Elsevier

Human Movement Science

Volume 31, Issue 2, April 2012, Pages 284-294
Human Movement Science

Movement Deviation Profile: A measure of distance from normality using a self-organizing neural network

https://doi.org/10.1016/j.humov.2010.06.003Get rights and content

Abstract

We introduce the Movement Deviation Profile (MDP), which is a single curve showing the deviation of an individual’s movement from normality. Joint angles, recorded from typically developing children over one gait cycle, were used to train a self-organizing map (SOM) which then generated MDP curves for patients with gait problems. The mean MDP over the gait cycle showed a high correlation (r2 = .927) with the Gait Deviation Index (GDI), a statistically significant difference between groups of patients with a range of functional levels (Gillette Functional Assessment Questionnaire Walking Scale 7–10) and a trend of increasing values for patients with cerebral palsy through hemiplegia I–IV, diplegia, triplegia, and quadriplegia. The small difference between the MDP and GDI can be explained by the SOM’s method of operation comparing biomechanical patterns to the nearest abstract reference pattern, and its flexibility to compensate for temporal shifts in movement data. The MDP is an alternative method of processing complex biomechanical data, potentially supporting clinical interpretation. The electronic addendum accompanying this article is a standalone program, which can be used to calculate the MDP from gait data, and can also be used in other applications where the deviation of multi-channel temporal data from a reference is required.

Introduction

Single number indices like the body mass index (originally proposed by Adolphe Quetelet in 1832) can capture efficiently the interaction between single variables (mass and height). Once the number of biological variables increases, and multiple interactions become more complex, the challenge is to reduce the amount of data while retaining most of the useful information. The commonly adopted practical solution in clinical gait analysis has been to train experts who develop the skills and understanding required to make efficient use of the rich data. A gait analyst typically interprets joint angles, moments, and powers describing the dynamics of at least 7 segments of the human body recorded during walking in patients with gait problems (Chambers & Sutherland, 2002). Interpretation of the large quantities of data is compromised by the subjectivity of the analyst whose performance is affected by training, skills, and experience (Watts, 1994, Skaggs et al., 2000). More recently, several indices which encapsulate the content of complex gait data have been developed in attempt to provide simple measures of gait function.

A widely used method of quantifying how far a patient’s gait lies from normality is the Gillette Gait Index (GGI, previously the “Normalcy Index”) described by Schutte et al. (2000). The GGI has been extensively evaluated (Romei et al., 2004, Wren et al., 2007) and has been used in gait analysis research (e.g., Gorton et al., 2009). A recognized limitation of the GGI is that the choice of the 16 kinematic parameters used to calculate the index was based on the subjective judgement of clinicians and so may not be the best possible set of parameters universally. Barton, Lees, Lisboa, and Attfield (2003) described the use of self-organizing neural networks to quantify the quality of gait. A combined set of kinematic and kinetic variables taken from normal gait were used to train the network iteratively, resulting in the emergence of an operational definition of normality. The neural network’s performance metric was then used to describe how far a patient’s gait was from normality. An advantage of this technique is that deviations from normality can be calculated at all points during the gait cycle, and a sensitivity analysis can be used to highlight which joints contribute most to the deviations. Schwartz and Rozumalski (2008) addressed the limitations of the GGI by developing a new and comprehensive measure of gait pathology. Singular value decomposition of 9 joint angle curves over the gait cycle of a large number of patients and unimpaired controls was used to derive a set of 15 gait features, which could model the full spectrum of normal and pathological gait. The scaled and standardized Euclidean distance of a patient from the mean of controls in the 15 dimensional gait feature space was termed the Gait Deviation Index (GDI), which has been validated against the GGI and Functional Assessment Questionnaire (FAQ) scores in 3922 patients. In addition, the GDI of 3128 strides from cerebral palsy patients grouped into sub-types showed a clear trend of reduction in the GDI as severity of disability increased. Most recently, Baker et al. (2009), inspired by the GDI, developed the Gait Profile Score (GPS), which is the RMS difference between a patient’s data and the mean of controls, taken over several kinematic variables along the entire gait cycle. Through validation against the GGI, GDI, FAQ, and Gross Motor Function Measure in a large number of patients the GPS was found to be a closely related alternative to the GDI. The GPS algorithm applied to single kinematic variables (joint angles) gives the Gait Variable Score (GVS) to each of 9 joints, and the simultaneous presentation of the 9 GVSs was termed the Movement Analysis Profile (MAP).

As an alternative to conventional analysis of complex data, artificial neural networks, specifically the self-organizing map (SOM) described by Kohonen, 1988, Kohonen, 2001, have been used to detect deviation from normality in a wide range of applications. Markou and Singh (2003) reviewed the use of the SOM as a generic novelty detector in fields including signal processing applications in engine health monitoring and detection of cracks based on the acoustic emissions of machinery. Health related applications of the SOM include the work of Fernández, Willshaw, Perazzo, Presedo, and Barro (2001), who used the SOM to detect ischemic periods from 12 channel electrocardiograms with a recognition accuracy of 90.5%. The cases above are examples where the exact pattern of abnormal states is not known, so pattern recognition algorithms cannot be used. Even without any a priori examples of abnormal states, subsequent to learning the normal state, the SOM can quantify the deviation from normality by calculating the multi-dimensional distance between the abnormal state vector and the closest matching normal state vector.

The use of the SOM’s quantization error to quantify the deviation of gait from normality has been described by Barton, Lisboa, Lees, and Attfield (2007) and was illustrated by examples from patients with cerebral palsy. The aims of the present paper were to present a refined procedure, which generates the Movement Deviation Profile (MDP) from information provided by the SOM, to validate the MDP by comparing it to the GDI, FAQ scores, and clinical diagnoses of patients from a large dataset, and to make the method available for wider use by supplying a free computer program to calculate the MDP.

Section snippets

Data preparation

The MDP is not restricted to the analysis of gait data, any multi-channel time series variables can be used, e.g., accelerometry data recorded during activity monitoring or even physiological signals such as electroencephalograms. In order to compare the MDP to the GDI, a gait dataset similar to the one used by Schwartz and Rozumalski (2008) was processed for calculation of the MDP. One barefoot stride was selected from each side of subjects seen in the Gillette Children’s Specialty Healthcare

The Movement Deviation Profile

To illustrate how the MDP can aid interpretation of gait analysis results, Fig. 2 shows the conventional presentation of 9 joint angles plotted against the gait cycle of Patient 1 over the mean and standard deviation of TD controls. An additional chart (surrounded by dotted line) shows the MDP of the same patient and the MDP of TD controls (mean ± one standard deviation). The shape of the MDP clearly indicates that the greatest deviation from normality occurs during initial and mid swing (60–80%

Discussion

The movement deviation profile for a subject produces the deviation from the movement of controls in the form of a single curve, which reflects the distance from normality during the whole duration of a movement. The unique value of the MDP lies in the detail provided by a measure of deviation from normal gait at every point in the gait cycle. However, for the purposes of validating the MDP against the GDI this distributed information was ignored by collapsing the MDP curve into a single mean

Conclusions

The Movement Deviation Profile employs a self-organizing map (a type of artificial neural network) to calculate a curve which gives the deviation of movement from normality. When compared to the GDI in a large dataset covering 166 TD controls and 7294 strides of patients with various gait problems, the MDP matched the GDI well. Several advantages of the MDP were identified which stem from the neural network’s way of operation, making it a complementary tool alongside similar indices including

References (19)

There are more references available in the full text version of this article.

Cited by (41)

  • Three decades of gait index development: A comparative review of clinical and research gait indices

    2022, Clinical Biomechanics
    Citation Excerpt :

    More so, a directed change in the direction of normalcy does not necessarily indicate overall gait improvement, thereby preventing correlations between HFI score and patient gait functionality from being made (Cimolin and Galli, 2014). Movement Deviation Profile uses a self-organizing map (a type of artificial neural network (Kohonen, 1990)) with various input measurements (joint angles, moments and powers of the lower limbs and pelvis) to learn a representation of the normal gait (Barton et al., 2012). The methodology output is a curve representing the distance to normality throughout a gait cycle.

  • Women with patellofemoral pain show altered motor coordination during lateral step down

    2020, Journal of Biomechanics
    Citation Excerpt :

    We previously showed through the Movement Deviation Profile (MDP) analysis that lateral step down (LSD) is one of the tasks (out of seven) that most differentiate kinematics between PFP and asymptomatic controls (Lopes Ferreira et al., 2019). MDP combines kinematics data into one single curve and calculates the deviation of patient’s motion from normality during the whole movement cycle (Barton et al., 2012). Therefore, MDP provides a broad view of movement changes without specifying the motion(s) responsible for differentiating the groups.

  • Quantifying workers’ gait patterns to identify safety hazards in construction using a wearable insole pressure system

    2020, Safety Science
    Citation Excerpt :

    Several approaches have been studied to measure gait abnormality in clinical and rehabilitation settings. Examples include but not limited to the Gillette Gait Index (GGI), formerly called the Normalcy Index (Wren et al., 2007), the Gait Deviation Index (Barton et al., 2015) and Movement Deviation Profile (Barton et al., 2012). Ultimately, these approaches provide a single score to quantify the disruption of multiple gait features between healthy participants and patients with disorders such as Parkinson or Cerebral Palsy.

View all citing articles on Scopus
View full text