% ± 3.0%, respectively. The hip model exhibited excellent classification accuracy for sitting (91.3%), standing (95.8%), walking (95.8%), and running (96.8%); acceptable classification accuracy for lying down (88.3%) and basketball (81.9%); and modest accuracy for dance (64.1%). The wrist model exhibited excellent classification accuracy for sitting (93.0%), standing (91.7%), and walking (95.8%); acceptable classification accuracy for basketball (86.0%); and modest accuracy for running (78.8%), lying down (74.6%) and dance (69.4%).Potential Impact: Both the hip and wrist algorithms achieved acceptable classification accuracy, allowing researchers to use either placement for activity recognition.

Human gait detection from wrist-worn accelerometer data

Human gait detection from wrist-worn accelerometer data