Detection of tripping gait patterns in the elderly using autoregressive features and support vector machines
Introduction
The estimated average growth rate of the elderly population in 52 major countries will increase to approximately 55% of their current population (Kinsella and Velkoff, 2001) by 2020. There is a significant need for continuous aged healthcare improvements to meet the projected increase in hospitalizations and aged-care institution admittance. One approach to alleviating the strain on present healthcare systems is to minimize the incidence of elderly falls which is the leading cause of unintentional death (11,000 fatalities annually in the US) in the elderly (National Safety Council, 2005). More than half of these falls occur from tripping (SAFE Aging Newsletter, 2005) which incurs high injury costs, for example, direct care for elderly fallers suffering from hip fractures cost governments billions annually (National Safety Council, 2005, Fildes, 1994). In view of this, more effective avenues must be pursued to either prevent or reduce the incidence of elderly falls.
Elderly falls are a multifactorial problem, a result of intrinsic physiological aging and underlying chronic disease interacting with the extrinsic environmental demands. Studies have demonstrated age-related differences in the biomechanical response of individuals following a trip (Pijnappels et al., 2005), however, the intrinsic factors that predispose persons to a higher risk of trip-related falls require further investigation (Van Dieen et al., 2005). Chou and Draganich (1998) have earlier identified lower toe clearance in the trailing foot as the main etiology of tripping when stepping over obstacles in the path of walking. More recent balance control studies have started to employ computational intelligence (CI) techniques (e.g., Hahn et al., 2005) thus motivating our interest in CI techniques for toe clearance studies, in particular the minimum toe clearance (MTC). MTC is defined as the minimum vertical distance between the lowest point on the shoe and the ground during mid-swing (Fig. 1) (Begg et al., 2007). This variable is believed to be a sensitive trip-risk predictor because it quantifies a region of safety when negotiating the walking environment (Winter, 1991, Begg et al., 2007). Prior studies employing statistical analysis have suggested that MTC height remains relatively unchanged (i.e., the differences were not found to be statistically significant) due to pathology or ageing (Ghelsen and Whaley, 1990, Elble et al., 1991, Winter, 1991, Dingwell et al., 1999, Begg et al., 2007).
Recent efforts however, have demonstrated that MTC data should be monitored to detect tripping gait patterns. Begg et al. (2005a) for example, used statistical quantities of MTC histogram plots to classify 20 individuals (10 healthy and 10 pathological) with a support vector machine (SVM) classifier yielding at best 90% leave one out (LOO) accuracies. Khandoker et al. (2006) then applied wavelet decomposition to a series of MTC data treated as a signal. This system used a SVM and required a feature selection algorithm to obtain 100% LOO accuracies from the optimal feature set. In these systems, individuals were required to walk for more than 10 min which even though yielded good detection proved strenuous for elderly individuals and required long periods of data collection. In view of this, we now investigate the design of a detection system which can achieve similar detection accuracies in a smaller number of strides. We hypothesize that explicit MTC modelling is the key to detecting tripping gait patterns in fewer gait cycles, a precursor to faster and more efficient tripping falls detection systems.
In this paper, we propose a two-stage detection system (AR–SVM) composed of an autoregressive (AR) model and a SVM classifier. The AR model is a linear parametric model used to model wide-based stationary random signals and has been limitedly applied to other gait studies such as the design of a postural stability criterion (Vance et al., 1999), shape deformations during gait (Veeraraghavan et al., 2005) and energy transfer (Gider et al., 2005) modelling. To the best of our knowledge, no work has applied the AR model for falls risk prediction. We conjecture that AR model coefficients will represent the average interdependencies between MTC samples, thereby capturing MTC signal variability characteristic of tripping gait patterns. SVMs on the other hand are powerful CI techniques used for classification and function estimation problems (Vapnik, 2000), possessing solid mathematical foundations (Schölkopf et al., 1999, Schölkopf and Smola, 2002) and demonstrating good classification capabilities even when data sets are not large. The central idea behind our proposed system is to first model a set of MTC signals with the AR process and use the AR model coefficients as SVM inputs to detect tripping fall patterns. The system's detection capability is further investigated to determine the minimum strides required for acceptable detection rates .
Section snippets
Participants
MTC gait data for this study were composed of 23 individuals comprising 13 healthy elderly with no falls history and 10 elderly who had more than one fall in the past year. Individuals were selected as being at risk of tripping if they suffered at least one tripping fall during normal walking activity. The tripping characteristics of these 10 fallers and the situations in which they occurred are depicted in Table 1. The healthy group had a mean age of 71.0 years and were on average
AR–SVM: A two-stage falls risk detection system
The proposed AR–SVM system is depicted in Fig. 2 where the system input is the raw MTC signal. The main detection idea is to first model the variability in the subject MTC signal using a linear predictor model estimated by the AR process. The model coefficients capture the MTC signal variability characteristics and are labelled as either being representative of healthy gait or tripping falls gait. These are then used as input features to train a SVM classifier so that it can recognize AR
Experimental methodology
We first derived seven raw data sets; , , , , , and each comprising 23 signals (corresponding to the 23 individuals) constructed from the first th samples of the respective subject's MTC data. The dyadic reduction in MTC signal lengths simulated different walking durations, for example represented approximately a minute's walking. AR models were next computed for all raw data sets over a range of using the Burg function implemented in MATLAB
AR model order selection
Examples of MTC signal plots for healthy and pathological individuals (Fig. 5) indicate larger MTC signal variability in individuals at risk of tripping. A comparison of the logarithm (log) of FPE criterion values across the AR model order, p as depicted in Fig. 6 for , and signal lengths show an almost constant log FPE for smaller model orders and sharp decreases as the order increased. For , the log FPE criterion remained almost constant up to before
Discussion
Contemporary gait analysis using non-specific kinematic variables to detect balance impairments in the elderly have previously required long walking durations (Baker, 2006). We have proposed a two-stage detection system requiring only MTC data during gait and demonstrated improved detection rates compared to previous systems. A major advantage of the proposed model is that it requires significantly smaller number of MTC data compared to other previous models (e.g., Khandoker et al., 2006)
Conclusion
We have proposed a new AR–SVM system for detecting elderly individuals at risk of falling using MTC data. Our system can accurately detect gait patterns of fallers from 32 consecutive strides which is a large improvement over previous methods that required more than 500 strides or at least 10 min of walking. The system is also highly sensitive, providing at least 95% detection accuracies down to 16 strides before suffering performance degradation. These results are encouraging because they
Conflict of interest
None.
References (39)
- et al.
Is low lower-limb kinematic variability always an index of stability?
Gait and Posture
(2007) - et al.
A machine learning approach for automated recognition of movement patterns using basic, kinetic and kinematic gait data
Journal of Biomechanics
(2005) - et al.
Minimum foot clearance during walking: strategies for minimisation of trip-related falls
Gait and Posture
(2007) - et al.
Placing the trailing foot closer to an obstacle reduces flexion of the hip, knee and ankle to increase the risk of tripping
Journal of Biomechanics
(1998) - et al.
Neuropathic gait shows only trends towards increased variability of sagittal plane kinematics during treadmill locomotion
Gait and Posture
(1999) - et al.
Push-off reactions in recovery after tripping discriminate young subjects, older non-fallers and older fallers
Gait and Posture
(2005) - et al.
Age-related intrinsic limitations in preventing a trip and regaining balance after a trip
Safety Science
(2005) Autoregressive model fitting for control
Annals of the Institute of Statistics and Mathematics
(1971)Gait analysis methods in rehabilitation
Journal of Neuro Engineering and Rehabilitation
(2006)- Begg, R., Lai, D., Taylor, S., Palaniswami, M., 2005a. SVM models in the diagnosis of balance impairments. In: 3rd...