A random forest classifier for the prediction of energy expenditure and type of physical activity from wrist and hip accelerometers

Longitudinal Changes in Children's Accelerometer-derived Activity Pattern Metrics

PURPOSE

The objective of this study was to quantify age-related changes in accelerometer-derived day-level physical activity and sedentary behavior pattern metrics (i.e., number, length, and temporal dispersion of bouts and breaks) across 3 yr of middle childhood. Differences by child sex and weekend versus weekday were examined.

METHOD

Children (N = 169, 54% female, 56% Hispanic; 8-12 yr old at enrollment) participated in a longitudinal study with six assessments across 3 yr. Day-level moderate-to-vigorous physical activity (MVPA; i.e., total minutes, number of short (<10 min) bouts, proportion of long (≥20 min) bouts, temporal dispersion) and sedentary behavior (i.e., total minutes, number of breaks, proportion of long (≥60 min) bouts, temporal dispersion) pattern metrics were measured using a waist-worn accelerometer (Actigraph GT3X).

RESULTS

Random intercept multilevel linear regression models showed that age-related decreases in the number of short MVPA bouts per were steeper for girls than for boys (b = -1.28; 95% confidence interval (CI), -1.93 to -0.64; P < 0.01) and on weekend days than on weekdays (b = -1.82; 95% CI, -2.36 to -1.29; P < 0.01). The evenness of the temporal dispersion of MVPA across the day increased more on weekend days than on weekdays as children got older (b = -0.02; 95% CI, -0.02 to -0.01; P < 0.01). Girls had steeper age-related decreases in the number of sedentary breaks per day (b = -2.89; 95% CI, -3.97 to -1.73; P < 0.01) and the evenness of the temporal dispersion of sedentary behavior across the day (b ≤ 0.01; 95% CI, <0.01 to 0.01; P < 0.01) than did boys. Changes in sedentary behavior metrics did not differ between weekend days and weekdays.

CONCLUSION

Strategies to protect against declines in short physical activity bouts and promote sedentary breaks, especially among girls and on weekend days, could reduce cardiometabolic risks.

Determining Physical Activity Characteristics from Wristband Data for Use in Automated Insulin Delivery Systems

Algorithms that can determine the type of physical activity (PA) and quantify the intensity can allow precision medicine approaches, such as automated insulin delivery systems that modulate insulin administration in response to PA. In this work, data from a multi-sensor wristband is used to design classifiers to distinguish among five different physical states (PS) (resting, activities of daily living, running, biking, and resistance training), and to develop models to estimate the energy expenditure (EE) of the PA for diabetes therapy. The data collected are filtered, features are extracted from the reconciled signals, and the extracted features are used by machine learning algorithms, including deep-learning techniques, to obtain accurate PS classification and EE estimation. The various machine learning techniques have different success rates ranging from 75.7% to 94.8% in classifying the five different PS. The deep neural network model with long short-term memory has 94.8% classification accuracy. We achieved 0.5 MET (Metabolic Equivalent of Task) root-mean-square error for EE estimation accuracy, relative to indirect calorimetry with randomly selected testing data (10% of collected data). We also demonstrate a 5% improvement in PS classification accuracy and a 0.34 MET decrease in the mean absolute error when using multi-sensor approach relative to using only accelerometer data.

Calibrating Wrist-Worn Accelerometers for Physical Activity Assessment in Preschoolers: Machine Learning Approaches

BACKGROUND

Physical activity (PA) level is associated with multiple health benefits during early childhood. However, inconsistency in the methods for quantification of PA levels among preschoolers remains a problem.

OBJECTIVE

This study aimed to develop PA intensity cut points for wrist-worn accelerometers by using machine learning (ML) approaches to assess PA in preschoolers.

METHODS

Wrist- and hip-derived acceleration data were collected simultaneously from 34 preschoolers on 3 consecutive preschool days. Two supervised ML models, receiver operating characteristic curve (ROC) and ordinal logistic regression (OLR), and one unsupervised ML model, k-means cluster analysis, were applied to establish wrist-worn accelerometer vector magnitude (VM) cut points to classify accelerometer counts into sedentary behavior, light PA (LPA), moderate PA (MPA), and vigorous PA (VPA). Physical activity intensity levels identified by hip-worn accelerometer VM cut points were used as reference to train the supervised ML models. Vector magnitude counts were classified by intensity based on three newly established wrist methods and the hip reference to examine classification accuracy. Daily estimates of PA were compared to the hip-reference criterion.

RESULTS

In total, 3600 epochs with matched hip- and wrist-worn accelerometer VM counts were analyzed. All ML approaches performed differently on developing PA intensity cut points for wrist-worn accelerometers. Among the three ML models, k-means cluster analysis derived the following cut points: ≤2556 counts per minute (cpm) for sedentary behavior, 2557-7064 cpm for LPA, 7065-14532 cpm for MPA, and ≥14533 cpm for VPA; in addition, k-means cluster analysis had the highest classification accuracy, with more than 70% of the total epochs being classified into the correct PA categories, as examined by the hip reference. Additionally, k-means cut points exhibited the most accurate estimates on sedentary behavior, LPA, and VPA as the hip reference. None of the three wrist methods were able to accurately assess MPA.

CONCLUSIONS

This study demonstrates the potential of ML approaches in establishing cut points for wrist-worn accelerometers to assess PA in preschoolers. However, the findings from this study warrant additional validation studies.

Discriminating Metabolic Health Status in a Cohort of Nursing Students: Protocol for a Cross-Sectional Study

BACKGROUND

Obesity is currently a worldwide health crisis. Nurses are integral members of the primary health care team and have an important role in managing obesity and administering physical activity (PA) for patients. However, research shows that nurses tend to be overweight or obese, have poor metabolic health, and do not meet PA recommendations. This is problematic because PA is linked to both physiological and psychological well-being and may also influence how nurses counsel their patients. Nursing students are the next generation of nurses; however, there is limited research examining PA (among other lifestyle factors) and metabolic health in nursing students.

OBJECTIVE

The goal of this research is to examine multiple lifestyle factors (including PA, nutrition, sleep, and stress) and determine whether these factors are associated with metabolic health in full-time undergraduate nursing students.

METHODS

An estimated 320 nursing students (18 years of age and older) will be assessed for their metabolic health. Metabolic status will be determined by measuring body mass index (BMI), waist-to-hip ratio (WHR), body fat percentage [skinfold measures (FitSystems Inc)], resting blood pressure [automated oscillatory (Omron Healthcare Inc)], and fasting blood glucose (glucometer). Lifestyle factors will also be measured, including PA and sleep [the International Physical Activity Questionnaire (IPAQ) and 7-day accelerometry (wGT3X-BT, Actigraph LLC)], nutrition [3-day diet log (Nutritionist Pro, Axxya Systems)], and stress [the Depression Anxiety Stress Scale, heart rate variability assessments, and salivary cortisol (ELISA, Eagle Biosciences)]. The association between metabolic status and PA, sleep quantity and quality, nutrition, and stress will be examined by linear regression analyses. Differences by year of study in metabolic health status, PA, sleep, nutrition, and stress will be examined by 1-way analyses of variance (ANOVAs). To determine the ability of PA, sleep, nutrition, and stress to discriminate prevalent overweight and obesity or poor metabolic status, logistic regression and receiver operating characteristic (ROC) curves will be constructed. Statistical analyses will be performed in Stata (version 16.1, StataCorp LLC).

RESULTS

Based on pilot data, we believe senior nursing students will have worse metabolic health (ie, higher BMI and WHR, increased body fat percentage, higher blood pressure, and increased fasting blood glucose) compared to first-year students. We hypothesize that poor PA participation, poor sleep quantity and quality, increased food intake, poor nutrition, and increased stress will be associated with worse metabolic health in full-time nursing students. The study received funding in February 2020. Due to the coronavirus disease 2019 (COVID-19) pandemic, work on this study has been delayed. We are currently completing our application for institutional research ethics approval. Data collection is projected to begin in January 2021, with data collection and analyses expected to be completed by May 2022.

CONCLUSIONS

This study will be the first published research to examine the relationship between lifestyle choices and metabolic status in nursing students attending a Canadian institution. More importantly, the results of this study will support the development of an informed intervention that will target the identified lifestyle factors, improving the physiological and mental health and well-being of nursing students.

INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID)

PRR1-10.2196/21342.

Wearable Device-Independent Next Day Activity and Next Night Sleep Prediction for Rehabilitation Populations

Wearable sensor-based devices are increasingly applied in free-living and clinical settings to collect fine-grained, objective data about activity and sleep behavior. The manufacturers of these devices provide proprietary software that labels the sensor data at specified time intervals with activity and sleep information. If the device wearer has a health condition affecting their movement, such as a stroke, these labels and their values can vary greatly from manufacturer to manufacturer. Consequently, generating outcome predictions based on data collected from patients attending inpatient rehabilitation wearing different sensor devices can be challenging, which hampers usefulness of these data for patient care decisions. In this article, we present a data-driven approach to combining datasets collected from different device manufacturers. With the ability to combine datasets, we merge data from three different device manufacturers to form a larger dataset of time series data collected from 44 patients receiving inpatient therapy services. To gain insights into the recovery process, we use this dataset to build models that predict a patient's next day physical activity duration and next night sleep duration. Using our data-driven approach and the combined dataset, we obtained a normalized root mean square error prediction of 9.11% for daytime physical activity and 11.18% for nighttime sleep duration. Our sleep result is comparable to the accuracy we achieved using the manufacturer's sleep labels (12.26%). Our device-independent predictions are suitable for both point-of-care and remote monitoring applications to provide information to clinicians for customizing therapy services and potentially decreasing recovery time.

A novel prognostic model based on multi-omics features predicts the prognosis of colon cancer patients

Background

As a common malignant tumor in the colon, colon cancer (CC) has high incidence and recurrence rates. This study is designed to build a prognostic model for CC.

Methods

The gene expression dataset, microRNA‐seq dataset, copy number variation (CNV) dataset, DNA methylation dataset, and transcription factor (TF) dataset of CC were downloaded from UCSC Xena database. Using limma package, the differentially methylated genes (DMGs), and differentially expressed genes (DEGs) and miRNAs (DEMs) were identified. Based on random forest method, prognostic model for each omics dataset were constructed. After the omics features related to prognosis were selected using logrank test, the prognostic model based on multi‐omics features was built. Finally, the clinical phenotypes correlated with prognosis were screened using Kaplan–Meier survival analysis, and the nomogram model was established.

Results

There were 1625 DEGs, 268 DEMs, and 386 DMGs between the tumor and normal samples. A total of 105, 29, 159, five, and six genes/sites significantly correlated with prognosis were identified in the gene expression dataset (GABRD), miRNA‐seq dataset (miR‐1271), CNV dataset (RN7SKP247), DNA methylation dataset (cg09170112 methylation site [located in SFSWAP]), and TF dataset (SIX5), respectively. The prognostic model based on multi‐omics features was more effective than those based on single omics dataset. The number of lymph nodes, pathologic_M stage, and pathologic_T stage were the clinical phenotypes correlated with prognosis, based on which the nomogram model was constructed.

Conclusion

The prognostic model based on multi‐omics features and the nomogram model might be valuable for the prognostic prediction of CC.

Prediction of medical expenditures of diagnosed diabetics and the assessment of its related factors using a random forest model, MEPS 2000-2015

OBJECTIVE

To predict the medical expenditures of individual diabetics and assess the related factors of it.

DESIGN AND SETTING

Cross-sectional study.

SETTING AND PARTICIPANTS

Data were collected from the US household component of the medical expenditure panel survey, 2000-2015.

MAIN OUTCOME MEASURE

Random forest (RF) model was performed with the programs of randomForest in R software. Spearman correlation coefficients (rs), mean absolute error (MAE) and mean-related error (MRE) was computed to assess the prediction of all the models.

RESULTS

Total medical expenditure was increased from $105 Billion in 2000 to $318 Billion in 2015. rs, MAE and MRE between the predicted and actual values of medical expenditures in RF model were 0.644, $0.363 and 0.043%. Top one factor in prediction was being treated by the insulin, followed by type of insurance, employment status, age and economical level. The latter four variables had no impact in predicting of medical expenditure by being treated by the insulin. Further, after the sub-analysis of gender and age-groups, the evaluating indicators of prediction were almost identical to each other. Top five variables of total medical expenditure among male were same as those among all the diabetics. Expenses for doctor visits, hospital stay and drugs were also predicted with RF model well. Treatment with insulin was the top one factor of total medical expenditure among female, 18-, 25- and 65-age-groups. Additionally, it indicated that RF model was little superior to traditional regression model.

CONCLUSIONS

RF model could be used in prediction of medical expenditure of diabetics and assessment of its related factors well.

Evaluating the performance of raw and epoch non-wear algorithms using multiple accelerometers and electrocardiogram recordings

Accurate detection of accelerometer non-wear time is crucial for calculating physical activity summary statistics. In this study, we evaluated three epoch-based non-wear algorithms (Hecht, Troiano, and Choi) and one raw-based algorithm (Hees). In addition, we performed a sensitivity analysis to provide insight into the relationship between the algorithms’ hyperparameters and classification performance, as well as to generate tuned hyperparameter values to better detect episodes of wear and non-wear time. We used machine learning to construct a gold-standard dataset by combining two accelerometers and electrocardiogram recordings. The Hecht and Troiano algorithms achieved poor classification performance, while Choi exhibited moderate performance. Meanwhile, Hees outperformed all epoch-based algorithms. The sensitivity analysis and hyperparameter tuning revealed that all algorithms were able to achieve increased classification performance by employing larger intervals and windows, while more stringently defining artificial movement. These classification gains were associated with the ability to lower the false positives (type I error) and do not necessarily indicate a more accurate detection of the total non-wear time. Moreover, our results indicate that with tuned hyperparameters, epoch-based non-wear algorithms are able to perform just as well as raw-based non-wear algorithms with respect to their ability to correctly detect true wear and non-wear episodes.

SmokingOpp: Detecting the Smoking 'Opportunity' Context Using Mobile Sensors

Context plays a key role in impulsive adverse behaviors such as fights, suicide attempts, binge-drinking, and smoking lapse. Several contexts dissuade such behaviors, but some may trigger adverse impulsive behaviors. We define these latter contexts as 'opportunity' contexts, as their passive detection from sensors can be used to deliver context-sensitive interventions. In this paper, we define the general concept of 'opportunity' contexts and apply it to the case of smoking cessation. We operationalize the smoking 'opportunity' context, using self-reported smoking allowance and cigarette availability. We show its clinical utility by establishing its association with smoking occurrences using Granger causality. Next, we mine several informative features from GPS traces, including the novel location context of smoking spots, to develop the SmokingOpp model for automatically detecting the smoking 'opportunity' context. Finally, we train and evaluate the SmokingOpp model using 15 million GPS points and 3,432 self-reports from 90 newly abstinent smokers in a smoking cessation study.

Application of Raw Accelerometer Data and Machine-Learning Techniques to Characterize Human Movement Behavior: A Systematic Scoping Review

BACKGROUND

Application of machine learning for classifying human behavior is increasingly common as access to raw accelerometer data improves. The aims of this scoping review are (1) to examine if machine-learning techniques can accurately identify human activity behaviors from raw accelerometer data and (2) to summarize the practical implications of these machine-learning techniques for future work.

METHODS

Keyword searches were performed in Scopus, Web of Science, and EBSCO databases in 2018. Studies that applied supervised machine-learning techniques to raw accelerometer data and estimated components of physical activity were included. Information on study characteristics, machine-learning techniques, and key study findings were extracted from included studies.

RESULTS

Of the 53 studies included in the review, 75% were published in the last 5 years. Most studies predicted postures and activity type, rather than intensity, and were conducted in controlled environments using 1 or 2 devices. The most common models were support vector machine, random forest, and artificial neural network. Overall, classification accuracy ranged from 62% to 99.8%, although nearly 80% of studies achieved an overall accuracy above 85%.

CONCLUSIONS

Machine-learning algorithms demonstrate good accuracy when predicting physical activity components; however, their application to free-living settings is currently uncertain.

Detecting prolonged sitting bouts with the ActiGraph GT3X

The ActiGraph has a high ability to measure physical activity; however, it lacks an accurate posture classification to measure sedentary behavior. The aim of the present study was to develop an ActiGraph (waist-worn, 30 Hz) posture classification to detect prolonged sitting bouts, and to compare the classification to proprietary ActiGraph data. The activPAL, a highly valid posture classification device, served as reference criterion. Both sensors were worn by 38 office workers over a median duration of 9 days. An automated feature selection extracted the relevant signal information for a minute-based posture classification. The machine learning algorithm with optimal feature number to predict the time in prolonged sitting bouts (≥5 and ≥10 minutes) was searched and compared to the activPAL using Bland-Altman statistics. The comparison included optimized and frequently used cut-points (100 and 150 counts per minute (cpm), with and without low-frequency-extension (LFE) filtering). The new algorithm predicted the time in prolonged sitting bouts most accurate (bias ≤ 7 minutes/d). Of all proprietary ActiGraph methods, only 150 cpm without LFE predicted the time in prolonged sitting bouts non-significantly different from the activPAL (bias ≤ 18 minutes/d). However, the frequently used 100 cpm with LFE accurately predicted total sitting time (bias ≤ 7 minutes/d). To study the health effects of ActiGraph measured prolonged sitting, we recommend using the new algorithm. In case a cut-point is used, we recommend 150 cpm without LFE to measure prolonged sitting and 100 cpm with LFE to measure total sitting time. However, both cpm cut-points are not recommended for a detailed bout analysis.

The validity of two widely used commercial and research-grade activity monitors, during resting, household and activity behaviours

Wearable devices are increasingly prevalent in research environments for the estimation of energy expenditure (EE) and heart rate (HR). The aim of this study was to validate the HR and EE estimates of the Fitbit charge 2 (FC2), and the EE estimates of the Sensewear armband mini (SWA). We recruited 59 healthy adults to participate in walking, running, cycling, sedentary and household tasks. Estimates of HR from the FC2 were compared to a HR chest strap (Polar) and EE to a stationary metabolic cart (Vyntus CPX). The SWA overestimated overall EE by 0.03 kcal/min−1 and was statistically equivalent to the criterion measure, with a mean absolute percentage error (MAPE) of 29%. In contrast, the FC2 was not equivalent overall (MAPE = 44%). In household tasks, MAPE values of 93% and 83% were observed for the FC2 and SWA, respectively. The FC2 HR estimates were equivalent to the criterion measure overall. The SWA is more accurate than the commercial-grade FC2. Neither device is consistently accurate across the range of activities used in this study. The HR data obtained from the FC2 is more accurate than its EE estimates and future research may focus more on this variable.

An Open-Source Monitor-Independent Movement Summary for Accelerometer Data Processing

Background

Physical behavior researchers using motion sensors often use acceleration summaries to visualize, clean, and interpret data. Such output is dependent on device specifications (e.g., dynamic range, sampling rate) and/or are proprietary, which invalidate cross-study comparison of findings when using different devices. This limits flexibility in selecting devices to measure physical activity, sedentary behavior, and sleep.

Purpose

Develop an open-source, universal acceleration summary metric that accounts for discrepancies in raw data among research and consumer devices.

Methods

We used signal processing techniques to generate a Monitor-Independent Movement Summary unit (MIMS-unit) optimized to capture normal human motion. Methodological steps included raw signal harmonization to eliminate inter-device variability (e.g., dynamic g-range, sampling rate), bandpass filtering (0.2-5.0 Hz) to eliminate non-human movement, and signal aggregation to reduce data to simplify visualization and summarization. We examined the consistency of MIMS-units using orbital shaker testing on eight accelerometers with varying dynamic range (±2 to ±8 g) and sampling rates (20-100 Hz), and human data (N = 60) from an ActiGraph GT9X.

Results

During shaker testing, MIMS-units yielded lower between-device coefficient of variations than proprietary ActiGraph and ENMO acceleration summaries. Unlike the widely used ActiGraph activity counts, MIMS-units were sensitive in detecting subtle wrist movements during sedentary behaviors.

Conclusions

Open-source MIMS-units may provide a means to summarize high-resolution raw data in a device-independent manner, thereby increasing standardization of data cleaning and analytical procedures to estimate selected attributes of physical behavior across studies.

Estimating energy expenditure from wrist and thigh accelerometry in free-living adults: a doubly labelled water study

BACKGROUND

Many large studies have implemented wrist or thigh accelerometry to capture physical activity, but the accuracy of these measurements to infer activity energy expenditure (AEE) and consequently total energy expenditure (TEE) has not been demonstrated. The purpose of this study was to assess the validity of acceleration intensity at wrist and thigh sites as estimates of AEE and TEE under free-living conditions using a gold-standard criterion.

METHODS

Measurements for 193 UK adults (105 men, 88 women, aged 40-66 years, BMI 20.4-36.6 kg m-2) were collected with triaxial accelerometers worn on the dominant wrist, non-dominant wrist and thigh in free-living conditions for 9-14 days. In a subsample (50 men, 50 women) TEE was simultaneously assessed with doubly labelled water (DLW). AEE was estimated from non-dominant wrist using an established estimation model, and novel models were derived for dominant wrist and thigh in the non-DLW subsample. Agreement with both AEE and TEE from DLW was evaluated by mean bias, root mean squared error (RMSE), and Pearson correlation.

RESULTS

Mean TEE and AEE derived from DLW were 11.6 (2.3) MJ day-1 and 49.8 (16.3) kJ day-1 kg-1. Dominant and non-dominant wrist acceleration were highly correlated in free-living (r = 0.93), but less so with thigh (r = 0.73 and 0.66, respectively). Estimates of AEE were 48.6 (11.8) kJ day-1 kg-1 from dominant wrist, 48.6 (12.3) from non-dominant wrist, and 46.0 (10.1) from thigh; these agreed strongly with AEE (RMSE ~12.2 kJ day-1 kg-1, r ~ 0.71) with small mean biases at the population level (~6%). Only the thigh estimate was statistically significantly different from the criterion. When combining these AEE estimates with estimated REE, agreement was stronger with the criterion (RMSE ~1.0 MJ day-1, r ~ 0.90).

CONCLUSIONS

In UK adults, acceleration measured at either wrist or thigh can be used to estimate population levels of AEE and TEE in free-living conditions with high precision.

Validity of the Polar M430 Activity Monitor in Free-Living Conditions: Validation Study

Background

Accelerometers, often in conjunction with heart rate sensors, are extensively used to track physical activity (PA) in research. Research-grade instruments are often expensive and have limited battery capacity, limited storage, and high participant burden. Consumer-based activity trackers are equipped with similar technology and designed for long-term wear, and can therefore potentially be used in research.

Objective

We aimed to assess the criterion validity of the Polar M430 sport watch, compared with 2 research-grade instruments (ActiGraph and Actiheart), worn on 4 different locations using 1- and 3-axis accelerometers.

Methods

A total of 50 participants wore 2 ActiGraphs (wrist and hip), 2 Actihearts (upper and lower chest position), and 1 Polar M430 sport watch for 1 full day. We compared reported time (minutes) spent in sedentary behavior and in light, moderate, vigorous, and moderate to vigorous PA, step counts, activity energy expenditure, and total energy expenditure between devices. We used Pearson correlations, intraclass correlations, mean absolute percentage errors (MAPEs), and Bland-Altman plots to assess criterion validity.

Results

Pearson correlations between the Polar M430 and all research-grade instruments were moderate or stronger for vigorous PA (r range .59-.76), moderate to vigorous PA (r range .51-.75), steps (r range .85-.87), total energy expenditure (r range .88-.94), and activity energy expenditure (r range .74-.79). Bland-Altman plots showed higher agreement for higher intensities of PA. MAPE was high for most outcomes. Only total energy expenditure measured by the hip-worn ActiGraph and both Actiheart positions had acceptable or close to acceptable errors with MAPEs of 6.94% (ActiGraph, 3 axes), 8.26% (ActiGraph, 1 axis), 14.54% (Actiheart, upper position), and 14.37% (Actiheart, lower position). The wrist-worn ActiGraph had a MAPE of 15.94% for measuring steps. All other outcomes had a MAPE of 22% or higher. For most outcomes, the Polar M430 was most strongly correlated with the hip-worn triaxial ActiGraph, with a moderate or strong Pearson correlation for sedentary behavior (r=.52) and for light (r=.7), moderate (r=.57), vigorous (r=.76), and moderate to vigorous (r=.75) PA. In addition, correlations were strong or very strong for activity energy expenditure (r=.75), steps (r=.85), and total energy expenditure (r=.91).

Conclusions

The Polar M430 can potentially be used as an addition to established research-grade instruments to collect some PA variables over a prolonged period. However, due to the high MAPE of most outcomes, only total energy expenditure can be trusted to provide close to valid results. Depending on the variable, the Polar M430 over- or underreported most metrics, and may therefore be better suited to report changes in PA over time for some outcomes, rather than as an accurate instrument for PA status in a population.

Motion Symmetry Evaluation Using Accelerometers and Energy Distribution

Analysis of motion symmetry constitutes an important area with many applications in engineering, robotics, neurology and biomedicine. This paper presents the use of microelectromechanical sensors (MEMS), including accelerometers and gyrometers, to acquire data via mobile devices so as to monitor physical activities and their irregularities. Special attention is devoted to the analysis of the symmetry of the motion of the body when the same exercises are performed by the right and the left limb. The analyzed data include the motion of the legs on a home exercise bike under different levels of load. The method is based on signal analysis using the discrete wavelet transform and the evaluation of signal segment features such as the relative energy at selected decomposition levels. The subsequent classification of the evaluated features is performed by k-nearest neighbours, a Bayesian approach, a support vector machine, and neural networks. The highest average classification accuracy attained is 91.0% and the lowest mean cross-validation error is 0.091, resulting from the use of a neural network. This paper presents the advantages of the use of simple sensors, their combination and intelligent data processing for the numerical evaluation of motion features in the rehabilitation and monitoring of physical activities.

A Dual-Accelerometer System for Classifying Physical Activity in Children and Adults

INTRODUCTION

Accurately monitoring 24-h movement behaviors is a vital step for progressing the time-use epidemiology field. Past accelerometer-based measurement protocols are either hindered by lack of wear time compliance, or the inability to accurately discern activities and postures. Recent work has indicated that skin-attached dual-accelerometers exhibit excellent 24-h uninterrupted wear time compliance. This study extends this work by validating this system for classifying various physical activities and sedentary behaviors in children and adults.

METHODS

Seventy-five participants (42 children) were equipped with two Axivity AX3 accelerometers; one attached to their thigh, and one to their lower back. Ten activity trials (e.g., sitting, standing, lying, walking, running) were performed while under direct observation in a lab setting. Various time- and frequency-domain features were computed from raw accelerometer data, which were then used to train a random forest machine learning classifier. Model performance was evaluated using leave-one-out cross-validation. The efficacy of the dual-sensor protocol (relative to single sensors) was evaluated by repeating the modeling process with each sensor individually.

RESULTS

Machine learning models were able to differentiate between six distinct activity classes with exceptionally high accuracy in both adults (99.1%) and children (97.3%). When a single thigh or back accelerometer was used, there was a pronounced drop in accuracy for nonambulatory activities (up to a 26.4% decline). When examining the features used for model training, those that took the orientation of both sensors into account concurrently were more important predictors.

CONCLUSIONS

When previous wear time compliance results are taken together with our findings, it represents a promising step forward for monitoring and understanding 24-h time-use behaviors. The next step will be to examine the generalizability of these findings in a free-living setting.

An Internet of Things Based Bed-Egress Alerting Paradigm Using Wearable Sensors in Elderly Care Environment

The lack of healthcare staff and increasing proportions of elderly population is alarming. The traditional means to look after elderly has resulted in 255,000 reported falls (only within UK). This not only resulted in extensive aftercare needs and surgeries (summing up to £4.4 billion) but also in added suffering and increased mortality. In such circumstances, the technology can greatly assist by offering automated solutions for the problem at hand. The proposed work offers an Internet of things (IoT) based patient bed-exit monitoring system in clinical settings, capable of generating a timely response to alert the healthcare workers and elderly by analyzing the wireless data streams, acquired through wearable sensors. This work analyzes two different datasets obtained from divergent families of sensing technologies, i.e., smartphone-based accelerometer and radio frequency identification (RFID) based accelerometer. The findings of the proposed system show good efficacy in monitoring the bed-exit and discriminate other ambulating activities. Furthermore, the proposed work manages to keep the average end-to-end system delay (i.e., communications of sensed data to Data Sink (DS)/Control Center (CC) + machine-based feature extraction and class identification + feedback communications to a relevant healthcare worker/elderly) below 1 10 th of a second.

Evaluating and Enhancing the Generalization Performance of Machine Learning Models for Physical Activity Intensity Prediction From Raw Acceleration Data

PURPOSE

To evaluate and enhance the generalization performance of machine learning physical activity intensity prediction models developed with raw acceleration data on populations monitored by different activity monitors.

METHOD

Five datasets from four studies, each containing only hip- or wrist-based raw acceleration data (two hip- and three wrist-based) were extracted. The five datasets were then used to develop and validate artificial neural networks (ANN) in three setups to classify activity intensity categories (sedentary behavior, light, and moderate-to-vigorous). To examine generalizability, the ANN models were developed using within dataset (leave-one-subject-out) cross validation, and then cross tested to other datasets with different accelerometers. To enhance the models' generalizability, a combination of four of the five datasets was used for training and the fifth dataset for validation. Finally, all the five datasets were merged to develop a single model that is generalizable across the datasets (50% of the subjects from each dataset for training, the remaining for validation).

RESULTS

The datasets showed high performance in within dataset cross validation (accuracy 71.9-95.4%, Kappa K = 0.63-0.94). The performance of the within dataset validated models decreased when applied to datasets with different accelerometers (41.2-59.9%, K = 0.21-0.48). The trained models on merged datasets consisting hip and wrist data predicted the left-out dataset with acceptable performance (65.9-83.7%, K = 0.61-0.79). The model trained with all five datasets performed with acceptable performance across the datasets (80.4-90.7%, K = 0.68-0.89).

CONCLUSIONS

Integrating heterogeneous datasets in training sets seems a viable approach for enhancing the generalization performance of the models. Instead, within dataset validation is not sufficient to understand the models' performance on other populations with different accelerometers.

On Placement, Location and Orientation of Wrist-Worn Tri-Axial Accelerometers during Free-Living Measurements

Wearable accelerometers have recently become a standalone tool for the objective assessment of physical activity (PA). In free-living studies, accelerometers are placed by protocol on a pre-defined body location (e.g., non-dominant wrist). However, the protocol is not always followed, e.g., the sensor can be moved between wrists or reattached in a different orientation. Such protocol violations often result in PA miscalculation. We propose an approach, PLOE ("Placement, Location and Orientation Evaluation method"), to determine the sensor position using statistical features from the raw accelerometer measurements. We compare the estimated position with the study protocol and identify discrepancies. We apply PLOE to the measurements collected from 45 older adults who wore ActiGraph GT3X+ accelerometers on the left and right wrist for seven days. We found that 15.6% of participants who wore accelerometers violated the protocol for one or more days. The sensors were worn on the wrong hand during 6.9% of the days of simultaneous wearing of devices. During the periods of discrepancies, the daily PA was miscalculated by more than 20%. Our findings show that correct placement of the device has a significant effect on the PA estimates. These results demonstrate a need for the evaluation of sensor position.

Identification of earlier predictors of pregnancy complications through wearable technologies in a Brazilian multicentre cohort: Maternal Actigraphy Exploratory Study I (MAES-I) study protocol

INTRODUCTION

Non-invasive tools capable of identifying predictors of maternal complications would be a step forward for improving maternal and perinatal health. There is an association between modification in physical activity (PA) and sleep-wake patterns and the occurrence of inflammatory, metabolic, pathological conditions related to chronic diseases. The actigraphy device is validated to estimate PA and sleep-wake patterns among pregnant women. In order to extend the window of opportunity to prevent, diagnose and treat specific maternal conditions, would it be possible to use actigraphy data to identify risk factors for the development of adverse maternal outcomes during pregnancy?

METHODS AND ANALYSIS

A cohort will be held in five centres from the Brazilian Network for Studies on Reproductive and Perinatal Health. Maternal Actigraphy Exploratory Study I (MAES-I) will enrol 400 low-risk nulliparous women who will wear the actigraphy device on their wrists day and night (24 hours/day) uninterruptedly from 19 to 21 weeks until childbirth. Changes in PA and sleep-wake patterns will be analysed throughout pregnancy, considering ranges in gestational age in women with and without maternal complications such as pre-eclampsia, preterm birth (spontaneous or provider-initiated), gestational diabetes, maternal haemorrhage during pregnancy, in addition to perinatal outcomes. The plan is to design a predictive model using actigraphy data for screening pregnant women at risk of developing specific adverse maternal and perinatal outcomes.

ETHICS AND DISSEMINATION

MAES-I has been reviewed and approved by each institutional review board and also by the National Council for Ethics in Research. Detailed information about the study is provided in the Brazilian Cohort website (www.medscinet.com/samba) and findings will be published in the scientific literature and institutional webpages.

Improving Hip-Worn Accelerometer Estimates of Sitting Using Machine Learning Methods

PURPOSE

This study aimed to improve estimates of sitting time from hip-worn accelerometers used in large cohort studies by using machine learning methods developed on free-living activPAL data.

METHODS

Thirty breast cancer survivors concurrently wore a hip-worn accelerometer and a thigh-worn activPAL for 7 d. A random forest classifier, trained on the activPAL data, was used to detect sitting, standing, and sit-stand transitions in 5-s windows in the hip-worn accelerometer. The classifier estimates were compared with the standard accelerometer cut point, and significant differences across different bout lengths were investigated using mixed-effect models.

RESULTS

Overall, the algorithm predicted the postures with moderate accuracy (stepping, 77%; standing, 63%; sitting, 67%; sit-to-stand, 52%; and stand-to-sit, 51%). Daily level analyses indicated that errors in transition estimates were only occurring during sitting bouts of 2 min or less. The standard cut point was significantly different from the activPAL across all bout lengths, overestimating short bouts and underestimating long bouts.

CONCLUSIONS

This is among the first algorithms for sitting and standing for hip-worn accelerometer data to be trained from entirely free-living activPAL data. The new algorithm detected prolonged sitting, which has been shown to be the most detrimental to health. Further validation and training in larger cohorts is warranted.

The Influence of Feature Representation of Text on the Performance of Document Classification

In this paper we perform a comparative analysis of three models for a feature representation of text documents in the context of document classification. In particular, we consider the most often used family of bag-of-words models, the recently proposed continuous space models word2vec and doc2vec, and the model based on the representation of text documents as language networks. While the bag-of-word models have been extensively used for the document classification task, the performance of the other two models for the same task have not been well understood. This is especially true for the network-based models that have been rarely considered for the representation of text documents for classification. In this study, we measure the performance of the document classifiers trained using the method of random forests for features generated with the three models and their variants. Multi-objective rankings are proposed as the framework for multi-criteria comparative analysis of the results. Finally, the results of the empirical comparison show that the commonly used bag-of-words model has a performance comparable to the one obtained by the emerging continuous-space model of doc2vec. In particular, the low-dimensional variants of doc2vec generating up to 75 features are among the top-performing document representation models. The results finally point out that doc2vec shows a superior performance in the tasks of classifying large documents.

Concurrent validity of the Fitbit for assessing sedentary behavior and moderate-to-vigorous physical activity

Background

Recent advances in sensor technologies have promoted the use of consumer-based accelerometers such as Fitbit Flex in epidemiological and clinical research; however, the validity of the Fitbit Flex in measuring sedentary behavior (SED) and physical activity (PA) has not been fully determined against previously validated research-grade accelerometers such as ActiGraph GT3X+. Therefore, the purpose of this study was to examine the concurrent validity of the Fitbit Flex against ActiGraph GT3X+ in a free-living condition.

Methods

A total of 65 participants (age: M = 42, SD = 14 years, female: 72%) each wore a Fitbit Flex and GT3X+ for seven consecutive days. After excluding sleep and non-wear time, time spent (min/day) in SED and moderate-to-vigorous PA (MVPA) were estimated using various cut-points for GT3X+ and brand-specific algorithms for Fitbit, respectively. Repeated measures one-way ANOVA and mean absolute percent errors (MAPE) served to examine differences and measurement errors in SED and MVPA estimates between Fitbit Flex and GT3X+, respectively. Pearson and Spearman correlations and Bland-Altman (BA) plots were used to evaluate the association and potential systematic bias between Fitbit Flex and GT3X+. PROC MIXED procedure in SAS was used to examine the equivalence (i.e., the 90% confidence interval with ±10% equivalence zone) between the devices.

Results

Fitbit Flex produced similar SED and low MAPE (mean difference [MD] = 37 min/day, P = .21, MAPE = 6.8%), but significantly higher MVPA and relatively large MAPE (MD = 59–77 min/day, P < .0001, MAPE = 56.6–74.3%) compared with the estimates from GT3X+ using three different cut-points. The correlations between Fitbit Flex and GT3X+ were consistently higher for SED (r = 0.90, ρ = 0.86, P < .01), but weaker for MVPA (r = 0.65–0.76, ρ = 0.69–0.79, P < .01). BA plots revealed that there is no apparent bias in estimating SED.

Conclusion

In comparison with the GT3X+ accelerometer, the Fitbit Flex provided comparatively accurate estimates of SED, but the Fitbit Flex overestimated MVPA under free-living conditions. Future investigations using the Fitbit Flex should be aware of present findings.

Calibration and validation of accelerometer-based activity monitors: A systematic review of machine-learning approaches

BACKGROUND

Objective measures using accelerometer-based activity monitors have been extensively used in physical activity (PA) and sedentary behavior (SB) research. To measure PA and SB precisely, the field is shifting towards machine learning-based (ML) approaches for calibration and validation of accelerometer-based activity monitors. Nevertheless, various parameters regarding the use and development of ML-based models, including data type (raw acceleration data versus activity counts), sampling frequency, window size, input features, ML technique, accelerometer placement, and free-living settings, affect the predictive ability of ML-based models. The effects of these parameters on ML-based models have remained elusive, and will be systematically reviewed here. The open challenges were identified and recommendations are made for future studies and directions.

METHOD

We conducted a systematic search of PubMed and Scopus databases to identify studies published before July 2017 that used ML-based techniques for calibration and validation of accelerometer-based activity monitors. Additional articles were manually identified from references in the identified articles.

RESULTS

A total of 62 studies were eligible to be included in the review, comprising 48 studies that calibrated and validated ML-based models for predicting the type and intensity of activities, and 22 studies for predicting activity energy expenditure.

CONCLUSIONS

It appears that various ML-based techniques together with raw acceleration data sampled at 20-30 Hz provide the opportunity of predicting the type and intensity of activities, as well as activity energy expenditure with comparable overall predictive accuracies regardless of accelerometer placement. However, the high predictive accuracy of laboratory-calibrated models is not reproducible in free-living settings, due to transitive and unseen activities together with differences in acceleration signals.

Compliance with wrist-worn accelerometers in primiparous early postpartum women

Background

There are few studies that objectively assess physical activity using accelerometry in postpartum women and none that do so before 3 months postpartum. It is not known whether accelerometry can be successfully used in the early postpartum period and thus benefit studies designed to assess the health benefits and risks of early physical activity. Wear compliance may be substantially lower several weeks after childbirth, given the overwhelming nature of the early postpartum period, particularly for first time mothers. The aims of this study were to 1) describe the methods used to facilitate protocol compliance of wrist-worn accelerometry, 2) describe device usage and wear time in early postpartum primiparous women and 3) to place the compliance characteristics of early postpartum primiparous women in our study in context with that of other studies of postpartum women and standards published by large, physical activity surveillance studies.

Methods

Participants were primiparous women who were enrolled at 3^rd trimester in a larger ongoing prospective cohort study, delivered vaginally, and lived in a 60 mile radius of the research site. The parent study was designed to evaluate the effects of early physical activity on pelvic floor health. Participants wore a wrist accelerometer (ActiGraph™ GT3XLink device) over two 7-day periods, 2-3 weeks and 5-6 weeks postpartum. We developed a protocol based on best practices to enhance compliance in this population. The Choi (2011) algorithm was used to determine wear time.

Results

Of all participants, 82.6% (166 of 201 eligible) and 70.1% (141 of 201 eligible) at 2-3 and 5-6 weeks, respectively, received and wore a functional device in the correct study time-frame for at least 7 days. Of participants that received a functional device, 94.3% (166/176) and 86.5% (141/163) wore the device for at least 7 days, with mean wear times of 1348.0 (135.8) minutes/day and 1313.5 (152) minutes/day, respectively. At 2-3 weeks, 96.1% and 90.4% met the NHANES and Whitehall II Study wear standards, respectively, while at 5-6 weeks, 93.9% and 84.1% did so.

Conclusion

Despite challenges in conducting physical activity research in postpartum women, adherence to wrist-worn accelerometry is high with this protocol.

Coaching and Prompting for Remote Physical Activity Promotion: Study Protocol of a Three-Arm Randomized Controlled Trial (Movingcall)

Background. Physical inactivity is currently seen as one of the biggest global public health issue. Remote physical activity (PA) promotion programs are expected to be effective if they are individually tailored and include behavior change techniques, personal coaching, and regular prompting. However, it is still not fully understood which intervention components are most effective. This paper describes the rationale and design of a study on an individually tailored remote PA promotion program comparing the efficacy of coaching and prompting with a single written advice. Methods. In total, 288 adults (age 20 to 65 years) were randomly assigned to three different intervention arms of a 6-month-long PA promotion program. A minimal intervention group received a single written PA recommendation. The two remaining groups either received telephone coaching sessions (n = 12 calls) with or without additional short message service (SMS) prompting (n = 48 SMSs for each participant). Data assessment took place at baseline, at the end of the intervention, and after a six-month follow-up-period. The primary outcome of the study was self-reported PA. Objectively assessed PA, psychosocial determinants of PA, well-being, body mass index (BMI), and adherence were assessed as secondary outcomes. Conclusion. Findings of this three-arm study will provide insight into the short and long-term effects of coaching and prompting for PA promotion.

Adapted Sojourn Models to Estimate Activity Intensity in Youth: A Suite of Tools

The challenges of using physical activity data from accelerometers have been compounded with the recent focus on wrist-worn monitors and raw acceleration (as opposed to activity counts).

PURPOSE

This study developed and systematically evaluated a suite of new accelerometer processing models for youth.

METHODS

Four adaptations of the Sojourn method were developed using data from a laboratory-based experiment in which youth (N = 54) performed structured activity routines. The adaptations corresponded to all possible pairings of hip or wrist attachment with activity counts (AC) or raw acceleration (RA), and they estimated time in sedentary behavior, light activity, and moderate-to-vigorous physical activity. Criterion validity was assessed using direct observation in an independent free-living sample (N = 27). Monitors were worn on both wrists to evaluate the effect of handedness on accuracy, and status quo methods for each configuration were also evaluated as benchmarks for comparison. Tests of classification accuracy (percent accuracy, κ statistics, and sensitivity and specificity) were used to summarize utility.

RESULTS

In the development sample, percent accuracy ranged from 68.5% (wrist-worn AC, κ = 0.42) to 71.6% (hip-worn RA, κ = 0.50). Accuracy was lower in the free-living evaluation, with values ranging from 49.3% (hip-worn RA, κ = 0.25) to 56.7% (hip-worn AC, κ = 0.36). Collectively, the suite predicted moderate-to-vigorous physical activity well, with the models averaging 96.5% sensitivity and 67.5% specificity. However, in terms of overall accuracy, the new models performed similarly to the status quo methods. There were no meaningful differences in performance at either wrist.

CONCLUSIONS

The new models offered minimal improvements over existing methods, but a major advantage is that further tuning of the models is possible with continued research.

Evaluating physiological signal salience for estimating metabolic energy cost from wearable sensors

Body-in-the-loop optimization algorithms have the capability to automatically tune the parameters of robotic prostheses and exoskeletons to minimize the metabolic energy expenditure of the user. However, current body-in-the-loop algorithms rely on indirect calorimetry to obtain measurements of energy cost, which are noisy, sparsely sampled, time-delayed, and require wearing a respiratory mask. To improve these algorithms, the goal of this work is to predict a user's steady-state energy cost quickly and accurately using physiological signals obtained from portable, wearable sensors. In this paper, we quantified physiological signal salience to discover which signals, or groups of signals, have the best predictive capability when estimating metabolic energy cost. We collected data from 10 healthy individuals performing 6 activities (walking, incline walking, backward walking, running, cycling, and stair climbing) at various speeds or intensities. Subjects wore a suite of physiological sensors that measured breath frequency and volume, limb accelerations, lower limb EMG, heart rate, electrodermal activity, skin temperature, and oxygen saturation; indirect calorimetry was used to establish the 'ground truth' energy cost for each activity. Evaluating Pearson's correlation coefficients and single and multiple linear regression models with cross validation (leave-one- subject-out and leave-one- task-out), we found that 1) filtering the accelerations and EMG signals improved their predictive power, 2) global signals (e.g., heart rate, electrodermal activity) were more sensitive to unknown subjects than tasks, while local signals (e.g., accelerations) were more sensitive to unknown tasks than subjects, and 3) good predictive performance was obtained combining a small number of signals (4-5) from multiple sensor modalities. NEW & NOTEWORTHY In this paper, we systematically compare a large set of physiological signals collected from portable sensors and determine which sensor signals contain the most salient information for predicting steady-state metabolic energy cost, robust to unknown subjects or tasks. This information, together with the comprehensive data set that is published in conjunction with this paper, will enable researchers and clinicians across many fields to develop novel algorithms to predict energy cost from wearable sensors.

Segmenting accelerometer data from daily life with unsupervised machine learning

Purpose

Accelerometers are increasingly used to obtain valuable descriptors of physical activity for health research. The cut-points approach to segment accelerometer data is widely used in physical activity research but requires resource expensive calibration studies and does not make it easy to explore the information that can be gained for a variety of raw data metrics. To address these limitations, we present a data-driven approach for segmenting and clustering the accelerometer data using unsupervised machine learning.

Methods

The data used came from five hundred fourteen-year-old participants from the Millennium cohort study who wore an accelerometer (GENEActiv) on their wrist on one weekday and one weekend day. A Hidden Semi-Markov Model (HSMM), configured to identify a maximum of ten behavioral states from five second averaged acceleration with and without addition of x, y, and z-angles, was used for segmenting and clustering of the data. A cut-points approach was used as comparison.

Results

Time spent in behavioral states with or without angle metrics constituted eight and five principal components to reach 95% explained variance, respectively; in comparison four components were identified with the cut-points approach. In the HSMM with acceleration and angle as input, the distributions for acceleration in the states showed similar groupings as the cut-points categories, while more variety was seen in the distribution of angles.

Conclusion

Our unsupervised classification approach learns a construct of human behavior based on the data it observes, without the need for resource expensive calibration studies, has the ability to combine multiple data metrics, and offers a higher dimensional description of physical behavior. States are interpretable from the distributions of observations and by their duration.

P. Abdul Majeed A, Taha Z, Chang SW, Ab. Nasir AF, Abdullah MR. A machine learning approach of predicting high potential archers by means of physical fitness indicators

k-nearest neighbour (k-NN) has been shown to be an effective learning algorithm for classification and prediction. However, the application of k-NN for prediction and classification in specific sport is still in its infancy. The present study classified and predicted high and low potential archers from a set of physical fitness variables trained on a variation of k-NN algorithms and logistic regression. 50 youth archers with the mean age and standard deviation of (17.0 ± 0.56) years drawn from various archery programmes completed a one end archery shooting score test. Standard fitness measurements of the handgrip, vertical jump, standing broad jump, static balance, upper muscle strength and the core muscle strength were conducted. Multiple linear regression was utilised to ascertain the significant variables that affect the shooting score. It was demonstrated from the analysis that core muscle strength and vertical jump were statistically significant. Hierarchical agglomerative cluster analysis (HACA) was used to cluster the archers based on the significant variables identified. k-NN model variations, i.e., fine, medium, coarse, cosine, cubic and weighted functions as well as logistic regression, were trained based on the significant performance variables. The HACA clustered the archers into high potential archers (HPA) and low potential archers (LPA). The weighted k-NN outperformed all the tested models at itdemonstrated reasonably good classification on the evaluated indicators with an accuracy of 82.5 ± 4.75% for the prediction of the HPA and the LPA. Moreover, the performance of the classifiers was further investigated against fresh data, which also indicates the efficacy of the weighted k-NN model. These findings could be valuable to coaches and sports managers to recognise high potential archers from a combination of the selected few physical fitness performance indicators identified which would subsequently save cost, time and energy for a talent identification programme.

Sensor-enabled Activity Class Recognition in Preschoolers: Hip versus Wrist Data

PURPOSE

Pattern recognition approaches to accelerometer data processing have emerged as viable alternatives to cut-point methods. However, few studies have explored the validity of pattern recognition approaches in preschoolers, and none have compared supervised learning algorithms trained on hip and wrist data. Purpose of this study was to develop, test, and compare activity class recognition algorithms trained on hip, wrist, and combined hip and wrist accelerometer data in preschoolers.

METHODS

Eleven children 3-6 yr of age (mean age, 4.8 ± 0.9 yr) completed 12 developmentally appropriate physical activity (PA) trials while wearing an ActiGraph GT3X+ accelerometer on the right hip and nondominant wrist. PA trials were categorized as sedentary, light activity games, moderate-to-vigorous games, walking, and running. Random forest (RF) and support vector machine (SVM) classifiers were trained using time and frequency domain features from the vector magnitude of the raw signal. Features were extracted from 15-s nonoverlapping windows. Classifier performance was evaluated using leave-one-out cross-validation.

RESULTS

Cross-validation accuracy for the hip, wrist, and combined hip and wrist RF models was 0.80 (95% confidence interval (CI), 0.79-0.82), 0.78 (95% CI, 0.77-0.80), and 0.82 (95% CI, 0.80-0.83), respectively. Accuracy for hip, wrist, and combined hip and wrist SVM models was 0.81 (95% CI, 0.80-0.83), 0.80 (95% CI, 0.79-0.80), and 0.85 (95% CI, 0.84-0.86), respectively. Recognition accuracy was consistently excellent for sedentary (>90%); moderate for light activity games, moderate-to-vigorous games, and running (69%-79%); and modest for walking (61%-71%).

CONCLUSIONS

Machine learning algorithms such as RF and SVM are useful for predicting PA class from accelerometer data collected in preschool children. Although classifiers trained on hip or wrist data provided acceptable recognition accuracy, the combination of hip and wrist accelerometer delivered better performance.

Wrist accelerometer shape feature derivation methods for assessing activities of daily living

Background

There has been an increasing interest in understanding the usefulness of wrist-based accelerometer data for physical activity (PA) assessment due to the ease of use and higher user compliance than other body placements. PA assessment studies have relied on machine learning methods which take accelerometer data in forms of variables, or feature vectors.

Methods

In this work, we introduce automated shape feature derivation methods to transform epochs of accelerometer data into feature vectors. As the first step, recurring patterns in the collected data are identified and placed in a codebook. Similarities between epochs of accelerometer data and codebook’s patterns are the basis of feature calculations. In this paper, we demonstrate supervised and unsupervised approaches to learn codebooks. We evaluated these methods and compared them with the standard statistical measures for PA assessment. The experiments were performed on 146 participants who wore an ActiGraph GT3X+ accelerometer on the right wrist and performed 33 activities of daily living.

Results

Our evaluations show that the shape feature derivation methods were able to perform comparably with the standard wrist model (F1-score: 0.89) for identifying sedentary PAs (F1-scores of 0.86 and 0.85 for supervised and unsupervised methods, respectively). This was also observed for identifying locomotion activities (F1-scores: 0.87, 0.83, and 0.81 for the standard wrist, supervised, unsupervised models, respectively). All the wrist models were able to estimate energy expenditure required for PAs with low error (rMSE: 0.90, 0.93, and 0.90 for the standard wrist, supervised, and unsupervised models, respectively).

Conclusion

The automated shape feature derivation methods offer insights into the performed activities by providing a summary of repeating patterns in the accelerometer data. Furthermore, they could be used as efficient alternatives (or additions) for manually engineered features, especially important for cases where the latter fail to provide sufficient information to machine learning methods for PA assessment.

Accelerometer Data Collection and Processing Criteria to Assess Physical Activity and Other Outcomes: A Systematic Review and Practical Considerations

BACKGROUND

Accelerometers are widely used to measure sedentary time, physical activity, physical activity energy expenditure (PAEE), and sleep-related behaviors, with the ActiGraph being the most frequently used brand by researchers. However, data collection and processing criteria have evolved in a myriad of ways out of the need to answer unique research questions; as a result there is no consensus.

OBJECTIVES

The purpose of this review was to: (1) compile and classify existing studies assessing sedentary time, physical activity, energy expenditure, or sleep using the ActiGraph GT3X/+ through data collection and processing criteria to improve data comparability and (2) review data collection and processing criteria when using GT3X/+ and provide age-specific practical considerations based on the validation/calibration studies identified.

METHODS

Two independent researchers conducted the search in PubMed and Web of Science. We included all original studies in which the GT3X/+ was used in laboratory, controlled, or free-living conditions published from 1 January 2010 to the 31 December 2015.

RESULTS

The present systematic review provides key information about the following data collection and processing criteria: placement, sampling frequency, filter, epoch length, non-wear-time, what constitutes a valid day and a valid week, cut-points for sedentary time and physical activity intensity classification, and algorithms to estimate PAEE and sleep-related behaviors. The information is organized by age group, since criteria are usually age-specific.

CONCLUSION

This review will help researchers and practitioners to make better decisions before (i.e., device placement and sampling frequency) and after (i.e., data processing criteria) data collection using the GT3X/+ accelerometer, in order to obtain more valid and comparable data.

PROSPERO REGISTRATION NUMBER

CRD42016039991.

Measuring Sedentary Behavior by Means of Muscular Activity and Accelerometry

Sedentary Behavior (SB) is among the most frequent human behaviors and is associated with a plethora of serious chronic lifestyle diseases as well as premature death. Office workers in particular are at an increased risk due to their extensive amounts of occupational SB. However, we still lack an objective method to measure SB consistent with its definition. We have therefore developed a new measurement system based on muscular activity and accelerometry. The primary aim of the present study was to calibrate the new-developed 8-CH-EMG+ for measuring occupational SB against an indirect calorimeter during typical desk-based office work activities. In total, 25 volunteers performed nine office tasks at three typical workplaces. Minute-by-minute posture and activity classification was performed using subsequent decision trees developed with artificial intelligence data processing techniques. The 8-CH-EMG+ successfully identified all sitting episodes (AUC = 1.0). Furthermore, depending on the number of electromyography channels included, the device has a sensitivity of 83–98% and 74–98% to detect SB and active sitting (AUC = 0.85–0.91). The 8-CH-EMG+ advances the field of objective SB measurements by combining accelerometry with muscular activity. Future field studies should consider the use of EMG sensors to record SB in line with its definition.

Defining and Predicting Pain Volatility in Users of the Manage My Pain App: Analysis Using Data Mining and Machine Learning Methods

Background

Measuring and predicting pain volatility (fluctuation or variability in pain scores over time) can help improve pain management. Perceptions of pain and its consequent disabling effects are often heightened under the conditions of greater uncertainty and unpredictability associated with pain volatility.

Objective

This study aimed to use data mining and machine learning methods to (1) define a new measure of pain volatility and (2) predict future pain volatility levels from users of the pain management app, Manage My Pain, based on demographic, clinical, and app use features.

Methods

Pain volatility was defined as the mean of absolute changes between 2 consecutive self-reported pain severity scores within the observation periods. The k-means clustering algorithm was applied to users’ pain volatility scores at the first and sixth month of app use to establish a threshold discriminating low from high volatility classes. Subsequently, we extracted 130 demographic, clinical, and app usage features from the first month of app use to predict these 2 volatility classes at the sixth month of app use. Prediction models were developed using 4 methods: (1) logistic regression with ridge estimators; (2) logistic regression with Least Absolute Shrinkage and Selection Operator; (3) Random Forests; and (4) Support Vector Machines. Overall prediction accuracy and accuracy for both classes were calculated to compare the performance of the prediction models. Training and testing were conducted using 5-fold cross validation. A class imbalance issue was addressed using a random subsampling of the training dataset. Users with at least five pain records in both the predictor and outcome periods (N=782 users) are included in the analysis.

Results

k-means clustering algorithm was applied to pain volatility scores to establish a threshold of 1.6 to differentiate between low and high volatility classes. After validating the threshold using random subsamples, 2 classes were created: low volatility (n=611) and high volatility (n=171). In this class-imbalanced dataset, all 4 prediction models achieved 78.1% (611/782) to 79.0% (618/782) in overall accuracy. However, all models have a prediction accuracy of less than 18.7% (32/171) for the high volatility class. After addressing the class imbalance issue using random subsampling, results improved across all models for the high volatility class to greater than 59.6% (102/171). The prediction model based on Random Forests performs the best as it consistently achieves approximately 70% accuracy for both classes across 3 random subsamples.

Conclusions

We propose a novel method for measuring pain volatility. Cluster analysis was applied to divide users into subsets of low and high volatility classes. These classes were then predicted at the sixth month of app use with an acceptable degree of accuracy using machine learning methods based on the features extracted from demographic, clinical, and app use information from the first month.

Dietary Intake and Physical Activity Assessment: Current Tools, Techniques, and Technologies for Use in Adult Populations

Accurate assessment of dietary intake and physical activity is a vital component for quality research in public health, nutrition, and exercise science. However, accurate and consistent methodology for the assessment of these components remains a major challenge. Classic methods use self-report to capture dietary intake and physical activity in healthy adult populations. However, these tools, such as questionnaires or food and activity records and recalls, have been shown to underestimate energy intake and expenditure as compared with direct measures like doubly labeled water. This paper summarizes recent technological advancements, such as remote sensing devices, digital photography, and multisensor devices, which have the potential to improve the assessment of dietary intake and physical activity in free-living adults. This review will provide researchers with emerging evidence in support of these technologies, as well as a quick reference for selecting the "right-sized" assessment method based on study design, target population, outcome variables of interest, and economic and time considerations.

THEME INFORMATION

This article is part of a theme issue entitled Innovative Tools for Assessing Diet and Physical Activity for Health Promotion, which is sponsored by the North American branch of the International Life Sciences Institute.

Physical activity characterization: does one site fit all?

BACKGROUND

It is evident that a growing number of studies advocate a wrist-worn accelerometer for the assessment of patterns of physical activity a priori, yet the veracity of this site rather than any other body-mounted location for its accuracy in classifying activity is hitherto unexplored.

OBJECTIVE

The objective of this review was to identify the relative accuracy with which physical activities can be classified according to accelerometer site and analytical technique.

METHODS

A search of electronic databases was conducted using Web of Science, PubMed and Google Scholar. This review included studies written in the English language, published between database inception and December 2017, which characterized physical activities using a single accelerometer and reported the accuracy of the technique.

RESULTS

A total of 118 articles were initially retrieved. After duplicates were removed and the remaining articles screened, 32 full-text articles were reviewed, resulting in the inclusion of 19 articles that met the eligibility criteria.

CONCLUSION

There is no 'one site fits all' approach to the selection of accelerometer site location or analytical technique. Research design and focus should always inform the most suitable location of attachment, and should be driven by the type of activity being characterized.

Fusion of Heart Rate, Respiration and Motion Measurements from a Wearable Sensor System to Enhance Energy Expenditure Estimation

This paper presents a new method that integrates heart rate, respiration, and motion information obtained from a wearable sensor system to estimate energy expenditure. The system measures electrocardiography, impedance pneumography, and acceleration from upper and lower limbs. A multilayer perceptron neural network model was developed, evaluated, and compared to two existing methods, with data from 11 subjects (mean age, 27 years, range, 21⁻65 years) who performed a 3-h protocol including submaximal tests, simulated work tasks, and periods of rest. Oxygen uptake was measured with an indirect calorimeter as a reference, with a time resolution of 15 s. When compared to the reference, the new model showed a lower mean absolute error (MAE = 1.65 mL/kg/min, R² = 0.92) than the two existing methods, i.e., the flex-HR method (MAE = 2.83 mL/kg/min, R² = 0.75), which uses only heart rate, and arm-leg HR+M method (MAE = 2.12 mL/kg/min, R² = 0.86), which uses heart rate and motion information. As indicated, this new model may, in combination with a wearable system, be useful in occupational and general health applications.

Application and Validation of Activity Monitors' Epoch Lengths and Placement Sites for Physical Activity Assessment in Exergaming

We assessed the agreement of two ActiGraph activity monitors (wGT3X vs. GT9X) placed at the hip and the wrist and determined an appropriate epoch length for physical activity levels in an exergaming setting. Forty-seven young adults played a 30-min exergame while wearing wGT3X and GT9X on both hip and wrist placement sites and a heart rate sensor below the chest. Intraclass correlation coefficient indicated that intermonitor agreement in steps and activity counts was excellent on the hip and good on the wrist. Bland-Altman plots indicated good intermonitor agreement in the steps and activity counts on both placement sites but a significant intermonitor difference was detected in steps on the wrist. Time spent in sedentary and physical activity intensity levels varied across six epoch lengths and depended on the placement sites, whereas time spent from a 1-s epoch of the hip-worn monitors most accurately matched the relative exercise intensity by heart rate. Hip placement site was associated with better step-counting accuracy for both activity monitors and more valid estimation of physical activity levels. A 1-s epoch was the most appropriate epoch length to detect short bursts of intense physical activity and may be the best choice for data processing and analysis in exergaming studies examining intermittent physical activities.

Application of feature selection and regression models for chlorophyll-a prediction in a shallow lake

As a representative index of the algal bloom, the concentration of chlorophyll-a (Chl-a) is a key parameter of concern for environmental managers. The relationships between environmental variables and Chl-a are complex and difficult to establish. Two machine learning methods, including support vector machine for regression (SVR) and random forest (RF), were used in this study to predict Chl-a concentration based on multiple variables. To improve the model accuracy and reduce the input number, two feature selection methods, including minimum redundancy and maximum relevance method (mRMR) and RF, were integrated with regression models. The results showed that the RF model had a higher predictive ability than the SVR model. Furthermore, the less computational time cost and unnecessary prior data transformation also indicated a better applicability of the RF model. The comparison between ensemble models of mRMR-RF and RF-RF showed that the RF-RF yielded a better performance with fewer variables. Seven variables selected from the candidate predictors could interpret most information, and their potential implications to Chl-a were discussed based on the level of importance. Overall, the RF-RF ensemble model can be considered as a useful approach to determine the significant stressors and achieve satisfactory prediction of Chl-a concentration.

A random forest and simulation approach for scheduling operation rooms: Elective surgery cancelation in a Chinese hospital urology department

Many hospitals encounter surgery cancelations for various reasons. We present a methodology applying data mining and simulation to optimize operating room (OR) scheduling in a urology department in West China Hospital. To the best of our knowledge, this is 1 of the first efforts to seek an optimal schedule solution based on cancelation risk of elective surgeries as well as OR allocation between elective and nonelective surgeries. First, chi-square test and random forest prediction modeling were used to predict potential elective surgeries with high cancelation risk, and the factors, including surgeon, number of days since admission of patient, first surgery or not, etc., that influence elective surgery cancelation were identified. Second, a simulation technology was designed to compare 7 different scheduling strategies. The results demonstrated that for elective surgery, cancelation rate low surgery first outperformed the others and increased the productivity of the ORs from 72% to 83%, while for nonelective surgery performed in a separate OR, there was no improvement because the supply was greater than necessary at present. However, in total, the selected strategies led to 7% higher productivity.

Statistical machine learning of sleep and physical activity phenotypes from sensor data in 96,220 UK Biobank participants

Current public health guidelines on physical activity and sleep duration are limited by a reliance on subjective self-reported evidence. Using data from simple wrist-worn activity monitors, we developed a tailored machine learning model, using balanced random forests with Hidden Markov Models, to reliably detect a number of activity modes. We show that physical activity and sleep behaviours can be classified with 87% accuracy in 159,504 minutes of recorded free-living behaviours from 132 adults. These trained models can be used to infer fine resolution activity patterns at the population scale in 96,220 participants. For example, we find that men spend more time in both low- and high- intensity behaviours, while women spend more time in mixed behaviours. Walking time is highest in spring and sleep time lowest during the summer. This work opens the possibility of future public health guidelines informed by the health consequences associated with specific, objectively measured, physical activity and sleep behaviours.