Wearable Sensor Data Research Articles

Machine Learning Algorithms Can Use Wearable Sensor Data to Accurately Predict Six-Week Patient-Reported Outcome Scores Following Joint Replacement in a Prospective Trial

BackgroundTracking patient-generated health data (PGHD) following total joint arthroplasty (TJA) may enable data-driven early intervention to improve clinical results. We aim to demonstrate the feasibility of combining machine learning (ML) with PGHD in TJA to predict patient-reported outcome measures (PROMs). MethodsTwenty-two TJA patients were recruited for this pilot study. Three activity trackers collected 35 features from 4 weeks before to 6 weeks following surgery. PROMs were collected at both endpoints (Hip and Knee Disability and Osteoarthritis Outcome Score, Knee Osteoarthritis Outcome Score, and Veterans RAND 12-Item Health Survey Physical Component Score). We used ML to identify features with the highest correlation with PROMs. The algorithm trained on a subset of patients and used 3 feature sets (A, B, and C) to group the rest into one of the 3 PROM clusters. ResultsFifteen patients completed the study and collected 3 million data points. Three sets of features with the highest R2 values relative to PROMs were selected (A, B and C). Data collected through the 11th day had the highest predictive value. The ML algorithm grouped patients into 3 clusters predictive of 6-week PROM results, yielding total sum of squares values ranging from 3.86 (A) to 1.86 (C). ConclusionThis small but critical proof-of-concept study demonstrates that ML can be used in combination with PGHD to predict 6-week PROM data as early as 11 days following TJA surgery. Further study is needed to confirm these findings and their clinical value.

Deep Learning Techniques for Improving Digital Gait Segmentation.

Wearable technology for the automatic detection of gait events has recently gained growing interest, enabling advanced analyses that were previously limited to specialist centres and equipment (e.g., instrumented walkway). In this study, we present a novel method based on dilated convolutions for an accurate detection of gait events (initial and final foot contacts) from wearable inertial sensors. A rich dataset has been used to validate the method, featuring 71 people with Parkinson's disease (PD) and 67 healthy control subjects. Multiple sensors have been considered, one located on the fifth lumbar vertebrae and two on the ankles. The aims of this study were: (i) to apply deep learning (DL) techniques on wearable sensor data for gait segmentation and quantification in older adults and in people with PD; (ii) to validate the proposed technique for measuring gait against traditional gold standard laboratory reference and a widely used algorithm based on wavelet transforms (WT); (iii) to assess the performance of DL methods in assessing high-level gait characteristics, with focus on stride, stance and swing related features. The results showed a high reliability of the proposed approach, which achieves temporal errors considerably smaller than WT, in particular for the detection of final contacts, with an inter-quartile range below 70 ms in the worst case. This study showes encouraging results, and paves the road for further research, addressing the effectiveness and the generalization of data-driven learning systems for accurate event detection in challenging conditions.

Unsupervised Factory Activity Recognition with Wearable Sensors Using Process Instruction Information

This paper presents an unsupervised method for recognizing assembly work done by factory workers by using wearable sensor data. Such assembly work is a common part of line production systems and typically involves the factory workers performing a repetitive work process made up of a sequence of manual operations, such as setting a board on a workbench and screwing parts onto the board. This study aims to recognize the starting and ending times for individual operations in such work processes through analysis of sensor data collected from the workers along with analysis of the process instructions that detail and describe the flow of operations for each work process. We propose a particle-filter-based factory activity recognition method that leverages (i) trend changes in the sensor data detected by a nonparametric Bayesian hidden Markov model, (ii) semantic similarities between operations discovered in the process instructions, (iii) sensor-data similarities between consecutive repetitions of individual operations, and (iv) frequent sensor-data patterns (motifs) discovered in the overall assembly work processes. We evaluated the proposed method using sensor data from six workers collected in actual factories, achieving a recognition accuracy of 80% (macro-averaged F-measure).

Fall Risk Classification in Community-Dwelling Older Adults Using a Smart Wrist-Worn Device and the Resident Assessment Instrument-Home Care: Prospective Observational Study.

Background Little is known about whether off-the-shelf wearable sensor data can contribute to fall risk classification or complement clinical assessment tools such as the Resident Assessment Instrument-Home Care (RAI-HC).Objective This study aimed to (1) investigate the similarities and differences in physical activity (PA), heart rate, and night sleep in a sample of community-dwelling older adults with varying fall histories using a smart wrist-worn device and (2) create and evaluate fall risk classification models based on (i) wearable data, (ii) the RAI-HC, and (iii) the combination of wearable and RAI-HC data.Methods A prospective, observational study was conducted among 3 faller groups (G0, G1, G2+) based on the number of previous falls (0, 1, ≥2 falls) in a sample of older community-dwelling adults. Each participant was requested to wear a smart wristband for 7 consecutive days while carrying out day-to-day activities in their normal lives. The wearable and RAI-HC assessment data were analyzed and utilized to create fall risk classification models, with 3 supervised machine learning algorithms: logistic regression, decision tree, and random forest (RF).Results Of 40 participants aged 65 to 93 years, 16 (40%) had no previous falls, whereas 8 (20%) and 16 (40%) had experienced 1 and multiple (≥2) falls, respectively. Level of PA as measured by average daily steps was significantly different between groups (P=.04). In the 3 faller group classification, RF achieved the best accuracy of 83.8% using both wearable and RAI-HC data, which is 13.5% higher than that of using the RAI-HC data only and 18.9% higher than that of using wearable data exclusively. In discriminating between {G0+G1} and G2+, RF achieved the best area under the receiver operating characteristic curve of 0.894 (overall accuracy of 89.2%) based on wearable and RAI-HC data. Discrimination between G0 and {G1+G2+} did not result in better classification performance than that between {G0+G1} and G2+.Conclusions Both wearable data and the RAI-HC assessment can contribute to fall risk classification. All the classification models revealed that RAI-HC outperforms wearable data, and the best performance was achieved with the combination of 2 datasets. Future studies in fall risk assessment should consider using wearable technologies to supplement resident assessment instruments.

A novel single-sensor-based method for the detection ofgait-cycle breakdown and freezing of gait in Parkinson's disease.

Objective measurement of walking speed and gait deficits are an important clinical tool in chronic illness management. We previously reported in Parkinson's disease that different types of gait tests can now be implemented and administered in the clinic or at home using Ambulosono smartphone-sensor technology, whereby movement sensing protocols can be standardized under voice instruction. However, a common challenge that remains for such wearable sensor systems is how meaningful data can be extracted from seemingly "noisy" raw sensor data, and do so with a high level of accuracy and efficiency. Here, we describe a novel pattern recognition algorithm for the automated detection of gait-cycle breakdown and freezing episodes. Ambulosono-gait-cycle-breakdown-and-freezing-detection (Free-D) integrates a nonlinear m-dimensional phase-space data extraction method with machine learning and Monte Carlo analysis for model building and pattern generalization. We first trained Free-D using a small number of data samples obtained from thirty participants during freezing of gait tests. We then tested the accuracy of Free-D via Monte Carlo cross-validation. We found Free-D to be remarkably effective at detecting gait-cycle breakdown, with mode error rates of 0% and mean error rates < 5%. We also demonstrate the utility of Free-D by applying it to continuous holdout traces not used for either training or testing, andfound it was able to identify gait-cycle breakdown and freezing events of varying duration. These results suggest that advanced artificial intelligence and automation tools can be developed to enhance the quality, efficiency, and the expansion of wearable sensor data processing capabilities to meet market and industry demand.

Impact Identification With Machine Learning From Low Frequency Wearable Sensor Data Among National Ice-hockey Players

Impact Identification With Machine Learning From Low Frequency Wearable Sensor Data Among National Ice-hockey Players

Toward Sensor-Based Sleep Monitoring with Electrodermal Activity Measures.

We use self-report and electrodermal activity (EDA) wearable sensor data from 77 nights of sleep of six participants to test the efficacy of EDA data for sleep monitoring. We used factor analysis to find latent factors in the EDA data, and used causal model search to find the most probable graphical model accounting for self-reported sleep efficiency (SE), sleep quality (SQ), and the latent factors in the EDA data. Structural equation modeling was used to confirm fit of the extracted graph to the data. Based on the generated graph, logistic regression and naïve Bayes models were used to test the efficacy of the EDA data in predicting SE and SQ. Six EDA features extracted from the total signal over a night’s sleep could be explained by two latent factors, EDA Magnitude and EDA Storms. EDA Magnitude performed as a strong predictor for SE to aid detection of substantial changes in time asleep. The performance of EDA Magnitude and SE in classifying SQ demonstrates promise for using a wearable sensor for sleep monitoring. However, our data suggest that obtaining a more accurate sensor-based measure of SE will be necessary before smaller changes in SQ can be detected from EDA sensor data alone.

Overcoming Limitations of the HoloLens for Use in Product Assembly

Augmented Reality (AR) is an emerging technology that could greatly increase training efficiency and effectiveness for assembly processes. In fact, studies show that AR may reduce time and errors by as much as 50%. While many devices are available to display AR content, the use of Head Mounted Displays (HMDs) is the most interesting for manufacturing and assembly as they free up the user&#x2019;s hands. Due to the emerging nature of this technology, there are many limitations including input, field of view, tracking and occlusion. The work presented in this paper explores the use of the Microsoft HoloLens to deliver AR work instructions for product assembly. An AR assembly application was developed to guide a trainee through the assembly of a mock aircraft wing. To ensure accurate instructions were displayed, novel techniques were developed to mitigate the HoloLens&#x2019; tracking and display limitations. Additionally, a data visualization application was developed to validate the training session and explore trends through data collected from the HoloLens along with wearable physiological sensor data.

Robust Human Activity Recognition Using Multimodal Feature-Level Fusion

Automated recognition of human activities or actions has great significance as it incorporates wide-ranging applications, including surveillance, robotics, and personal health monitoring. Over the past few years, many computer vision-based methods have been developed for recognizing human actions from RGB and depth camera videos. These methods include space-time trajectory, motion encoding, key poses extraction, space-time occupancy patterns, depth motion maps, and skeleton joints. However, these camera-based approaches are affected by background clutter and illumination changes and applicable to a limited field of view only. Wearable inertial sensors provide a viable solution to these challenges but are subject to several limitations such as location and orientation sensitivity. Due to the complementary trait of the data obtained from the camera and inertial sensors, the utilization of multiple sensing modalities for accurate recognition of human actions is gradually increasing. This paper presents a viable multimodal feature-level fusion approach for robust human action recognition, which utilizes data from multiple sensors, including RGB camera, depth sensor, and wearable inertial sensors. We extracted the computationally efficient features from the data obtained from RGB-D video camera and inertial body sensors. These features include densely extracted histogram of oriented gradient (HOG) features from RGB/depth videos and statistical signal attributes from wearable sensors data. The proposed human action recognition (HAR) framework is tested on a publicly available multimodal human action dataset UTD-MHAD consisting of 27 different human actions. K-nearest neighbor and support vector machine classifiers are used for training and testing the proposed fusion model for HAR. The experimental results indicate that the proposed scheme achieves better recognition results as compared to the state of the art. The feature-level fusion of RGB and inertial sensors provides the overall best performance for the proposed system, with an accuracy rate of 97.6%.

A New Weight and Sensitivity Based Variable Maximum Distance to Average Vector Algorithm for Wearable Sensor Data Privacy Protection

The problem of privacy protection of wearable devices when publishing data can be solved based on the variable-maximum distance average vector. This paper proposes a new weight and sensitivity based variable maximum distance average vector (WSV-MDAV) method aiming to solve the problems that may be contained in the existing privacy protection algorithm. The proposed approach considers the difference of the importance among all the identifiers by setting corresponding weight coefficient W. Given a specific weight to each attribute in the table, we can subsequently get a distance metric based on weight. Similarly, the different sensitivity constraint S for different sensitive attributes is also available for our proposed method. Using the WSV-MDAV algorithm we propose a new privacy protection model for the data publishing of wearable device by introducing the concept of the differential privacy. The numerical results show that the proposed WSV-MDAV algorithm improves the privacy protection performance and reduces the information loss compared to the traditional method.

Outlier Detection in Wearable Sensor Data for Human Activity Recognition (HAR) Based on DRNNs

Wearable sensors provide a user-friendly and non-intrusive mechanism to extract user-related data that paves the way to the development of personalized applications. Within those applications, human activity recognition (HAR) plays an important role in the characterization of the user context. Outlier detection methods focus on finding anomalous data samples that are likely to have been generated by a different mechanism. This paper combines outlier detection and HAR by introducing a novel algorithm that is able both to detect information from secondary activities inside the main activity and to extract data segments of a particular sub-activity from a different activity. Several machine learning algorithms have been previously used in the area of HAR based on the analysis of the time sequences generated by wearable sensors. Deep recurrent neural networks (DRNNs) have proven to be optimally adapted to the sequential characteristics of wearable sensor data in previous studies. A DRNN-based algorithm is proposed in this paper for outlier detection in HAR. The results are validated both for intra- and inter-subject cases and both for outlier detection and sub-activity recognition using two different datasets. A first dataset comprising 4 major activities (walking, running, climbing up, and down) from 15 users is used to train and validate the proposal. Intra-subject outlier detection is able to detect all major outliers in the walking activity in this dataset, while inter-subject outlier detection only fails for one participant executing the activity in a peculiar way. Sub-activity detection has been validated by finding out and extracting walking segments present in the other three activities in this dataset. A second dataset using four different users, a different setting and different sensor devices is used to assess the generalization of results.

An edge-stream computing infrastructure for real-time analysis of wearable sensors data

The fast development of IoT in general and wearable smart sensors in particular in the context of wellness and healthcare are demanding for definition of specific infrastructure supporting real time data analysis for anomaly detection, event identification, situation awareness just to mention few. The explosion in the development and adoption of these smart wearable sensors has contributed to the definition of the Internet of Medical Things (IoMT), which is revolutionizing the way healthcare is tackled worldwide. Data produced by wearable sensors continuously grow and could be spread among clinical centers, hospitals, research labs, yielding to a Big Data management problem. In this paper we propose a technological and architectural solution, based on Open Source big data technologies to perform real-time analysis of wearable sensor data streams. The proposed architecture is composed of four distinct layers: a sensing layer, a pre-processing layer (Raspberry Pi), a cluster processing layer (Kafka’s broker and Flink’s mini-cluster) and a persistence layer (Cassandra database). A performance evaluation of each layer has been carried out by considering CPU and memory usage for accomplishing a simple anomaly detection task using the REALDISP dataset.

Robust Sparse Representation-Based Classification Using Online Sensor Data for Monitoring Manual Material Handling Tasks

Sensor-based online process monitoring has extensive applications, such as in manufacturing and service industries. In real environments, though, sensor data are often contaminated with noise, leading to severe challenges in accurate data analysis. In the existing literature, noise is generally modeled as Gaussian to analyze sensor data for various applications, for example in fault detection and diagnostics. However, in some applications, such as due to challenging field conditions, sensor data may be disturbed by high levels of outliers such that the Gaussian assumption of sensor noise is inadequate, thus leading to large estimation errors. This paper focuses on online classification applications. A robust sparse representation classification method is proposed, which considers non-Gaussian noise, and thus can effectively analyze sensor data with higher levels of outliers. Case studies were completed, based on both numerically simulated sensor data and actual wearable sensor data from occupational manual material handling process monitoring. The proposed classification method could effectively analyze sensor data with non-Gaussian noise, and outperformed commonly used methods in the literature. Thus, this new method may be advantageous for solving classification problems in challenging field conditions, to address the difficulties of high levels of sensor outliers. Note to Practitioners —This paper proposes a fast, robust classification method for online sensor data classification. The proposed method is designed to cope with high levels of sensor outliers. The robustness of the method and its computational efficiency make it particularly appealing for online sensor data classification in challenging field conditions in which the presence of sensor outliers causes practical difficulties for most existing classification algorithms.

TeamSense

Continuous monitoring with unobtrusive wearable social sensors is becoming a popular method to assess individual affect states and team effectiveness in human research. A large number of applications have demonstrated the effectiveness of applying wearable sensing in corporate settings; for example, in short periodic social events or in a university campus. However, little is known of how we can automatically detect individual affect and group cohesion for long duration missions. Predicting negative affect states and low cohesiveness is vital for team missions. Knowing team members' negative states allows timely interventions to enhance their effectiveness. This work investigates whether sensing social interactions and individual behaviors with wearable sensors can provide insights into assessing individual affect states and group cohesion. We analyzed wearable sensor data from a team of six crew members who were deployed on a four-month simulation of a space exploration mission at a remote location. Our work proposes to recognize team members' affect states and group cohesion as a binary classification problem using novel behavior features that represent dyadic interaction and individual activities. Our method aggregates features from individual members into group levels to predict team cohesion. Our results show that the behavior features extracted from the wearable social sensors provide useful information in assessing personal affect and team cohesion. Group task cohesion can be predicted with a high performance of over 0.8 AUC. Our work demonstrates that we can extract social interactions from sensor data to predict group cohesion in longitudinal missions. We found that quantifying behavior patterns including dyadic interactions and face-to-face communications are important in assessing team process.

Feature Representation and Data Augmentation for Human Activity Classification Based on Wearable IMU Sensor Data Using a Deep LSTM Neural Network.

Wearable inertial measurement unit (IMU) sensors are powerful enablers for acquisition of motion data. Specifically, in human activity recognition (HAR), IMU sensor data collected from human motion are categorically combined to formulate datasets that can be used for learning human activities. However, successful learning of human activities from motion data involves the design and use of proper feature representations of IMU sensor data and suitable classifiers. Furthermore, the scarcity of labelled data is an impeding factor in the process of understanding the performance capabilities of data-driven learning models. To tackle these challenges, two primary contributions are in this article: first; by using raw IMU sensor data, a spectrogram-based feature extraction approach is proposed. Second, an ensemble of data augmentations in feature space is proposed to take care of the data scarcity problem. Performance tests were conducted on a deep long term short term memory (LSTM) neural network architecture to explore the influence of feature representations and the augmentations on activity recognition accuracy. The proposed feature extraction approach combined with the data augmentation ensemble produces state-of-the-art accuracy results in HAR. A performance evaluation of each augmentation approach is performed to show the influence on classification accuracy. Finally, in addition to using our own dataset, the proposed data augmentation technique is evaluated against the University of California, Irvine (UCI) public online HAR dataset and yields state-of-the-art accuracy results at various learning rates.

Personal Health Data: Access and Perceived Value in Denmark.

This study explores how accessible and valuable Personal Health Data are in Denmark. This paper uses a qualitative inquiry which was adopted to provide information about (1) the accessibility of data available, (2) and the perceived value of data by recruiting 8 healthy Danish individuals who were instructed to access their personal health data, and were then prompted to discuss how accessible and valuable they perceived their personal health data to be. In total, participants accessed 31 datasets and wearable sensor data through 23 web applications and 8 mobile applications. They reported on search and access challenges in interviews and through journaling. Our results suggest that participants were satisfied with the access they had to their personal health data, however the participants expressed disappointment in ways the data was presented for them by the services and platforms. Thus, we concluded that the perceived value of personal data were found to be dependent on the usability and personalization features of the services, rather than on the data itself.

Relationships between full-day arm movement characteristics and developmental status in infants with typical development as they learn to reach: An observational study.

Background: Advances in wearable sensor technology now allow us to quantify the number, type and kinematic characteristics of bouts of infant arm movement made across a full day in the natural environment. Our aim here was to determine whether the amount and kinematic characteristics of arm movements made across the day in the natural environment were related to developmental status in infants with typical development as they learned to reach for objects using their arms. Methods: We used wearable sensors to measure arm movement across days and months as infants developed arm reaching skills. In total, 22 infants with typical development participated, aged between 38 and 203 days. Of the participants, 2 infants were measured once and the other 20 infants were measured once per month for 3 to 6 visits. The Bayley Scales of Infant Development was used to measure developmental level. Results: Our main findings were: 1) infant arm movement characteristics as measured by full-day wearable sensor data were related to Bayley motor, cognitive and language scores, indicating a relationship between daily movement characteristics and developmental status; 2) infants who moved more had larger increases in language and cognitive scores across visits; and 3) larger changes in movement characteristics across visits were related to higher motor scores. Conclusions: This was a preliminary, exploratory, small study of the potential importance of infant arm movement characteristics as measured by full-day wearable sensor data. Our results support full-day arm movement activity as an area of interest for future study as a biomarker of neurodevelopmental status and as a target for early intervention.

Relationships between full-day arm movement characteristics and developmental status in infants with typical development as they learn to reach: An observational study

Background: Advances in wearable sensor technology now allow us to quantify the number, type and kinematic characteristics of bouts of infant arm movement made across a full day in the natural environment. Our aim here was to determine whether the amount and kinematic characteristics of arm movements made across the day in the natural environment were related to developmental status in infants with typical development as they learned to reach for objects using their arms. Methods: We used wearable sensors to measure arm movement across days and months as infants developed arm reaching skills. In total, 22 infants with typical development participated, aged between 38 and 203 days. Of the participants, 2 infants were measured once and the other 20 infants were measured once per month for 3 to 6 visits. The Bayley Scales of Infant Development was used to measure developmental level. Results: Our main findings were: 1) infant arm movement characteristics as measured by full-day wearable sensor data were related to Bayley motor, cognitive and language scores, indicating a relationship between daily movement characteristics and developmental status; 2) infants who moved more had larger increases in language and cognitive scores across visits; and 3) larger changes in movement characteristics across visits were related to higher motor scores. Conclusions: This was a preliminary, exploratory, small study of the potential importance of infant arm movement characteristics as measured by full-day wearable sensor data. Our results support full-day arm movement activity as an area of interest for future study as a biomarker of neurodevelopmental status and as a target for early intervention.

Wearable Sensor Application for Integrated Early Warning and Health Surveillance

Objective: The Wearable Sensor Application developed by Pacific Northwest National Laboratory (PNNL) provides an early warning system for stressors to individual and group health using physiologic and environmental indicators. The application integrates health monitoring parameters from wearable sensors, e.g., temperature and heart rate, with relevant environmental parameters, e.g., weather and landscape data, and calculates the corresponding physiological strain index. The information is presented to the analyst in a group and individual view with real-time alerting of abnormal health parameters. This application is the first of its kind being developed for integration into the Defense Threat Reduction Agency's Biosurveillance Ecosystem (BSVE).Introduction: Wearable devices are a low cost, minimally invasive way to monitor health. Sensor data provides real-time physiological indictors of an individual’s health status without the requirement of health care professionals or facilities. Information gleamed from wearable sensors can be used to better understand physiological stressors and prodromal symptoms. In addition, this data can be used to monitor individuals that are in high risk of health-related problems.However, raw data from wearable sensors can be overwhelming to process and laborious to monitor for an individual and, even more so, for a group of individuals. Often specific combination of ranges of sensor readings are indicative of changes to health status and need to be evaluated together or used to calculate specific signal parameters. In addition, the environment surrounding the individual needs to be considered when interpreting the data. To address these issues, PNNL has developed an application that collects, analyzes, and integrates wearable sensor data with geographic landscape and weather information to provide a real-time early alert and situational awareness tool for monitoring the health of groups and individuals.Methods: The prototype application described here was a product of PNNL’s BSVE Application Development Competition. The final product that will be deployed in the BSVE is currently under development by PNNL and will vary slightly in the exact design and architecture described.Data. Wearable sensor data was collected from the Rim2Rim (R2R) Watch Study of individuals hiking the Grand Canyon in Arizona [1]. Weather information was obtained from nearby weather stations and mapping features were derived from Google Maps.Calculations. A physiological Strain Index (PSI) was calculated using core temperature estimates derived through a Kalman Filter approach and heart rate [2,3].Application. The prototype backend application development was based in Python with a MongoDB. The front-end development was built using a scalable architecture and modular approach with components in React and D3.Results: A prototype application was developed this past summer through the PNNL BSVE App Competition (Fig 1). The application was aimed at visualizing wearable sensor data from the Grand Canyon R2R hike dataset. Simulated real-time analysis was used to calculate health status of individuals hiking based on measured physiological parameters and to alert to individuals with signs of physiologic health stress. Visualization tools were incorporated to enable sensor data for individuals and the group to be viewed simultaneously along with pertinent weather, geographic, and elevation data.Many features described in the prototype application will be incorporated into the final BSVE application. The key changes will be 1) the ability to select given time periods for viewing historical data as well as the real-time data collection, 2) environmental data and map view will come from BSVE internal data sources, and 3) the alerts will provide more information and have their own page for reviewing.Conclusions: The Wearable Sensor Application developed by PNNL for integration into the BSVE provides an early warning system for individual and group health using physiologic and environmental parameters. The application highlights health status from wearable sensors and relevant environmental parameters while monitoring a calculated physiological strain index. With this tool, an analyst can easily monitor the health of individuals and groups with the aid of real-time alerting tool for early detection of abnormal health parameters.

Deep Learning to Predict Falls in Older Adults Based on Daily-Life Trunk Accelerometry.

Early detection of high fall risk is an essential component of fall prevention in older adults. Wearable sensors can provide valuable insight into daily-life activities; biomechanical features extracted from such inertial data have been shown to be of added value for the assessment of fall risk. Body-worn sensors such as accelerometers can provide valuable insight into fall risk. Currently, biomechanical features derived from accelerometer data are used for the assessment of fall risk. Here, we studied whether deep learning methods from machine learning are suited to automatically derive features from raw accelerometer data that assess fall risk. We used an existing dataset of 296 older adults. We compared the performance of three deep learning model architectures (convolutional neural network (CNN), long short-term memory (LSTM) and a combination of these two (ConvLSTM)) to each other and to a baseline model with biomechanical features on the same dataset. The results show that the deep learning models in a single-task learning mode are strong in recognition of identity of the subject, but that these models only slightly outperform the baseline method on fall risk assessment. When using multi-task learning, with gender and age as auxiliary tasks, deep learning models perform better. We also found that preprocessing of the data resulted in the best performance (AUC = 0.75). We conclude that deep learning models, and in particular multi-task learning, effectively assess fall risk on the basis of wearable sensor data.