Stumbling prediction based on plantar pressure distribution

Abstract

BACKGROUND:

Stumbles are common accidents that can result in falls and serious injuries, particularly in the workplace where back and forth movements are involved and in offices where high heels are imperative. Currently, the characteristics of plantar pressure during a stumble and the differences between stumbling and a normal gait remain unclear.

OBJECTIVE:

This paper is aimed at providing insights into the feasibility of the data mining technique for interventions in stumble-related occupational safety issues.

METHODS:

The characteristics of plantar pressure distribution during stumbling and normal gait were analyzed by using the power spectrum density (PSD) and the Support Vector Machine (SVM). The PSD, a novel pattern recognition feature, was used to mathematically describe the image signal. The SVM, a powerful data mining technique, was used as the classifier to recognize a stumble. Dynamic plantar pressures were measured from twelve healthy participants as they walked.

RESULTS:

The plantar pressures of the stumbling gaits had significantly different patterns compared to the normal ones, from either a qualitative or quantitative perspective. The mean recognition accuracy of the proposed method reached 96.7%.

CONCLUSIONS:

This study helps better understand stumbles and provides a theoretical basis for stumble-related occupational injuries. In addition, the stumble is the precursor of a fall and the research on stumble recognition would be of value to predict and provide warnings of falls and to design anti-fall devices for potential victims.

Keywords

Gait recognition SVM power spectrum density artificial intelligence

1 Introduction

Physical workers, particularly those whose responsibilities require them to walk back and forth, and white-collar workers, who must wear high heels in the office, often suffer from stumbles and consequently falling. According to the National Information Services and Systems, a nationwide information acquisition system based on outpatient/emergency injuries in China, fall patients accounted for the greatest proportion of unintentional injuries, approximately 31.2% ∼36.1% [1]. Slips, trips and falls can lead to serious injuries, and even to death, particularly when not noticed in advance. A serious consequence of a fall is the ‘long-lie’, which is identified as involuntarily remaining on the ground for an hour or more following a fall [2]. The ‘long-lie’ is a common occurrence and has shown that many victims lack the ability to rise up by themselves, even from non-injurious falls, and thus remain lying on the ground for longer than an hour, which further compounds the severity [3]. Therefore, the consequences of a fall can be rather serious and fatal. Stumble-related occupational injuries and accidents have brought about great challenges and burdens to the economy. The earlier the fall that is reported, the lower the rate of morbidity-mortality will be [4]. Functional fall training is a new promising method for fall prevention. This intervention can increase participant adherence by decreasing the amount of total training time required to improve balance recovery skills [5]. Falls have been identified to be the first cause of injury-related declines in health which consequently lead to higher levels of morbidity and mortality in the elderly [6]. The aging process and related chronic diseases that affect older adults lead to balance disorders in this population [7]. Therefore, studies have evaluated the level of perceived balance and pain levels in older subjects [8]. Others have studied whether there is a difference in foot clearances between different steps in the short-term stairs of the elderly, thereby reducing the occurrence of falls [9]. The stumble is the precursor of a fall, and the research on stumble recognition would be of value in recognizing and warning of falls and in designing anti-fall devices for potential victims.

Another issue regarding stumbling is recover a normal gait. There are several recovery strategies during a stumble: short swing blockage [10], treadmill speed reversal [11], lifted on the gait track [12] or dropped on the treadmill band [13]. There are two categories of stumbling reactions described [14]. The first category is the elevating strategy, more frequent in early swing perturbations, and consists of an elevation of the swing limb to overtake the obstacle [12]. The longer step length will lead to the toe clearance increasing. The second category is the lowering strategy, which consists of bringing the foot to the ground as quickly as possible [12]. Different people have different responses to stumbling. The performance of the recovery is measured by its time duration and energy cost [15].

Some studies used image processing, accelerometers, and inertial measurement units to detect falls [16 –18]; however, such a motion analysis system is not available in occupational or community settings and is not appropriate for giving early warning. Stumbling is one of the abnormal gaits and plantar pressure is an essential feature of gait analysis. In addition, plantar pressure measurement equipment is not intrusive, and the plantar pressure data are easy to collect. However, few researchers have investigated the characteristics of plantar pressure during a stumble. The proposed study is anticipated to provide a theoretical basis of plantar pressure during stumbling. In this study, peak pressure, mean pressure and the power spectrum density of each frame were used. Our method can be further integrated into wearable devices and predict gait trends based on plantar pressure data, thus allowing early warnings of possible falls and providing the ability avoid falls.

2 Method

2.1 Subjects

Twelve healthy male participants (age 23.9±2.3 years, height 174.4±2.3 cm, weight 63.9±8.5 kg, and body mass index 21.0±2.8 kg/m²) without discernible gait abnormalities were recruited for this research. All the participants were from the University of XX (anonymous for peer review). Compared with healthy female adults, the bone strength of the healthy male adults’ proximal femur is higher, which can resist the impact force caused by falls and reduce the chance of proximal femoral fractures. Therefore, we chose male adults as the participants. The purpose of the research was explained to each participant before they were asked to give their consent to take part in the experiment and sign off on it medically and ethically approval.

2.2 Equipment

We used the F-scan VersaTek system produced by the Tekscan Company to collect the plantar pressure data. The F-scan VersaTek system can measure and display the pressure between the soles and the shoes in real time (Fig. 1). The sampling frequency was 50 frames per second (up to 100 Hz). To simplify the experimental procedure and minimize the effect of inter sensor variation, the experiment shoes, which can fit all participants, was chosen. The pressure measurement insoles were trimmed to fit our shoe size. The calibration of the insoles was performed according to the instructions in the Tekscan user manual [19].

Fig.1

Experiment facility.

2.3 Protocol

Before the formal experiment, the participants wore the experiment shoes and were asked to walk at least ten minutes to ensure familiarity with the equipment. After calibration, the participants were asked to wear the experiment shoes and to walk along a slope with an angle of 10° at their normal speed for three minutes. The reason why we chose a slope rather than flat ground was that people are more prone to stumble on a slope [20]. We chose the intermediate 3s as the target experiment time; thus, each trial has a frame sequence of 3×50 frames and contains 2-3 gait cycles. We preset a small obstacle on the slope where the subjects were asked to walk so that during this procedure the stumble would be triggered. We asked each participant to walk 10 trials at their normal speed (5 trials normal and 5 trials abnormal, in which normal trials do not trigger the stumble, while abnormal trials need to trigger the stumble). The data we finally obtained were 10 (trials)×12 (participants) = 120 trials.

2.4 Analysis

Gait cycle was identified according to the periodic variety of plantar pressure [21]. Plantar pressure distribution shows similarity at the same gait phase among continuous walking. Gait cycle in this research was split into five phases: swing phase, initial contact phase, forefoot contact phase, foot flat phase and forefoot push off contact phase [22], as shown in Fig. 2.

Fig.2

Five distinct phases of foot roll-over.

Feature extraction is an important step before data mining. In the statistic characteristics of images, image signal is considered as a random signal. The mathematical descriptions of a random signal are the distribution function of the amplitude or the phase, the probability density function or a series of central moments and power spectrum, etc. The power spectrum density (PSD) was chosen as our feature because PSD has a firm mathematical foundation, a wide application scope and good robustness. Since each frame has a PSD value, it can represent the information of the whole plantar pressure image.A two-dimensional pressure image can be considered a two-dimensional discrete function g (h, v) consisting of H × V pixels. H is the number of the horizontal pixels, and V is the number of the vertical pixels. The PSD of an image is defined as: $e = (\sum_{n = 0}^{N - 1} \sum_{m = 0}^{M - 1} | g (h, v) |) / s$ (1) where g (h, v) is the color value of a pixel whose coordinate is (h, v) and s = H × V is the number of the pixels.

SVMs was used as the classifier. The SVMs are state-of-the-art classifiers that have gained popularity within pattern recognition recently [23]. A brief review of the theory behind the SVMs algorithm is provided below.

Consider the problem of separating the set of training data {(x₁, y₁) , (x₂, y₂) , ⋯ , (x_m, y_m)} into two classes, where x_i ∈ R^N is a series of PSD value and y_i∈ { 0, 1 } is its class label: normal and stumbling gait. If we assume that the two classes can be separated by a hyperplane ω · x + b = 0 in some space and that we have no prior knowledge about the data distribution, the optimal hyperplane is the one that maximizes the margin. The optimal values for w and b can be found by solving a constrained minimization problem using Lagrange multipliers α_i (i = 1, ⋯ , m). $f (x) = sgn (\sum_{i = 1}^{m} α_{i} y_{i} K (x_{i}, x) + b)$ (2) where α_i and b are found by using a support vector machines learning algorithm. Those x_i with nonzero α_i are the “support vectors”. For K (x, y) = x · y, this corresponds to constructing an optimal separating hyperplane in the input space R^N.

The K-fold cross-validation (K-CV) method was employed to evaluate the performance of the recognition strategy [24]. In this approach, the feature sets were divided into 10 folds with equal size. Then, 9 folds of feature sets were used as training data to train the classifier, and the rest were used as testing data to test the classifier. To make each fold used once as testing data, we repeated this procedure 10 times.

Recognition accuracy (RA) was used to represent the quantification of the recognition performance. The RA was calculated by: $RA = \frac{N_{c}}{N_{t}} \times 100 %$ (3) where N_c means the number of correctly classified data and N_t means the total number of testing data. Finally, we calculated the mean value of the K RAs as the final recognition accuracy.

3 Results

Data analysis was performed using MATLAB®. Mean pressure data are displayed in Fig. 3(a). The 0% ∼30% gait cycle represents the swing phase. The mean pressure of normal gait showed similar pattern to that of stumbling gait, both in the shape of ‘M’. The maximum of the mean pressure occurred approximately 40% of the gait cycle and was found to be approximately 40 kPa in both situations. However, the second maximum mean pressure of normal gait was obviously greater than the latter, while the standard deviation was smaller.

Fig.3

Trajectory curves of different measures during gait cycle. Mean value (solid line) plus and minus standard deviation (dotted line) considering all trials from the repeated measure design (1 subject, 5 normal gait, 5 stumbling gait). Note: 0% ∼30% of the gait cycle represents swing phase.

Peak pressure data are displayed in Fig. 3(b). The peak pressure of normal gait and that of stumbling gait are different both in shape and in magnitude. The maximum of the peak pressure for the former was approximately 85% of the gait cycle and was approximately 1000 kPa, while it was approximately 40% of the gait cycle and was approximately 1000 kPa for the latter. Like the mean pressure, the standard deviation of the normal gait was obviously smaller than the latter. The PSD data were displayed in Fig. 3(c). The curve of the PSD of the normal gait was smoother. Compared to the normal gait, the standard deviation of the stumbling gait was much greater.

The SVM classifier was trained for each subject separately. The results of the stumble recognition are shown in Table 1. For four of them, the recognition accuracy reached 90%, and for other eight participants, it reached 100%. In addition, the mean value of recognition accuracy is 96.7%. The experiment results indicated that our approach can be used to recognize a stumble.

Table 1

Recognition results of the participants

Subject	RA (%)	Subject	RA (%)
S ₁	90	S ₇	100
S ₂	100	S ₈	100
S ₃	100	S ₉	90
S ₄	100	S ₁₀	100
S ₅	100	S ₁₁	90
S ₆	90	S ₁₂	100
Mean RA(%): 96.7

Note: RA means recognition accuracy.

4 Discussion

Most of the research focused on the recovery strategy, yet few studies have investigated the characteristics of plantar pressure during a stumble and the differences in the plantar pressure between stumble and normal gait. Plantar pressure is an essential characteristic during walking, either in stumbling or in a normal gait, and it has good potential to improve our understanding of the stumble before falling [25]. Plantar pressure has been included in studies of patients with diabetes [26] as a risk for falling [27, 28] and as risks for various lower limb musculoskeletal disorders [29, 30]. Traditionally, force platforms were used to assess the plantar pressure, but the measurements are generally confined to a particular location, often a treadmill in a laboratory [25]. Insole-based pressure sensors represent an interesting methodology that can be used to study the plantar pressure outdoors. In this study, we used the F-scan VersaTek system to collect the plantar pressure data, which can be used in many outdoor situations. This is more meaningful for occupational injury prevention in workshops or offices. In the future, wearable plantar pressure devices via wireless are expected, and it will be more convenient.

The plantar pressure parameters used most often are mean pressure, peak pressure and pressure-time integral [31 –34]. Keijsers et al. found that peak pressure, mean pressure and pressure-time integral were highly correlated [31]. However, in our study, when comparing normal gait and stumbling gait, mean pressure and peak pressure differed in pattern perhaps because only normal gait was considered in Keijser’s study. When encountering a stumble, the participants decelerated, prolonged the contact phase and increased their pressure during the off-contact phase to maintain balance. In addition, we proposed PSD, a novel feature. The PSD has a firm mathematical foundation, a wide application scope and good robustness. The experiment results showed that using PSD as the feature can recognize a stumble.

For the gait study, the analysis of several consecutive steps is generally not possible with motion analysis systems during walking [25], and most studies are based on an average of between 3 and 10 steps [35, 36]. Our study segmented the gait cycle according to the periodic variety of plantar pressure. We calculated the PSD of every frame and classified the data using the SVMs. In our study, the mean value of recognition accuracy was 96.7%. Therefore, even though the gait duration in this study was a little bit short, the approach we proposed can be used to distinguish the stumbling gait and normal gait and thus can be used to design anti-fall devices for the workers and staffs.

Although our proposed method is promising, it has its limitations. First, each trial lasts only 3 s, while a large number of data can fully reflect the differences in plantar pressure of the fall and normal gait. In addition, for some special groups, for example adolescents and the aging, one gait cycle may last more than 3 s. More data are needed to improve the recognition accuracy in the future. Second, when people walk on uneven terrains or places with small holes, their foot may not contact the ground completely. Therefore, useful information will be missed due to the absence of sensor feedbacks. In addition, the signals were transmitted to the base station via wire, which was inconvenient. In practice, the soles should transmit the signals via wireless. Past researches have mentioned that to obtain a more complete understanding of avoidance strategies employed by people, responses in environments with multiple obstacles of different types must be explored, for which they have set up three different types of obstacles in the experiment [37]. Weichun Hsu et al. studied participants walking and crossing obstacles of three different heights [38]. Because the obstacle height we set is too low, obstacle crossing can be relatively easy. More detailed experimental parameters, e.g., the number and type of obstacles, would be investigated in subsequent research. Furthermore, past researches have found that fall anxiety can regulate changes in gait patterns, thereby reducing the risk of falling due to a trip incident [39]. In our experiment, participants were aware of the presence and location of obstacles in advance. Future research would explore other conditions where an obstacle is not expected beforehand.

5 Conclusion

In this paper, we studied the differences in plantar pressure between the stumbling gait and the normal gait. We proposed a novel PSD feature and the SVMs data mining technique to differentiate a stumble. The experiment results showed that the stumbling groups had different plantar pressure patterns compared to the normal ones. The mean value of the recognition accuracy of our classification method was 96.7%, which is high enough for most industrial and office applications. It is promising to use the PSD of the plantar pressure images as the feature and use the SVMs as a classifier to detect stumbles. Given that the stumble is the precursor of a fall, this research would be meaningful in recognizing falls and issuing warnings and to help design anti-fall devices for potential victims.

Conflict of interest

None to report.

Footnotes

Acknowledgments

This research is supported by Special funds for the basic R&D undertakings by welfare research institutions (522016Y-4680), National Key R&D Program of China (2017) YFF0206602, General Administration of Quality Supervision, Inspection and Quarantine of the People’s Republic of China (201510042). The authors also appreciate the support from the State Scholarship Fund from China Scholarship Council (201208110144), the National Natural Science Foundation of China (51005016), and Fundamental Research Funds for the Central Universities, China (FRF-TP-14-026A2).

References

, Duan

, Wang

, et al. Analysis on data from Chinese National Injury Surveillance System, 2008-2013. Chin J Epidemiol. 2015;36(1):12–6.

Wild

, Nayak

, Isaacs

. How dangerous are falls in old people at home? Br Med J (Clin Res Ed). 1981;282(6260):266–8.

Reece

, Simpson

. Preparing older people to cope after a fall. Physiotherapy. 1996;82(4):227–35.

Gurley

, Lum

, Sande

, et al. Persons found in their homes helpless or dead. N

Engl J Med

. 1996;334(26):1710–6.

Silver

, Mokha

, Peacock

. Exploring fall training adaptations while walking. Work. 2016;54:699–707.

Stenhagen

, Nordell

, Elmstahl

. Falls in elderly people: Amultifactorial analysis of risk markers using data from the Swedish general population study ‘Good ageing in Skane’. Aging Clin Exp Res. 2013;25:59–67.

Patti

, Mammina

, Cataldo

, et al. Relationship between hypertension and accidental falls: The potential positive effects of physical exercise on blood pressure. Journal of Biological Research (Italy). 2015;88:113–4.

Patti

, Karsten

, Bianco

, et al. The effects of physical training without equipment on pain perception and balance in the elderly: A randomized controlled trial. Work. 2017;57:23–30.

Kunzler

, Rocha

, Santos

, et al. Should we consider steps with variable height for a safer stair negotiation in older adults? Work. 2018;59:15–21.

10.

Dietz

, Quintern

, Boos

, et al. Obstruction of the swing phase during gait: Phase-dependent bilateral leg muscle coordination. Brain Res. 1986;384(1):166–9.

11.

Dietz

, Quintern

, Sillem

. Stumbling reactions in man: Significance of proprioceptive and pre-programmed mechanisms. J Physiol. 1987;386:149–63.

12.

Eng

, Winter

, Patla

. Strategies for recovery from a trip in early and late swing during human walking. Exp

Brain Res

. 1994;102:339–49.

13.

Schillings

, Van Wezel

, Duysens

. Mechanically induced stumbling during human treadmill walking. J Neurosci Methods. 1996;67:11–7.

14.

Forner Cordero

, Koopman

. Multiple-step strategies to recover from stumbling perturbations. Gait & Posture. 2003;18:47–59.

15.

Cordero

, Koopman

HJFM

, Helm

FCTVD

. Energy analysis of human stumbling: The limitations of recovery. Gait & Posture. 2005;21(3):243–54.

16.

Foroughi

, Naseri

, Saberi

, et al. An eigenspace-based approach for human fall detection using integrated time motion image and neural network [A]. Proceedings of 2008 9th International Conference on Signal Processing (ICSP’2008) [C]. IEEE Beijing Section: IEEE BEIJING SECTION. 2008;1499-503.

17.

Lai

, Chang

, Chao

, et al. Detection of cognitive injured body region using multiple triaxial accelerometers for elderly falling. IEEE Sens J. 2011;11:763–70.

18.

Shi

, Zou

, Jin

, et al. PCA/ICA-based SVM for fall recognition using MEMS motion sensing data [P]. APCCAS 2008 [C]. Circuits and Systems. 2008:69-72.

19.

Tekscan Research Software User Manual. Version 5.24 Rev. D, Tekscan Inc, 2004.

20.

Maynard

, Brogmus

. Reducing slips, trips and falls in stairways. Occupational Hazards. 2007;81-4.

21.

Warren

, Maher

, Higbie

. Temporal patterns of plantar pressures and lower-leg muscle activity during walking: Effect of speed. Gait & Posture. 2004;19(1):91–100.

22.

Willems

, De Ridder

, Roosen

. The effect of a long-distance run on plantar pressure distribution during running. Gait & Posture. 2012;35:405–9.

23.

Zheng

, Huang

, Tan

, et al. A cascade fusion scheme for gait and cumulative foot pressure image recognition. Pattern Recognition. 2012;45(10):3603–10.

24.

Chen

, Wang

, Huang

, et al. A foot-wearable interface for locomotion mode recognition based on discrete contact force distribution. Mechatronics. 2015;32:12–21.

25.

Mann

, Malisoux

, Urhausen

, et al. Plantar pressure measurements and running-related injury: A systematic review of methods and possible associations. Gait & Posture. 2016;47:1.

26.

Allet

, Armand

, Golay

, et al. Gait characteristics of diabetic patients: A systematic review. Diabetes/Metabolism Research & Reviews. 2008;24(3):173–91.

27.

Moghadam

, Ashayeri

, Salavati

, et al. Reliability of center of pressure measures of postural stability in healthy older adults: Effects of postural task difficulty and cognitive load. Gait & Posture. 2011;33(4):651.

28.

Maki

, Mcilroy

. Control of rapid limb movements for balance recovery: Age-related changes and implications for fall prevention. Age & Ageing. 2006;35(Suppl 2):ii12.

29.

Nawata

, Nishihara

, Hayashi

, et al. Plantar pressure distribution during gait in athletes with functional instability of the ankle joint: Preliminary report. Journal of Orthopaedic Science. 2005;10(3):298–301.

30.

Morrison

. Plantar Pressure during Running in Subjects With Chronic Ankle Instability. Foot & Ankle International. 2010;31(11):994–1000.

31.

Keijsers

NLW

, Stolwijk

, Pataky

. Linear dependence of peak, mean, and pressure-time integral values in plantar pressure images. Gait & Posture. 2010;31(1):140.

32.

Tessutti

, Trombini-Souza

, Ribeiro

, et al. In-shoe plantar pressure distribution during running on natural grass and asphalt in recreational runners. Journal of Science and Medicine in Sport. 2010;13(1):151–5.

33.

Mickle

, Cliff

, Munro

, et al. Relationship between plantar pressures, physical activity and sedentariness among preschool children. Journal of Science and Medicine in Sport. 2011;14(1):36–41.

34.

Bergstra

, Kluitenberg

, Dekker

, et al. Running with a minimalist shoe increases plantar pressure in the forefoot region of healthy female runners. Journal of Science and Medicine in Sport. 2014;18(4):463–8.

35.

Dixon

. Use of pressure insoles to compare in-shoe loading for modern running shoes. Ergonomics. 2008;51(10):1503–14.

36.

Godi

, Turcato

, Schieppati

, Nardone A. Test-retest reliability of an insole plantar pressure system to assess gait along linear and curved trajectories. J Neuroeng Rehabil. 2014;11:95.

37.

Hak

, Hettinga

, et al. The concept of margins of stability can be used to better understand a change in obstacle crossing strategy with an increase in age. Journal of Biomechanics. 2019;84:147–52.

38.

Weichun

Hsu

, Mingwei

Liu

, et al. Biomechanical risk factors for tripping during obstacle—Crossing with the trailing limb in patients with type II diabetes mellitus. Gait & Posture. 2016;45:103–9.

39.

Brown

, Doan

, et al. Anxiety-mediated gait adaptations reduce errors of obstacle negotiation among younger and older adults: Implications for fall risk. Gait & Posture. 2006;24:418–23.