Intra-articular corticosteroid knee injections (ICIs) were used as a tool to determine the sensitivity of wearable inertial sensors and machine learning algorithms in identifying meaningful changes in gait patterns amidst day-to-day fluctuations in out-of-laboratory gait. Specifically, three overarching aims were proposed; I) Determine if three gait trials could define an everyday typical gait pattern, II) investigate if post-injection atypical strides are significantly different from pre-injection atypical strides and III) explore the relationship between changes in pain and atypical strides. Nine knee OA patients (7M/2F) were recruited from St. Joseph’s Healthcare Hamilton. Participants completed a total of four walking trials prior to the ICI and three following. Participants were fitted with two wearable sensors on each shank just below the knee, and one sensor on the lower back during every trial. Data from these sensors were processed to train and test a one-class support vector machine (OCSVM). Individual gait models were created based on three out of the four pre-injection trials. Each trained model was tested on a withheld pre-injection trial and three post-injection trials to determine the number of typical and atypical gait cycles. Self-reported pain was analyzed throughout the study and compared to the percent of atypical strides seen during each walk. It was found that three gait trials could not define a typical gait model and that post-injection atypical strides were not significantly different from with-held pre-injection atypical strides. Finally, large variations and fluctuations in self-reported pain were observed on a week-to-week basis, which were not significantly correlated to atypical strides observed. This study was the first to investigate the sensitivity of wearable inertial sensors and machine learning algorithms to detect changes in real-world gait patterns and provides foundational work for using wearable sensors to monitor and triage knee OA patients. / Thesis / Master of Science (MSc)

Walking is the most common form of human locomotion and the systematic study thereof is known as gait analysis. Measurement and assessment thereof have application in many fields including clinical diagnosis, rehabilitation and biomechanics. The process of gait evaluation is typically done using an optical motion analysis system combined with stationary force platforms. This is considered the gold standard, but unfortunately, has several drawbacks. It is expensive, requires dedicated laboratories with spatial restrictions, calls for lengthy set up and post-processing times and cannot be used in 'real-world' environments. Alternative systems based on wearable sensors have been developed to overcome these limitations. The Council for Scientific and Industrial Research (CSIR) has therefore developed a prototype wearable sensor unit consisting of an inertial measurement unit (IMU). The objective of the current study is, therefore, to advance the prototype to a wearable multi-sensor system for quantitative gait analysis. The focus is on the development of the pre- and post-processing algorithms and methods used to transform the measurements into interpretable information. The focus outlined includes establishing techniques for synchronising the data from the sensors offline, pre-processing the signals, developing algorithms for stride and gait event detection, selecting an appropriate gait model and defining methods for estimating gait parameters. The determined parameters were the spatio-temporal and joint kinematics (hip, knee and ankle). The algorithms and new system were validated against the Vicon motion capture system through gait analyses. The twenty able-bodied volunteers that took part were required to walk across the laboratory six times at three self-selected walking speeds (slow, normal and fast). For the sake of simplicity and due to various limitations, only data in the sagittal plane of the right lower limb of each volunteer was used to validate the wearable system and associated algorithms. The results obtained were then evaluated against several validation criteria. The absolute mean difference between the estimated timing of detected gait events of the two systems was consistently small (between 0.021 and 7.25% of the gait cycle overall). The spatially dependent parameters, stride length and walking speed, had significant maximum mean absolute percentage errors (31.9 and 34.5% respectively), but with little variation. Excluding outliers, that of the temporal parameters, stride time and cadence, was significantly lower (5.7 and 5.6% respectively). The kinematic results were substantially comparable with a minimum correlation co-efficient of 0.86 and a maximum RMSE of 7.8 degrees with little variation implying repeatability. Although there were some discrepancies between the outputs, the wearable sensor system and its corresponding algorithms were considered feasible and potentially beneficial to developing countries like South Africa. Recommendations for future work include synchronising data between the wearable and reference system for stride-to-stride comparisons and validating algorithms using a known reliable wearable system. / Dissertation (MEng)--University of Pretoria, 2017. / Mechanical and Aeronautical Engineering / MEng / Unrestricted

Wearable sensor systems composed of small and light sensing nodes have the potential to revolutionise healthcare. While uptake has increased over time in a variety of application areas, it has been slowed by problems such as lack of infrastructure and the functional capabilities of the systems themselves. An important application of wearable sensors is the detection of falls, particularly for elderly or otherwise vulnerable people. However, existing solutions do not provide the detection accuracy required for the technology to gain the trust of medical professionals. This thesis aims to improve the state of the art in automated human fall detection algorithms through the use of a machine learning based algorithm combined with novel data annotation and feature extraction methods. Most wearable fall detection algorithms are based on thresholds set by observational analysis for various fall types. However, such algorithms do not generalise well for unseen datasets. This has thus led to many fall detection systems with claims of high performance but with high rates of False Positive and False Negative when evaluated on unseen datasets. A more appropriate approach, as proposed in this thesis, is a machine learning based algorithm for fall detection. The work in this thesis uses a C4.5 Decision Tree algorithm and computes input features based on three fall stages: pre-impact, impact and post-impact. By computing features based on these three fall stages, the fall detection algorithm can learn patterns unique to falls. In total, thirteen features were selected across the three fall stages out of an original set of twenty-eight features. Further to the identification of fall stages and selection of appropriate features, an annotation technique named micro-annotation is proposed that resolves annotation-related ambiguities in the evaluation of fall detection algorithms. Further analysis on factors that can impact the performance of a machine learning based algorithm were investigated. The analysis defines a design space which serves as a guideline for a machine learning based fall detection algorithm. The factors investigated include sampling frequency, the number of subjects used for training, and sensor location. The optimal values were found to be10Hz, 10 training subjects, and a single sensor mounted on the chest. Protocols for falls and Activities of Daily Living (ADL) were designed such that the developed algorithms are able to cope under a variety of real world activities and events. A total of 50 subjects were recruited to participate in the data gathering exercise. Four common types of falls in the sagittal and coronal planes were simulated by the volunteers; and falls in the sagittal plane were additionally induced by applying a lateral force to blindfolded volunteers. The algorithm was evaluated based on leave one subject out cross validation in order to determine its ability to generalise to unseen subjects. The current state of the art in the literature shows fall detectors with an F-measure below 90%. The commercial Tynetec fall detector provided an F-measure of only 50% when evaluated here. Overall, the fall detection algorithm using the proposed micro-annotation technique and fall stage features provides an F-measure of 93% at 10Hz, exceeding the performance provided by the current state of the art.

The use of wearable sensors in consumer health and medicine is a rapidly developing topic of interest. The main purpose of the series of studies in this thesis is to identify novel uses of technology that can provide clinicians and scientists clinically feasible, low cost approaches to obtain meaningful information about functional limb symmetry in patients with knee injuries. In Study 1, individuals undergoing knee surgery were evaluated as they walked and stepped down onto a force platform in a manner similar to how one would step off a curb to cross a street. When subjects stepped onto their uninvolved leg, peak vertical ground reaction force was greater and occurred earlier than when stepping onto their involved leg. Asymmetries were greater in those with higher quadriceps neuromuscular impairment. In Study 2, the reliability and validity of using wearable accelerometer sensors was evaluated for estimating single leg vertical hop height in healthy people and individuals after ACL reconstruction surgery. The reliability and concurrent validity of using accelerometers to estimate single leg hop height were excellent, and were similar for healthy and ACL-reconstructed subjects. Error for this method was low, in particular when the accelerometer was worn at the lower leg. Asymmetry in hop height was greater in those with higher quadriceps neuromuscular impairment. In Study 3, wearable accelerometers were compared to a system of motion capture cameras and force platform as a method to assess functional movement asymmetry in healthy people and individuals after ACL reconstruction. While walking and stepping down, accelerometers worn at the waist were able to detect underlying movement asymmetry when it exists in people after ACL reconstruction. Acceleration at the waist was strongly associated with vertical ground reaction force and moderately associated with knee extension moments. Collectively, these studies provide evidence that functional movement symmetry can be measured with simple, inexpensive methods that can be used in a variety of clinical or field-based settings.

Wearable sensors are at the heart of the digital health revolution. Integral to the use of these sensors for monitoring conditions impacting balance and mobility are accurate estimates of joint angles. To this end a simple and novel method of estimating hip joint angles from small wearable magnetic and inertial sensors is proposed and its performance is established relative to optical motion capture in a sample of human subjects. Improving upon previous work, this approach does not require precise sensor placement or specific calibration motions, thereby easing deployment outside of the research laboratory. Specific innovations include the determination of sensor to segment rotations based on functionally determined joint centers, and the development of a novel filtering algorithm for estimating the relative orientation of adjacent body segments. Hip joint angles and range of motion determined from the proposed approach and an existing method are compared to those from an optical motion capture system during walking at a variety of speeds and tasks designed to exercise the hip through its full range of motion. Results show that the proposed approach estimates flexion/extension angle more accurately (RMSE from 7.08 to 7.29 deg) than the existing method (RMSE from 11.64 deg to 14.33 deg), with similar performance for the other anatomical axes. Agreement of each method with optical motion capture is further characterized by considering correlation and regression analyses. Mean ranges of motion for the proposed method are not largely different from those reported by motion capture, and showed similar values to the existing method. Results indicate that this algorithm provides a promising approach for estimating hip joint angles using wearable inertial sensors, and would allow for use outside of constrained research laboratories.

Chronic stress is endemic to modern society. If we could monitor our mental state, we may be able to develop insights about how we respond to stress. However, it is unfeasible to continuously annotate stress levels all the time. In the studies conducted for this dissertation, a minimally-invasive wearable sensor platform and physiological data processing methods were developed to analyze a number of physiological correlates of mental stress. We present a minimally obtrusive wearable sensor system that incorporates embedded and wireless communication technologies. The system is designed such that it provides a balance between data collection and user comfort. The system records the following stress related physiological and contextual variables: heart rate variability (HRV), respiratory activity, electrodermal activity (EDA), electromyography (EMG), body acceleration, and geographical location. We assume that if the respiratory influences on HRV can be removed, the residual HRV will be more salient to stress in comparison with raw HRV. We develop three signal processing methods to separate HRV into a respiration influenced and residual HRV. The first method consists of estimating respiration-induced portion of HRV using a linear system identification method (autoregressive moving average model with exogenous inputs). The second method consists of decomposing HRV into respiration-induced principal dynamic mode and residual using nonlinear dynamics decomposition method (principal dynamic mode analysis). The third method consists of splitting HRV into respiration-induced power spectrum and residual in frequency domain using spectral weighting method. These methods were validated on a binary discrimination problem of two psychophysiological conditions: mental stress and relaxation. The linear system identification method, nonlinear dynamics decomposition method, and spectral weighting method classified stress and relaxation conditions at 85.2 %, 89.2 %, and 81.5 % respectively. When tonic and phasic EDA features were combined with the linear system identification method, the nonlinear dynamics decomposition method, and the spectral weighting method, the average classification rates were increased to 90.4 %, 93.2 %, and 88.1 % respectively. To evaluate the developed wearable sensors and signal processing methods on multiple subjects, we performed case studies. In the first study, we performed experiments in a laboratory setting. We used the wearable sensors and signal processing methods to discriminate between stress and relaxation conditions. We achieved 81 % average classification rate in the first case study. In the second study, we performed experiments to detect stress in ambulatory settings. We collected data from the subjects who wore the sensors during regular daily activities. Relaxation and stress conditions were allocated during daily activities. We achieved a 72 % average classification rate in ambulatory settings. Together, the results show achievements in recognizing stress from wearable sensors in constrained and ambulatory conditions. The best results for stress detection were achieved by removing respiratory influence from HRV and combining features from EDA.

abstract: There are many inconsistencies in the literature regarding how to estimate the Lyapunov Exponent (LyE) for gait. In the last decade, many papers have been published using Lyapunov Exponents to determine differences between young healthy and elderly adults and healthy and frail older adults. However, the differences in methodologies of data collection, input parameters, and algorithms used for the LyE calculation has led to conflicting numerical values for the literature to build upon. Without a unified methodology for calculating the LyE, researchers can only look at the trends found in studies. For instance, LyE is generally lower for young adults compared to elderly adults, but these values cannot be correlated across studies to create a classifier for individuals that are healthy or at-risk of falling. These issues could potentially be solved by standardizing the process of computing the LyE. This dissertation examined several hurdles that must be overcome to create a standardized method of calculating the LyE for gait data when collected with an accelerometer. In each of the following investigations, both the Rosenstein et al. and Wolf et al. algorithms as well as three normalization methods were applied in order to understand the extent at which these factors affect the LyE. First, the a priori parameters of time delay and embedding dimension which are required for phase space reconstruction were investigated. This study found that the time delay can be standardized to a value of 10 and that an embedding dimension of 5 or 7 should be used for the Rosenstein and Wolf algorithm respectively. Next, the effect of data length on the LyE was examined using 30 to 1300 strides of gait data. This analysis found that comparisons across papers are only possible when similar amounts of data are used but comparing across normalization methods is not recommended. And finally, the reliability and minimum required number of strides for each of the 6 algorithm-normalization method combinations in both young healthy and elderly adults was evaluated. This research found that the Rosenstein algorithm was more reliable and required fewer strides for the calculation of the LyE for an accelerometer. / Dissertation/Thesis / Appendix A / Doctoral Dissertation Biomedical Engineering 2019

No description available.

No description available.

In this study, inertial measurement units (IMUs) were used to train a random forest classifier to correctly classify different jump types in volleyball. Athlete motion data were collected in a controlled setting using three IMUs, one on the waist and one on each ankle. There were 11 participants who at the time played volleyball at the collegiate level in the United States, seven male and four female. Each performed the same number of jumps across the eight jump types--five BASIC jumps and three each of the other seven--resulting in 26 jumps per subject for a total of 286. The data were processed using a max-bin method and trained using a leave-one-out cross-validation method to produce a classifier that can determine jump type with an accuracy of 0.967, as measured by an F1-score.