Physical Activity Recognition Using Posterior-Adapted Class-Based Fusion of Multiaccelerometer Data

Emerging methods for measuring physical activity using accelerometry in children and adolescents with neuromotor disorders: a narrative review

BACKGROUND

Children and adolescents with neuromotor disorders need regular physical activity to maintain optimal health and functional independence throughout their development. To this end, reliable measures of physical activity are integral to both assessing habitual physical activity and testing the efficacy of the many interventions designed to increase physical activity in these children. Wearable accelerometers have been used for children with neuromotor disorders for decades; however, studies most often use disorder-specific cut points to categorize physical activity intensity, which lack generalizability to a free-living environment. No reviews of accelerometer data processing methods have discussed the novel use of machine learning techniques for monitoring physical activity in children with neuromotor disorders.

METHODS

In this narrative review, we discuss traditional measures of physical activity (including questionnaires and objective accelerometry measures), the limitations of standard analysis for accelerometry in this unique population, and the potential benefits of applying machine learning approaches. We also provide recommendations for using machine learning approaches to monitor physical activity.

CONCLUSIONS

While wearable accelerometers provided a much-needed method to quantify physical activity, standard cut point analyses have limitations in children with neuromotor disorders. Machine learning models are a more robust method of analyzing accelerometer data in pediatric neuromotor disorders and using these methods over disorder-specific cut points is likely to improve accuracy of classifying both type and intensity of physical activity. Notably, there remains a critical need for further development of classifiers for children with more severe motor impairments, preschool aged children, and children in hospital settings.

Evaluation of Wrist Accelerometer Cut-Points for Classifying Physical Activity Intensity in Youth

Background

Wrist worn accelerometers are convenient to wear and provide greater compliance. However, methods to transform the resultant output into predictions of physical activity (PA) intensity have been slow to evolve, with most investigators continuing the practice of applying intensity-based thresholds or cut-points. The current study evaluated the classification accuracy of seven sets of previously published youth-specific cut-points for wrist worn ActiGraph accelerometer data.

Methods

Eighteen children and adolescents [mean age (± SD) 14.6 ± 2.4 years, 10 boys, 8 girls] completed 12 standardized activity trials. During each trial, participants wore an ActiGraph GT3X+ tri-axial accelerometer on the wrist and energy expenditure (Youth METs) was measured directly using the Oxycon Mobile portable calorimetry system. Seven previously published sets of ActiGraph cut-points were evaluated: Crouter regression vertical axis, Crouter regression vector magnitude, Crouter ROC curve vertical axis, Crouter ROC curve vector magnitude, Chandler ROC curve vertical axis, Chandler ROC curve vector magnitude, and Hildebrand ENMO. Classification accuracy was evaluated via weighted Kappa. Confusion matrices were generated to summarize classification accuracy and identify patterns of misclassification.

Results

The cut-points exhibited only moderate agreement with directly measured PA intensity, with Kappa ranging from 0.45 to 0.58. Although the cut-points classified sedentary behavior accurately (> 95%), classification accuracy for the light (3-51%), moderate (12-45%), and vigorous-intensity trials (30-88%) was generally poor. All cut-points underestimated the true intensity of the walking trials, with error rates ranging from 35 to 100%, while the intensity of activity trials requiring significant upper body and/or arm movements was consistently overestimated. The Hildebrand cut-points which serve as the default option in the popular GGIR software package misclassified 30% of the light intensity trials as sedentary and underestimated the intensity of moderate and vigorous intensity trials 75% of the time.

Conclusion

Published ActiGraph cut-points for the wrist, developed specifically for school-aged youth, do not provide acceptable classification accuracy for estimating daily time spent in light, moderate, and vigorous intensity physical activity. The development and deployment of more robust accelerometer data reduction methods such as functional data analysis and machine learning approaches continues to be a research priority.

Decision fusion in healthcare and medicine: a narrative review

OBJECTIVE

To provide an overview of the decision fusion (DF) technique and describe the applications of the technique in healthcare and medicine at prevention, diagnosis, treatment and administrative levels.

BACKGROUND

The rapid development of technology over the past 20 years has led to an explosion in data growth in various industries, like healthcare. Big data analysis within the healthcare systems is essential for arriving to a value-based decision over a period of time. Diversity and uncertainty in big data analytics have made it impossible to analyze data by using conventional data mining techniques and thus alternative solutions are required. DF is a form of data fusion techniques that could increase the accuracy of diagnosis and facilitate interpretation, summarization and sharing of information.

METHODS

We conducted a review of articles published between January 1980 and December 2020 from various databases such as Google Scholar, IEEE, PubMed, Science Direct, Scopus and web of science using the keywords decision fusion (DF), information fusion, healthcare, medicine and big data. A total of 141 articles were included in this narrative review.

CONCLUSIONS

Given the importance of big data analysis in reducing costs and improving the quality of healthcare; along with the potential role of DF in big data analysis, it is recommended to know the full potential of this technique including the advantages, challenges and applications of the technique before its use. Future studies should focus on describing the methodology and types of data used for its applications within the healthcare sector.

Enhanced Human Activity Recognition Using Wearable Sensors via a Hybrid Feature Selection Method

The study of human activity recognition (HAR) plays an important role in many areas such as healthcare, entertainment, sports, and smart homes. With the development of wearable electronics and wireless communication technologies, activity recognition using inertial sensors from ubiquitous smart mobile devices has drawn wide attention and become a research hotspot. Before recognition, the sensor signals are typically preprocessed and segmented, and then representative features are extracted and selected based on them. Considering the issues of limited resources of wearable devices and the curse of dimensionality, it is vital to generate the best feature combination which maximizes the performance and efficiency of the following mapping from feature subsets to activities. In this paper, we propose to integrate bee swarm optimization (BSO) with a deep Q-network to perform feature selection and present a hybrid feature selection methodology, BAROQUE, on basis of these two schemes. Following the wrapper approach, BAROQUE leverages the appealing properties from BSO and the multi-agent deep Q-network (DQN) to determine feature subsets and adopts a classifier to evaluate these solutions. In BAROQUE, the BSO is employed to strike a balance between exploitation and exploration for the search of feature space, while the DQN takes advantage of the merits of reinforcement learning to make the local search process more adaptive and more efficient. Extensive experiments were conducted on some benchmark datasets collected by smartphones or smartwatches, and the metrics were compared with those of BSO, DQN, and some other previously published methods. The results show that BAROQUE achieves an accuracy of 98.41% for the UCI-HAR dataset and takes less time to converge to a good solution than other methods, such as CFS, SFFS, and Relief-F, yielding quite promising results in terms of accuracy and efficiency.

A Framework for Maternal Physical Activities and Health Monitoring Using Wearable Sensors

We propose a physical activity recognition and monitoring framework based on wearable sensors during maternity. A physical activity can either create or prevent health issues during a given stage of pregnancy depending on its intensity. Thus, it becomes very important to provide continuous feedback by recognizing a physical activity and its intensity. However, such continuous monitoring is very challenging during the whole period of maternity. In addition, maintaining a record of each physical activity, and the time for which it was performed, is also a non-trivial task. We aim at such problems by first recognizing a physical activity via the data of wearable sensors that are put on various parts of body. We avoid the use of smartphones for such task due to the inconvenience caused by wearing it for activities such as “eating”. In our proposed framework, a module worn on body consists of three sensors: a 3-axis accelerometer, 3-axis gyroscope, and temperature sensor. The time-series data from these sensors are sent to a Raspberry-PI via Bluetooth Low Energy (BLE). Various statistical measures (features) of this data are then calculated and represented in features vectors. These feature vectors are then used to train a supervised machine learning algorithm called classifier for the recognition of physical activity from the sensors data. Based on such recognition, the proposed framework sends a message to the care-taker in case of unfavorable situation. We evaluated a number of well-known classifiers on various features developed from overlapped and non-overlapped window size of time-series data. Our novel dataset consists of 10 physical activities performed by 61 subjects at various stages of maternity. On the current dataset, we achieve the highest recognition rate of 89% which is encouraging for a monitoring and feedback system.

Activity-Aware Vital SignMonitoring Based on a Multi-Agent Architecture

Vital sign monitoring outside the clinical environment based on wearable sensors ensures better support in assessing a patient’s health condition, and in case of health deterioration, automatic alerts can be sent to the care providers. In everyday life, the users can perform different physical activities, and considering that vital sign measurements depend on the intensity of the activity, we proposed an architecture based on the multi-agent paradigm to handle this issue dynamically. Different types of agents were proposed that processed different sensor signals and recognized simple activities of daily living. The system was validated using a real-life dataset where subjects wore accelerometer sensors on the chest, wrist, and ankle. The system relied on ontology-based models to address the data heterogeneity and combined different wearable sensor sources in order to achieve better performance. The results showed an accuracy of 95.25% on intersubject activity classification. Moreover, the proposed method, which automatically extracted vital sign threshold ranges for each physical activity recognized by the system, showed promising results for remote health status evaluation.

Smart integration of sensors, computer vision and knowledge representation for intelligent monitoring and verbal human-computer interaction

The details presented in this article revolve around a sophisticated monitoring framework equipped with knowledge representation and computer vision capabilities, that aims to provide innovative solutions and support services in the healthcare sector, with a focus on clinical and non-clinical rehabilitation and care environments for people with mobility problems. In contemporary pervasive systems most modern virtual agents have specific reactions when interacting with humans and usually lack extended dialogue and cognitive competences. The presented tool aims to provide natural human-computer multi-modal interaction via exploitation of state-of-the-art technologies in computer vision, speech recognition and synthesis, knowledge representation, sensor data analysis, and by leveraging prior clinical knowledge and patient history through an intelligent, ontology-driven, dialogue manager with reasoning capabilities, which can also access a web search and retrieval engine module. The framework’s main contribution lies in its versatility to combine different technologies, while its inherent capability to monitor patient behaviour allows doctors and caregivers to spend less time collecting patient-related information and focus on healthcare. Moreover, by capitalising on voice, sensor and camera data, it may bolster patients’ confidence levels and encourage them to naturally interact with the virtual agent, drastically improving their moral during a recuperation process.

Posture and Physical Activity Detection: Impact of Number of Sensors and Feature Type

Studies using wearable sensors to measure posture, physical activity (PA), and sedentary behavior typically use a single sensor worn on the ankle, thigh, wrist, or hip. Although the use of single sensors may be convenient, using multiple sensors is becoming more practical as sensors miniaturize.

PURPOSE

We evaluated the effect of single-site versus multisite motion sensing at seven body locations (both ankles, wrists, hips, and dominant thigh) on the detection of physical behavior recognition using a machine learning algorithm. We also explored the effect of using orientation versus orientation-invariant features on performance.

METHODS

Performance (F1 score) of PA and posture recognition was evaluated using leave-one-subject-out cross-validation on a 42-participant data set containing 22 physical activities with three postures (lying, sitting, and upright).

RESULTS

Posture and PA recognition models using two sensors had higher F1 scores (posture, 0.89 ± 0.06; PA, 0.53 ± 0.08) than did models using a single sensor (posture, 0.78 ± 0.11; PA, 0.43 ± 0.03). Models using two nonwrist sensors for posture recognition (F1 score, 0.93 ± 0.03) outperformed two-sensor models including one or two wrist sensors (F1 score, 0.85 ± 0.06). However, two-sensor models for PA recognition with at least one wrist sensor (F1 score, 0.60 ± 0.05) outperformed other two-sensor models (F1 score, 0.47 ± 0.02). Both posture and PA recognition F1 scores improved with more sensors (up to seven; 0.99 for posture and 0.70 for PA), but with diminishing performance returns. Models performed best when including orientation-based features.

CONCLUSIONS

Researchers measuring posture should consider multisite sensing using at least two nonwrist sensors, and researchers measuring PA should consider multisite sensing using at least one wrist sensor and one nonwrist sensor. Including orientation-based features improved both posture and PA recognition.

Deep Learning-Based Multimodal Data Fusion: Case Study in Food Intake Episodes Detection Using Wearable Sensors

BACKGROUND

Multimodal wearable technologies have brought forward wide possibilities in human activity recognition, and more specifically personalized monitoring of eating habits. The emerging challenge now is the selection of most discriminative information from high-dimensional data collected from multiple sources. The available fusion algorithms with their complex structure are poorly adopted to the computationally constrained environment which requires integrating information directly at the source. As a result, more simple low-level fusion methods are needed.

OBJECTIVE

In the absence of a data combining process, the cost of directly applying high-dimensional raw data to a deep classifier would be computationally expensive with regard to the response time, energy consumption, and memory requirement. Taking this into account, we aimed to develop a data fusion technique in a computationally efficient way to achieve a more comprehensive insight of human activity dynamics in a lower dimension. The major objective was considering statistical dependency of multisensory data and exploring intermodality correlation patterns for different activities.

METHODS

In this technique, the information in time (regardless of the number of sources) is transformed into a 2D space that facilitates classification of eating episodes from others. This is based on a hypothesis that data captured by various sensors are statistically associated with each other and the covariance matrix of all these signals has a unique distribution correlated with each activity which can be encoded on a contour representation. These representations are then used as input of a deep model to learn specific patterns associated with specific activity.

RESULTS

In order to show the generalizability of the proposed fusion algorithm, 2 different scenarios were taken into account. These scenarios were different in terms of temporal segment size, type of activity, wearable device, subjects, and deep learning architecture. The first scenario used a data set in which a single participant performed a limited number of activities while wearing the Empatica E4 wristband. In the second scenario, a data set related to the activities of daily living was used where 10 different participants wore inertial measurement units while performing a more complex set of activities. The precision metric obtained from leave-one-subject-out cross-validation for the second scenario reached 0.803. The impact of missing data on performance degradation was also evaluated.

CONCLUSIONS

To conclude, the proposed fusion technique provides the possibility of embedding joint variability information over different modalities in just a single 2D representation which results in obtaining a more global view of different aspects of daily human activities at hand, and yet preserving the desired performance level in activity recognition.

Free-living Evaluation of Laboratory-based Activity Classifiers in Preschoolers

Machine learning classification models for accelerometer data are potentially more accurate methods to measure physical activity in young children than traditional cut point methods. However, existing algorithms have been trained on laboratory-based activity trials, and their performance has not been investigated under free-living conditions.

PURPOSE

This study aimed to evaluate the accuracy of laboratory-trained hip and wrist random forest and support vector machine classifiers for the automatic recognition of five activity classes: sedentary (SED), light-intensity activities and games (LIGHT_AG), walking (WALK), running (RUN), and moderate to vigorous activities and games (MV_AG) in preschool-age children under free-living conditions.

METHODS

Thirty-one children (4.0 ± 0.9 yr) were video recorded during a 20-min free-living play session while wearing an ActiGraph GT3X+ on their right hip and nondominant wrist. Direct observation was used to continuously code ground truth activity class and specific activity types occurring within each class using a bespoke two-stage coding scheme. Performance was assessed by calculating overall classification accuracy and extended confusion matrices summarizing class-level accuracy and the frequency of specific activities observed within each class.

RESULTS

Accuracy values for the hip and wrist random forest algorithms were 69.4% and 59.1%, respectively. Accuracy values for hip and wrist support vector machine algorithms were 66.4% and 59.3%, respectively. Compared with the laboratory cross validation, accuracy decreased by 11%-15% for the hip classifiers and 19%-21% for the wrist classifiers. Classification accuracy values were 72%-78% for SED, 58%-79% for LIGHT_AG, 71%-84% for MV_AG, 9%-15% for WALK, and 66%-75% for RUN.

CONCLUSION

The accuracy of laboratory-based activity classifiers for preschool-age children was attenuated when tested on new data collected under free-living conditions. Future studies should train and test machine learning activity recognition algorithms using accelerometer data collected under free-living conditions.

Smartphone Motion Sensor-Based Complex Human Activity Identification Using Deep Stacked Autoencoder Algorithm for Enhanced Smart Healthcare System

Human motion analysis using a smartphone-embedded accelerometer sensor provided important context for the identification of static, dynamic, and complex sequence of activities. Research in smartphone-based motion analysis are implemented for tasks, such as health status monitoring, fall detection and prevention, energy expenditure estimation, and emotion detection. However, current methods, in this regard, assume that the device is tightly attached to a pre-determined position and orientation, which might cause performance degradation in accelerometer data due to changing orientation. Therefore, it is challenging to accurately and automatically identify activity details as a result of the complexity and orientation inconsistencies of the smartphone. Furthermore, the current activity identification methods utilize conventional machine learning algorithms that are application dependent. Moreover, it is difficult to model the hierarchical and temporal dynamic nature of the current, complex, activity identification process. This paper aims to propose a deep stacked autoencoder algorithm, and orientation invariant features, for complex human activity identification. The proposed approach is made up of various stages. First, we computed the magnitude norm vector and rotation feature (pitch and roll angles) to augment the three-axis dimensions (3-D) of the accelerometer sensor. Second, we propose a deep stacked autoencoder based deep learning algorithm to automatically extract compact feature representation from the motion sensor data. The results show that the proposed integration of the deep learning algorithm, and orientation invariant features, can accurately recognize complex activity details using only smartphone accelerometer data. The proposed deep stacked autoencoder method achieved 97.13% identification accuracy compared to the conventional machine learning methods and the deep belief network algorithm. The results suggest the impact of the proposed method to improve a smartphone-based complex human activity identification framework.

A Comparative Analysis of Hybrid Deep Learning Models for Human Activity Recognition

Recent advances in artificial intelligence and machine learning (ML) led to effective methods and tools for analyzing the human behavior. Human Activity Recognition (HAR) is one of the fields that has seen an explosive research interest among the ML community due to its wide range of applications. HAR is one of the most helpful technology tools to support the elderly's daily life and to help people suffering from cognitive disorders, Parkinson's disease, dementia, etc. It is also very useful in areas such as transportation, robotics and sports. Deep learning (DL) is a branch of ML based on complex Artificial Neural Networks (ANNs) that has demonstrated a high level of accuracy and performance in HAR. Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) are two types of DL models widely used in the recent years to address the HAR problem. The purpose of this paper is to investigate the effectiveness of their integration in recognizing daily activities, e.g., walking. We analyze four hybrid models that integrate CNNs with four powerful RNNs, i.e., LSTMs, BiLSTMs, GRUs and BiGRUs. The outcomes of our experiments on the PAMAP2 dataset indicate that our proposed hybrid models achieve an outstanding level of performance with respect to several indicative measures, e.g., F-score, accuracy, sensitivity, and specificity.

Machine learning to quantify habitual physical activity in children with cerebral palsy

AIM

To investigate whether activity-monitors and machine learning models could provide accurate information about physical activity performed by children and adolescents with cerebral palsy (CP) who use mobility aids for ambulation.

METHOD

Eleven participants (mean age 11y [SD 3y]; six females, five males) classified in Gross Motor Function Classification System (GMFCS) levels III and IV, completed six physical activity trials wearing a tri-axial accelerometer on the wrist, hip, and thigh. Trials included supine rest, upper-limb task, walking, wheelchair propulsion, and cycling. Three supervised learning algorithms (decision tree, support vector machine [SVM], random forest) were trained on features in the raw-acceleration signal. Model-performance was evaluated using leave-one-subject-out cross-validation accuracy.

RESULTS

Cross-validation accuracy for the single-placement models ranged from 59% to 79%, with the best performance achieved by the random forest wrist model (79%). Combining features from two or more accelerometer placements significantly improved classification accuracy. The random forest wrist and hip model achieved an overall accuracy of 92%, while the SVM wrist, hip, and thigh model achieved an overall accuracy of 90%.

INTERPRETATION

Models trained on features in the raw-acceleration signal may provide accurate recognition of clinically relevant physical activity behaviours in children and adolescents with CP who use mobility aids for ambulation in a controlled setting.

WHAT THIS PAPER ADDS

Machine learning may assist clinicians in evaluating the efficacy of surgical and therapy-based interventions. Machine learning may help researchers better understand the short- and long-term benefits of physical activity for children with more severe motor impairments.

Machine Learning Models for Classifying Physical Activity in Free-Living Preschool Children

Machine learning (ML) activity classification models trained on laboratory-based activity trials exhibit low accuracy under free-living conditions. Training new models on free-living accelerometer data, reducing the number of prediction windows comprised of multiple activity types by using shorter windows, including temporal features such as standard deviation in lag and lead windows, and using multiple sensors may improve the classification accuracy under free-living conditions. The objective of this study was to evaluate the accuracy of Random Forest (RF) activity classification models for preschool-aged children trained on free-living accelerometer data. Thirty-one children (mean age = 4.0 ± 0.9 years) completed a 20 min free-play session while wearing an accelerometer on their right hip and non-dominant wrist. Video-based direct observation was used to categorize the children’s movement behaviors into five activity classes. The models were trained using prediction windows of 1, 5, 10, and 15 s, with and without temporal features. The models were evaluated using leave-one-subject-out-cross-validation. The F-scores improved as the window size increased from 1 to 15 s (62.6%–86.4%), with only minimal improvements beyond the 10 s windows. The inclusion of temporal features increased the accuracy, mainly for the wrist classification models, by an average of 6.2 percentage points. The hip and combined hip and wrist classification models provided comparable accuracy; however, both the models outperformed the models trained on wrist data by 7.9 to 8.2 percentage points. RF activity classification models trained with free-living accelerometer data provide accurate recognition of young children’s movement behaviors under real-world conditions.

Machine Learning to Quantify Physical Activity in Children with Cerebral Palsy: Comparison of Group, Group-Personalized, and Fully-Personalized Activity Classification Models

Pattern recognition methodologies, such as those utilizing machine learning (ML) approaches, have the potential to improve the accuracy and versatility of accelerometer-based assessments of physical activity (PA). Children with cerebral palsy (CP) exhibit significant heterogeneity in relation to impairment and activity limitations; however, studies conducted to date have implemented “one-size fits all” group (G) models. Group-personalized (GP) models specific to the Gross Motor Function Classification (GMFCS) level and fully-personalized (FP) models trained on individual data may provide more accurate assessments of PA; however, these approaches have not been investigated in children with CP. In this study, 38 children classified at GMFCS I to III completed laboratory trials and a simulated free-living protocol while wearing an ActiGraph GT3X+ on the wrist, hip, and ankle. Activities were classified as sedentary, standing utilitarian movements, or walking. In the cross-validation, FP random forest classifiers (99.0–99.3%) exhibited a significantly higher accuracy than G (80.9–94.7%) and GP classifiers (78.7–94.1%), with the largest differential observed in children at GMFCS III. When evaluated under free-living conditions, all model types exhibited significant declines in accuracy, with FP models outperforming G and GP models in GMFCS levels I and II, but not III. Future studies should evaluate the comparative accuracy of personalized models trained on free-living accelerometer data.

Laboratory-based and free-living algorithms for energy expenditure estimation in preschool children: A free-living evaluation

Machine learning models to predict energy expenditure (EE) from accelerometer data have traditionally been trained on data from laboratory-based activity trials. However, accuracy is typically attenuated when implemented in free-living scenarios. Currently, no studies involving preschool children have evaluated the accuracy of EE prediction models trained on laboratory (LAB) under free-living conditions.

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.

Step by Step Towards Effective Human Activity Recognition: A Balance between Energy Consumption and Latency in Health and Wellbeing Applications

Human activity recognition (HAR) is a classification process that is used for recognizing human motions. A comprehensive review of currently considered approaches in each stage of HAR, as well as the influence of each HAR stage on energy consumption and latency is presented in this paper. It highlights various methods for the optimization of energy consumption and latency in each stage of HAR that has been used in literature and was analyzed in order to provide direction for the implementation of HAR in health and wellbeing applications. This paper analyses if and how each stage of the HAR process affects energy consumption and latency. It shows that data collection and filtering and data segmentation and classification stand out as key stages in achieving a balance between energy consumption and latency. Since latency is only critical for real-time HAR applications, the energy consumption of sensors and devices stands out as a key challenge for HAR implementation in health and wellbeing applications. Most of the approaches in overcoming challenges related to HAR implementation take place in the data collection, filtering and classification stages, while the data segmentation stage needs further exploration. Finally, this paper recommends a balance between energy consumption and latency for HAR in health and wellbeing applications, which takes into account the context and health of the target population.

Optimization of Physical Activity Recognition for Real-Time Wearable Systems: Effect of Window Length, Sampling Frequency and Number of Features

The aim of this study was to develop an optimized physical activity classifier for real-time wearable systems with the focus on reducing the requirements on device power consumption and memory buffer. Classification parameters evaluated in this study were the sampling frequency of the acceleration signal, window length of the classification fragment, and the number of classification features, found with different feature selection methods. For parameter evaluation, a decision tree classifier was created based on the acceleration signals recorded during tests, where 25 healthy test subjects performed various physical activities. Overall average F1-score achieved in this study was about 0.90. Similar F1-scores were achieved with the evaluated window lengths of 5 s (0.92 ± 0.02) and 3 s (0.91 ± 0.02), while classification performance with 1 s were lower (0.87 ± 0.02). Tested sampling frequencies of 50 Hz, 25 Hz, and 13 Hz had similar results with most classified activity types, with an exception of outdoor cycling, where differences were significant. Using forward sequential feature selection enabled the decreasing of the number of features from initial 110 features to about 12 features without lowering the classification performance. The results of this study have been used for developing more efficient real-time physical activity classifiers.

Prediction of Relative Physical Activity Intensity Using Multimodal Sensing of Physiological Data

This study examined the feasibility of a non-laboratory approach that uses machine learning on multimodal sensor data to predict relative physical activity (PA) intensity. A total of 22 participants completed up to 7 PA sessions, where each session comprised 5 trials (sitting and standing, comfortable walk, brisk walk, jogging, running). Participants wore a wrist-strapped sensor that recorded heart-rate (HR), electrodermal activity (Eda) and skin temperature (Temp). After each trial, participants provided ratings of perceived exertion (RPE). Three classifiers, including random forest (RF), neural network (NN) and support vector machine (SVM), were applied independently on each feature set to predict relative PA intensity as low (RPE ≤ 11), moderate (RPE 12–14), or high (RPE ≥ 15). Then, both feature fusion and decision fusion of all combinations of sensor modalities were carried out to investigate the best combination. Among the single modality feature sets, HR provided the best performance. The combination of modalities using feature fusion provided a small improvement in performance. Decision fusion did not improve performance over HR features alone. A machine learning approach using features from HR provided acceptable predictions of relative PA intensity. Adding features from other sensing modalities did not significantly improve performance.

Effects of a training programme of functional electrical stimulation (FES) powered cycling, recreational cycling and goal-directed exercise training on children with cerebral palsy: a randomised controlled trial protocol

INTRODUCTION

Children with cerebral palsy (CP) experience declines in gross motor ability as they transition from childhood to adolescence, which can result in the loss of ability to perform sit-to-stand transfers, ambulate or participate in leisure activities such as cycling. Functional electrical stimulation (FES) cycling is a novel technology that may provide opportunities for children with CP to strengthen their lower limbs, improve functional independence and increase physical activity participation. The proposed randomised controlled trial will test the efficacy of a training package of FES cycling, adapted cycling and goal-directed functional training to usual care in children with CP who are susceptible to functional declines.

METHODS AND ANALYSIS

Forty children with CP (20 per group), aged 6-8 years and classified as Gross Motor Function Classification System (GMFCS) levels II-IV will be recruited across South East Queensland. Participants will be randomised to either an immediate intervention group, who will undertake 8 weeks of training, or a waitlist control group. The training group will attend two 1 hour sessions per week with a physiotherapist, consisting of FES cycling and goal-directed, functional exercises and a 1 hour home exercise programme per week, consisting of recreational cycling. Primary outcomes will be the gross motor function measure and Canadian occupational performance measure, and secondary outcomes will include the five times sit-to-stand test, habitual physical activity (accelerometry), power output during cycling and Participation and Environment Measure-Children and Youth. Outcomes will be assessed at baseline, postintervention (8 weeks) and 8 weeks following the intervention (retention).

ETHICS AND DISSEMINATION

Ethical approval has been obtained from Griffith University (2018/037) and the Children's Health Queensland Hospital and Health Service (CHQHHS) Human Research Ethics Committee (HREC/17/QRCH/88). Site-specific approval was obtained from CHQHHS research governance (SSA/17/QRCH/145). Results from this trial will be disseminated via publication in relevant peer-reviewed journals.

TRIAL REGISTRATION NUMBER

ACTRN12617000644369p.

TSE-CNN: A Two-Stage End-to-End CNN for Human Activity Recognition

Human activity recognition has been widely used in healthcare applications such as elderly monitoring, exercise supervision, and rehabilitation monitoring. Compared with other approaches, sensor-based wearable human activity recognition is less affected by environmental noise and therefore is promising in providing higher recognition accuracy. However, one of the major issues of existing wearable human activity recognition methods is that although the average recognition accuracy is acceptable, the recognition accuracy for some activities (e.g., ascending stairs and descending stairs) is low, mainly due to relatively less training data and complex behavior pattern for these activities. Another issue is that the recognition accuracy is low when the training data from the test subject are limited, which is a common case in real practice. In addition, the use of neural network leads to large computational complexity and thus high power consumption. To address these issues, we proposed a new human activity recognition method with two-stage end-to-end convolutional neural network and a data augmentation method. Compared with the state-of-the-art methods (including neural network based methods and other methods), the proposed methods achieve significantly improved recognition accuracy and reduced computational complexity.

Activity Monitoring with a Wrist-Worn, Accelerometer-Based Device

This study condenses huge amount of raw data measured from a MEMS accelerometer-based, wrist-worn device on different levels of physical activities (PAs) for subjects wearing the device 24 h a day continuously. In this study, we have employed the device to build up assessment models for quantifying activities, to develop an algorithm for sleep duration detection and to assess the regularity of activity of daily living (ADL) quantitatively. A new parameter, the activity index (AI), has been proposed to represent the quantity of activities and can be used to categorize different PAs into 5 levels, namely, rest/sleep, sedentary, light, moderate, and vigorous activity states. Another new parameter, the regularity index (RI), was calculated to represent the degree of regularity for ADL. The methods proposed in this study have been used to monitor a subject’s daily PA status and to access sleep quality, along with the quantitative assessment of the regularity of activity of daily living (ADL) with the 24-h continuously recorded data over several months to develop activity-based evaluation models for different medical-care applications. This work provides simple models for activity monitoring based on the accelerometer-based, wrist-worn device without trying to identify the details of types of activity and that are suitable for further applications combined with cloud computing services.