Charence Wong

Deep learning for health informatics

With a massive influx of multimodality data, the role of data analytics in health informatics has grown rapidly in the last decade. This has also prompted increasing interests in the generation of analytical, data driven models based on machine learning in health informatics. Deep learning, a technique with its foundation in artificial neural networks, is emerging in recent years as a powerful tool for machine learning, promising to reshape the future of artificial intelligence. Rapid improvements in computational power, fast data storage, and parallelization have also contributed to the rapid uptake of the technology in addition to its predictive power and ability to generate automatically optimized high-level features and semantic interpretation from the input data. This article presents a comprehensive up-to-date review of research employing deep learning in health informatics, providing a critical analysis of the relative merit, and potential pitfalls of the technique as well as its future outlook. The paper mainly focuses on key applications of deep learning in the fields of translational bioinformatics, medical imaging, pervasive sensing, medical informatics, and public health.

A deep learning approach to on-node sensor data analytics for mobile or wearable devices

The increasing popularity of wearable devices in recent years means that a diverse range of physiological and functional data can now be captured continuously for applications in sports, wellbeing, and healthcare. This wealth of information requires efficient methods of classification and analysis where deep learning is a promising technique for large-scale data analytics. While deep learning has been successful in implementations that utilize high-performance computing platforms, its use on low-power wearable devices is limited by resource constraints. In this paper, we propose a deep learning methodology, which combines features learned from inertial sensor data together with complementary information from a set of shallow features to enable accurate and real-time activity classification. The design of this combined method aims to overcome some of the limitations present in a typical deep learning framework where on-node computation is required. To optimize the proposed method for real-time on-node computation, spectral domain preprocessing is used before the data are passed onto the deep learning framework. The classification accuracy of our proposed deep learning approach is evaluated against state-of-the-art methods using both laboratory and real world activity datasets. Our results show the validity of the approach on different human activity datasets, outperforming other methods, including the two methods used within our combined pipeline. We also demonstrate that the computation times for the proposed method are consistent with the constraints of real-time on-node processing on smartphones and a wearable sensor platform.

Wearable Sensing for Solid Biomechanics: A Review

Understanding the solid biomechanics of the human body is important to the study of structure and function of the body, which can have a range of applications in health care, sport, well-being, and workflow analysis. Conventional laboratory-based biomechanical analysis systems and observation-based tests are designed only to capture brief snapshots of the mechanics of movement. With recent developments in wearable sensing technologies, biomechanical analysis can be conducted in less-constrained environments, thus allowing continuous monitoring and analysis beyond laboratory settings. In this paper, we review the current research in wearable sensing technologies for biomechanical analysis, focusing on sensing and analytics that enable continuous, long-term monitoring of kinematics and kinetics in a free-living environment. The main technical challenges, including measurement drift, external interferences, nonlinear sensor properties, sensor placement, and muscle variations, that can affect the accuracy and robustness of existing methods and different methods for reducing the impact of these sources of errors are described in this paper. Recent developments in motion estimation in kinematics, mobile force sensing in kinematics, sensor reduction for electromyography, and the future direction of sensing for biomechanics are also discussed.

Gait Parameter Estimation From a Miniaturized Ear-Worn Sensor Using Singular Spectrum Analysis and Longest Common Subsequence

This paper presents a new approach to gait analysis and parameter estimation from a single miniaturized ear-worn sensor embedded with a triaxial accelerometer. Singular spectrum analysis combined with the longest common subsequence algorithm has been used as a basis for gait parameter estimation. It incorporates information from all axes of the accelerometer to estimate parameters including swing, stance, and stride times. Rather than only using local features of the raw signals, the periodicity of the signals is also taken into account. The hypotheses tested by this study include: 1) how accurate is the ear-worn sensor in terms of gait parameter extraction compared to the use of an instrumented treadmill; 2) does the ear-worn sensor provide a feasible option for assessment and quantification of gait pattern changes. Key gait events for normal subjects such as heel contact and toe off are validated with a high-speed camera, as well as a force-plate instrumented treadmill. Ten healthy adults walked for 20 min on a treadmill with an increasing incline of 2% every 2 min. The upper and lower limits of the absolute errors using 95% confidence intervals for swing, stance, and stride times were obtained as 35.5 ±3.99 ms, 36.9 ±3.84 ms, and 17.9 ±2.29 ms, respectively.

Deep learning for human activity recognition: A resource efficient implementation on low-power devices

Human Activity Recognition provides valuable contextual information for wellbeing, healthcare, and sport applications. Over the past decades, many machine learning approaches have been proposed to identify activities from inertial sensor data for specific applications. Most methods, however, are designed for offline processing rather than processing on the sensor node. In this paper, a human activity recognition technique based on a deep learning methodology is designed to enable accurate and real-time classification for low-power wearable devices. To obtain invariance against changes in sensor orientation, sensor placement, and in sensor acquisition rates, we design a feature generation process that is applied to the spectral domain of the inertial data. Specifically, the proposed method uses sums of temporal convolutions of the transformed input. Accuracy of the proposed approach is evaluated against the current state-of-the-art methods using both laboratory and real world activity datasets. A systematic analysis of the feature generation parameters and a comparison of activity recognition computation times on mobile devices and sensor nodes are also presented.

Elderly Risk Assessment of Falls with BSN

Due to the natural aging process, the risks associated with falling can increase significantly. For the elderly, this usually marks a rapid deterioration of their health. While there are identified strategies that can be adopted to reduce the number of falls, it is still not possible to prevent all falls. Clinically, the Tinetti Gait and Balance Assessment has been widely used to assess the risk of falls in elderly by examining balance and gait. This paper presents our initial results of using an ear-worn BSN sensor to detect aspects of the Tinetti Gait and Balance Assessment to predict the risk of falls compared to a healthy control cohort. For this study, data was collected from a control cohort of 12 healthy volunteers and a cohort of 16 elderly fallers of varying degrees of risk. The results derived have shown that it is possible to directly detect some aspects of the Tinetti Gait and Balance Assessment and the Timed Up and Go test, demonstrating the potential value of using the platform for continuous assessment in a home environment.

Forearm functional movement recognition using spare channel surface electromyography

Myoelectric signal analysis provides insight into neural control during muscle contraction and it has been widely used to identify the intention of performing different movements for patients with disabilities. Previous studies have demonstrated that detailed neural control information could be extracted from high-density surface electromyography (EMG) signals. However, this imposes practical constraints for routine applications. In this paper, we present an analysis framework using low-density EMG with example experiments demonstrating the control of forearm functional movement Eight channel surface EMG signals are used with subjects performing 6 different forearm and hand movements. Data analysis consisting of feature selection and pattern classification based on KNN, linear discriminant analysis and support vector machine is then performed. High classification accuracy has been achieved for all the subjects, illustrating the practical value of the method proposed.

Gait Analysis From a Single Ear-Worn Sensor: Reliability and Clinical Evaluation for Orthopaedic Patients

Objective assessment of detailed gait patterns after orthopaedic surgery is important for post-surgical follow-up and rehabilitation. The purpose of this paper is to assess the use of a single ear-worn sensor for clinical gait analysis. A reliability measure is devised for indicating the confidence level of the estimated gait events, allowing it to be used in free-walking environments and for facilitating clinical assessment of orthopaedic patients after surgery. Patient groups prior to or following anterior cruciate ligament (ACL) reconstruction and knee replacement were recruited to assess the proposed method. The ability of the sensor for detailed longitudinal analysis is demonstrated with a group of patients after lower limb reconstruction by considering parameters such as temporal and force-related gait asymmetry derived from gait events. The results suggest that the ear-worn sensor can be used for objective gait assessments of orthopaedic patients without the requirement and expense of an elaborate laboratory setup for gait analysis. It significantly simplifies the monitoring protocol and opens the possibilities for home-based remote patient assessment.

Motion Reconstruction from Sparse Accelerometer Data Using PLSR

Detailed motion reconstruction is a prerequisite of biomotion analysis and physical function assessment for a variety of scenarios. For example, biomechanical analysis can be used to assess physical activity to diagnose pathological conditions, to provide an objective measure of biomechanics for peri-operative care, and to monitor patients with mobility issues. Unfortunately, current motion capture systems cannot perform biomechanical analysis continuously in the patients natural environment. In this paper, a pose estimation scheme from a sparse network of accelerometer-based wearable sensors, which does not impose restrictions upon the patients daily life, is presented. In the proposed method, a marker-based motion capture system is used for acquiring the 3D motion data, and partial least squares regression (PLSR) is used to establish the implicit model between 3D body pose and the wearable sensor measurements. A linear constant velocity process model and measurement model are designed and a Kalman filter is then deployed to estimate the posture. Experimental results demonstrate the strength of the technique and how it can be used to estimate detailed 3D motion from a sparse set of sensors.

Markerless motion capture using appearance and inertial data

Current monitoring techniques for biomechanical analysis typically capture a snapshot of the state of the subject due to challenges associated with long-term monitoring. Continuous long-term capture of biomechanics can be used to assess performance in the workplace and rehabilitation at home. Noninvasive motion capture using small low-power wearable sensors and camera systems have been explored, however, drift and occlusions have limited their ability to reliably capture motion over long durations. In this paper, we propose to combine 3D pose estimation from inertial motion capture with 2D pose estimation from vision to obtain more robust posture tracking. To handle the changing appearance of the human body due to pose variations and illumination changes, our implementation is based upon Least Soft-Threshold Squares Tracking. Constraints on the variation of the appearance model and estimated pose from an inertial motion capture system are used to correct 2D and 3D estimates simultaneously. We evaluate the performance of our method with three state-of-the-art trackers, Incremental Visual Tracking, Multiple Instance Learning, and Least Soft-Threshold Squares Tracking. In our experiments, we track the movement of the upper limbs. While the results indicate an improvement in tracking accuracy at some joint locations, they also show that the result can be further improved. Conclusions and further work required to improve our results are discussed.