您当前的位置: 首页 > 网页快照
Prediction of fall risk among community-dwelling older adults using a wearable system - Scientific Reports
Methods . All methods were performed in accordance with the guidelines of the Declaration of Helsinki, and approved by the Institutional Review Board (or Ethics Committee) of Virginia Tech Institutional review Board (protocol code 11-1088 and 10-04-2013 as the date of approval). The study was conducted in four different community centers in Northern Virginia (Dale City, Woodbridge, Leesburg, and Manassas), using the same set of instruments i.e., Inertial Measurement Unit (IMU) on different days. All participants provided written consent before beginning the study. Participants wore comfortable attire and had to perform a 10?m walk. Ten-meter walk was chosen based on the assumption that at least 10?s of continuous walking activity can be detected during the activity of daily living. The participants were instructed to walk at their normal speed. All participants stood behind the start line quietly for 5?s, the experimenter started data collection and gave an auditory signal “GO” to the participant to start walking at their own normal walking speed. When the participant crossed the 10?m finish line, the participant stood quietly for 5?s until they were asked to come back. The walking trial was repeated twice for all participants. The sampling rate was 100?Hz. Rest of 3?min was provided between each measurement. Participants . A total of 171 older community-dwelling volunteers (age 56–90?years; mean age 74.3?±?7.6?years) participated in this study. All participants were asked to walk 10?m with one wireless inertial sensor affixed at sternum level (Fig.? 5 ). For the model development, we evaluated 127 community-dwelling older adults’ trunk kinematics using a wearable sensor during walking to unobtrusively assess fall risks that may be amenable to predicting fallers and non-fallers using linear and nonlinear measures. Among the 127 participants, there were 101 non-fallers and 26 fallers. The predictive model accuracy was tested on 44 community-dwelling individuals with six months follow up of their fall history (35 non-fallers vs. 9 fallers). Participants’ history of falls had been recorded for the last 2?years, with emphasis on the frequency and characteristics of falls. Fall history was obtained by self-report, and any subject with at least two falls in the prior year was classified as a faller and the others as non-faller. Figure 5 Placement of a wearable IMU system. Full size image Gait feature extractions . Trunk accelerations in the anterior–posterior (AP), medio-lateral (ML), and vertical (V) directions were analyzed. Gait event times were identified using an inertial measurement unit (IMU) positioned over the sternum 37 . A modified continuous wavelet transform (CWT) method was utilized as a gait detection algorithm 38 . The wavelet transform supports time–frequency decomposition of non-stationary signals and does not require preprocessing of the signal, making it ideally suited for a peak detection algorithm 38 , 39 , 40 . The resultant acceleration, a signal invariant to axis alignment, was analyzed to mitigate any alignment errors reliant on IMU placement. Furthermore, due to the placement of the inertial sensor, the gaussian (gaus1) mother wavelet was deemed inappropriate for the inertial data 38 . Instead, a symlet (sym4) mother wavelet with an order of 4 and a scale between 35 and 70, was employed over the resultant acceleration signal to detect the instant events 15 . Heel contacts (HC) were identified as the maxima of the CWT differentiated signal (Fig.? 6 ). The toe-off (TO) events were processed by a windowing technique in which the HC points and the subsequent zero crossings of the CWT differentiated signal determined an appropriate window size where the instant of the first minima in the AP acceleration signal was considered as a TO event (Fig.? 6 ) 41 , 42 . As per the placement of the inertial sensor and the extracted resultant acceleration, the CWT method previously employed, in which the maxima of a further CWT differentiated signal was considered the final contact event, could not be relied upon to determine the TO time. Figure 6 Detection of HC events using the CWT differentiation method. Peaks (blue) equate to HC events; the local minima in the AP acceleration (red) equate to TO events. Full size image Moreover, because of the inherent gait deficiencies associated with the community-dwelling older adults and the intermittent “shuffling of gait,” a window detection method was better suited for the extracted signal 41 . Finally, the right and left HC events were designated by the sign of the vertical angular velocity at the instant of the first HC in which every other HC equated to a stride 38 . The signal was preprocessed with a 4th order low pass Butterworth filter and a cutoff frequency of 2?Hz 42 , 43 . Trunk acceleration-based measures of gait spatiotemporal and variability parameters have been used extensively to identify gait characteristics in both healthy and pathologic populations and are often used to quantify fall risk 38 , 41 , 43 . Gait variability was assessed by the RMS of trunk acceleration components—the anteroposterior (AP), mediolateral (ML), and vertical (V) directions—and by statistical measures of variability from temporal gait parameters: Standard deviation (SD) and coefficient of variation (CV). CV denotes the variability of a specific gait parameter normalized to its mean value; it is represented as a percentage (CV?=?SD/mean?×?100). The first and last stride during the initiation and termination of gait were excluded from analysis; the local average and the local SD of each time series was computed for each spatiotemporal parameter, as well. Table 4 provides further operational definitions for each parameter. The normalized RMS of trunk acceleration was implemented to distinguish the proportion of trunk acceleration variability in a particular direction compared to the total acceleration variability. The RMS norm is a normalization method to mitigate the correlation with walking speed 44 . To compute the RMS norm of the trunk acceleration, the RMS of each acceleration component is divided by the vector norm of all the components (AP, ML, V). Furthermore, scaling behavior of walking patterns were assessed. Harmonic ratio (HR) was computed by decomposing the AP, ML and V acceleration signals into harmonics using discrete Fourier transformation 45 . For HR, the summed amplitudes of the first 10 even harmonics were divided by the sum of amplitudes of the first 10 odd harmonics for the AP and V directions and vice versa for ML acceleration. Since AP and V have two periods every stride, showing dominance of second harmonics and subsequent even harmonics, whereas, ML accelerations have only one period per stride, reflecting a dominance of the first and subsequent odd harmonics 45 . Higher HR is an indication of increased smoothness of gait. Approximate entropy (ApEn) quantifies the ensemble amount of randomness, or irregularity 46 , contained in a physiological time series. It uses a moving window procedure to determine the probability that short sequences of data points are repeated (within a defined tolerance) throughout the time-series. Here, we use ApEn to quantify the regularity of 3-D trunk accelerations during walking in community-dwelling older adults. Previous research reports that ApEn can be used to detect subtle changes in signal variability that are averaged out in traditional statistical measures of gait stability 33 , 47 . The algorithm for estimating ApEn was first reported by Pincus 48 . Sample entropy (SaEn) indexes the regularity of a time series by calculating the probability that having repeated itself for a window length? m , it will remain similar for? m ?+?1 data points, excluding any self-matches and within a matching tolerance r . Greater SaEn values delineate irregularity, in which a set of similar points are considered unique as they will likely not be followed by a similar set of matching points within a specified tolerance. Higher values are considered part of a healthy, robust system able to adapt to challenges and unexpected perturbations. Lower values of SaEn are associated with higher regularity of the time series, in which there is a greater likelihood that datasets of matching epochs in a time series will be followed by another match within a specified tolerance. Lower values denote a possible rigid, disease state—unable to adapt to challenges or walking perturbations. SaEn was computed with the resultant acceleration time series. Parameters? m ?and? r ?were chosen accordingly obtaining? m ?=?2 and? r ?=?0.25 for both directions. Multiscale entropy (MSE) is a regularity measure that quantifies the information content of postural/gait fluctuations over a range of physiologically relevant time scales while sample entropy is computed for every consecutive coarse-grained time series. The entropy values are then plotted as a function of the time scales in which the area under the curve reveals the signal’s complexity index. A complex signal is associated with a time evolution with a rich structure on multiple scales. For white noise, which is irregular on small time scales but not structurally complex, the entropy decreases for larger time scales. For a complex signal, such as pink 1/f noise, the entropy remains high on different scales. For the computation of MSE the input parameters m?=?2 and r?=?0.25 were chosen similar to the SaEn algorithm. Recurrence quantitative analysis (RQA): recurrence quantitative analysis is a nonlinear analysis technique 49 , 50 recently used in gait signal analysis 50 . The local recurrence of data points during gait in the reconstructed state space allows RQA to quantify deterministic structures and associated non stationarities 51 . In this study, an embedding dimension of 5 and a delay of 10 was chosen 50 . The recurrence plot was made with radius of 40% of the maximum distance and cells below this threshold were identified as recurrent points. RQA measures such as entropy, recurrence, determinism, and MaxLine were computed for this study. All gait descriptors were calculated using custom MATLAB scripts. A list of both linear and nonlinear gait variability descriptors used in this study is provided below (Table 5 ). Table 5 List of 58 linear and nonlinear gait variability descriptors used in fall classification. Full size table Random forest predictive model . In this study, we conducted three experiments for predictive model development and validation using random forest (RF) 53 , a well-studied supervised machine learning algorithm as the classifier. RF creates the forest with a number of trees. With more trees in the forest, it is more likely to provide robust predictions with high accuracy 54 . Each decision tree is created from randomly chosen features and test-data participants and utilizing a set of rules to predict fall risk. Finally, votes are calculated for each predicted output from decision trees, and majority voting is considered a final prediction. Some advantages of RF are that it can handle missing values 55 , and it provides robust prediction without overfitting 54 . As seen in Fig.? 7 , Experiment I explores the applicability of RF on all 58 gait parameters (both linear and nonlinear—Table 5 ) trained on 127 participants and blind tested on the 44 subjects. In Experiment II, our focus was to employ two feature engineering steps to improve the RF classifier. The first step was unsupervised feature selection. Two sample t test was applied to evaluate the potential risk of source discrepancy in gait parameters using a training dataset. Specifically, the trained data was randomly split into two groups, and the p value for each variable was used to evaluate the risk factor of each gait parameters. This procedure was repeated n (=?1000) times. The averaged p values represented the ranked potential risk of source discrepancy for each predictor. The second step applied principal component analysis (PCA) to orthogonalize the original features into less correlated principal components (PCs). Because the gait features are derived from 10?m walk with dynamic motion, inherently, the features may have similar characteristics. One of the limitations of having highly correlated features is the trained RF may be destabilized which will weaken its clinical value. We hypothesized PCA approach may address this issue. Usually, a few PCs may sufficiently account for most of variability in the original feature space. In Experiment II, PCs capturing 99% of the variability in the original dataset were derived for the RF classifier. Experiment I and II used linear and nonlinear features independently to assess the contributions from feature engineering. Experiment III was then conducted on RF model in conjunction with feature engineering using combined linear and nonlinear features. Figure 7 Workflow of three designed experiments (OOB: Out of Bag RF strategy). Full size image .
From:
监测目标主题     
(1)  
(1)  
系统抽取主题     
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)  
(1)