Gait characterization in golden retriever muscular dystrophy dogs using linear discriminant analysis

Duchenne muscular dystrophy (DMD) is an X-linked disorder caused by various mutations in the gene encoding for dystrophin, resulting in the absence of the functional protein in muscle fibres [1]. DMD patients display progressive muscle weakness leading to the loss of independent mobility in young adolescents and respiratory and heart failure in young adults. While gene, cell, and pharmacological therapies have all been investigated [2, 3], there is currently no curative therapy for DMD. However, several issues such as efficacy of studied drugs, locoregional or systemic medication pathways or dosage, remain to be further explored. The preclinical DMD model of choice is the dystrophin-deficient golden retriever muscular dystrophy (GRMD) dog, which closely mimics many aspects of the human disease [4, 5]. Scale is a key factor that influences the translation of data from animal models to humans. Thus, when studying mechanical impacts, molecular diffusion and/or cell migration data acquired in GRMD dogs is much more relevant to human DMD than that obtained in smaller animal models such as the mdx mutant mouse [6]. The downside of the GRMD model is that it is more expensive to purchase and house, and the use of "man's best friend" for research purposes entails additional political and ethical considerations. Accordingly, limited numbers of GRMD dogs are generally used in studies. At first glance, this appears to be a major impediment to the preclinical evaluation of therapies, particularly given that GRMD dogs exhibit considerable inter-individual phenotypic variability [6‐8]. However, there are numerous similarities between canine and human diseases, and considerable inter-individual variability is also observed among human patients [9]. Moreover, given the rarity of muscular dystrophies, and for obvious ethical reasons, clinical studies in man generally involve limited numbers of patients and pose similar challenges to those performed in GRMD dogs. Tools that allow the establishment of better readouts of disease progression and treatment response are essential to overcome limitations imposed by small sample sizes and wide inter-individual variability in studies using the GRMD model, and to better predict the pathogenesis of muscular dystrophy and treatment efficacy in humans.

Studies by several groups have conducted gait analysis in GRMD dogs [10‐14]. Using accelerometry analysis in these animals we have demonstrated less regular and less powerful acceleration, decreased stride length and frequency, and a redistribution of power from the cranio-caudal to the medio-lateral axis [11]. Using the main gait variables, we developed a global gait index that consistently detected early changes in gait patterns in GRMD, as well as the progressive deterioration of gait quality. This index was based on the use of principal component analysis (PCA) and the computation of Euclidean distances at multiple time points with respect to an age-matched control group [10]. Given the inherent complexity of the methodology, we believe that this approach is not best suited to the problem at hand, and in fact may hinder the evaluation of therapies in preclinical studies in GRMD dogs. To increase sensitivity and aid interpretation of the outcomes, we sought to design a simpler and more appropriate analytical method using linear discriminant analysis (LDA). Like PCA, LDA is an orthogonal transformation and data reduction technique, but unlike PCA, LDA seeks to minimize intra-group variance and maximize inter-group variance. Moreover, LDA yields a predictive model based on control group data. This is achieved by computation of group membership based on experimental data and assignment rules, which allow the prediction of group membership for future observations. Here, we describe a new method of 3D accelerometric gait analysis using LDA. We discuss the choice of dependent and independent variables and describe how to represent the results in a manner that can be understood by a broad range of users, including those unfamiliar with LDA. Finally, we demonstrate the validity of this method by reanalysing data from a previous study regarding immunosuppressive treatment in GRMD dogs.

We previously demonstrated that PCA, although a non-supervised analysis strategy, can distinguish the gait phenotype of GRMD dogs from that of healthy controls with reasonable accuracy, and does so independently of age [11]. We show in the present study that similar results could be obtained using LDA. However, when the age of the dogs were not take into account, the model built using LDA (Model 1) was found not to adequately represent the progression of the disease, which evolves over the course of the postnatal growth phase. On the other hand, addition of age, in months, as a supplementary dependent variable in LDA, the model obtained (Model 2) exhibited a very high rate of misclassification. This likely reflected the slow rate of disease progression over the short age intervals analysed.

Nonetheless, in Model2, the discriminant axis mainly associated with age and disease progression, F2, presented a low canonical correlation, indicating that the model was not sufficiently sensitive to accurately evaluate the impact of progressive dystrophy on gait in growing GRMD dogs. One means of improving the model was to increase the age interval (e.g. from 0.5 to 1 month), which would certainly reduce the misclassification error, but would also decrease the sensitivity of the model. The aim of this study was not to demonstrate altered gait in GRMD versus healthy dogs, but rather to develop a method to better identify gait alterations and their progression with age in GRMD dogs. To improve the descriptive properties of the model while keeping its predictive capabilities, we have artificially stretched the model (Model 3) in the direction of the progression of age, by introducing age (in days) as an explanatory (or dependent) variable in the LDA analysis, in addition to including age (in months) as a dependent variable. The dogs included in the study were not born on the same day. Additionally, sometimes some dogs were not able to walk, thus their acquisition session was delayed. By contrast, acquisition sessions were always done on the same day of the week, every 15 days. Thus, the age of the animals in days in the same age group in months could vary from about 7 days. The intrinsic variability of the age in days spread the model in the factorial plan helping to discriminate the different groups of age, although it could be assumed that this variability does not significantly impact the measurements. In Model 3, the two first canonical discriminant factors, F1 and F2, accounted for 99.7% of variance. F1 was predominantly correlated with age whereas F2 was discriminant for gait phenotype. Model 3 presented a low overall misclassification error and high canonical correlations of 0.99 and of 0.95 were calculated for F1 and F2, respectively.

In order to investigate further the suitability of Model 3 as a tool for gait analysis in preclinical studies with GRMD dogs, we used it to predict the genotypes of a cohort of 81 healthy and GRMD dogs starting from their gait accelerometry characteristics. For each dog of the cohort, a number was randomly attributed by an external scientist blinded to the data. Using this approach, when “the data were unblinded”, the predicted gait phenotype matched the actual phenotype and genotype in all cases but one. These findings strongly support the robustness and accuracy of the method.

Some members of our group tested, in a previous study, the effects of immunosuppressive treatments (oral administration of cyclosporine A and corticosteroids) on muscular dystrophy and overall health in GRMD dogs [12]. Indeed, immunomodulatory treatments have been employed in several studies assessing the effectiveness of gene, cell, or pharmacological therapies in dog models of DMD to supress the immune response to the viral vector, donor cells, and/or the transgene product [18‐20]. The obtained results vary considerably depending on the variables measured. Although Barthélémy and colleagues [12] reported a more severe disease progression in terms of isometric force and histology, they also found an improvement in gait classification using a PCA-based gait index calculated from the seven accelerometric variables used in the present study. This longitudinal analysis was complex, as PCA, unlike DA, is an unsupervised approach. The authors performed PCA for each age category and calculated the corresponding Euclidean distance for each GRMD dog to the centroid of age-matched healthy dogs, and, thereafter, plotted the evolution of this distance with age. Using Model 3, we found that immunosuppression in GRMD dogs had a significant beneficial effect on gait, although this was both partial and temporary. This finding was in line with those of Barthélémy and coworkers [12]. Furthermore, LDA revealed that significant effects of treatment on gait observed at 4 months of age were preceded by signs of improvement that were evident as early as 3 months, supporting the sensitivity of our model.

Using LDA we have generated a highly sensitive model of gait alterations due to muscular dystrophy in the GRMD dog. Our model shows a high degree of discriminatory accuracy, distinguishing the gait phenotype of GRMD dogs from that of healthy dogs as early as 2.5 months of age, and thus overcoming some of the difficulties in analysing a progressive disease that occurs during the growth phase of postnatal development. Moreover, we designed a new means of representing the outcomes of our analysis that allows for easier interpretation of the results. This is a key strength of our study: because preclinical results are the base upon which phase 1/2 clinical trials in human patients are prepared and designed, preclinical outcomes should be completely understandable to all those involved in the design and the testing of potential treatments from preclinical through to clinical phases. In our experience, graphical displays depicting only the two variables (F1 and F2) are very likely to be unclear, and sometimes misleading, to practitioners unfamiliar with DA or PCA.

We previously demonstrated that accelerometry combined with PCA constitutes a reliable follow-up tool for gait analysis in preclinical therapeutic trials using GRMD dogs. However, this approach has some limitations, including limited sensitivity, and high complexity of the data generated. The model presented here greatly improves upon this method by employing LDA.