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

Abstract

Background: Accelerometric analysis of gait abnormalities in golden retriever muscular dystrophy (GRMD) dogs is of limited sensitivity, and produces highly complex data. The use of discriminant analysis may enable simpler and more sensitive evaluation of treatment benefits in this important preclinical model.

Methods: Accelerometry was performed twice monthly between the ages of 2 and 12 months on 8 healthy and 20 GRMD dogs. Seven accelerometric parameters were analysed using linear discriminant analysis (LDA). Manipulation of the dependent and independent variables produced three distinct models. The ability of each model to detect gait alterations and their pattern change with age was tested using a leave-one-out cross-validation approach.

Results: Selecting genotype (healthy or GRMD) as the dependent variable resulted in a model (Model 1) allowing a good discrimination between the gait phenotype of GRMD and healthy dogs. However, this model was not sufficiently representative of the disease progression. In Model 2, age in months was added as a supplementary dependent variable (GRMD_2 to GRMD_12 and Healthy_2 to Healthy_9.5), resulting in a high overall misclassification rate (83.2%). To improve accuracy, a third model (Model 3) was created in which age was also included as an explanatory variable. This resulted in an overall misclassification rate lower than 12%. Model 3 was evaluated using blinded data pertaining to 81 healthy and GRMD dogs. In all but one case, the model correctly matched gait phenotype to the actual genotype. Finally, we used Model 3 to reanalyse data from a previous study regarding the effects of immunosuppressive treatments on muscular dystrophy in GRMD dogs. Our model identified significant effect of immunosuppressive treatments on gait quality, corroborating the original findings, with the added advantages of direct statistical analysis with greater sensitivity and more comprehensible data representation.

Conclusions: Gait analysis using LDA allows for improved analysis of accelerometry data by applying a decision-making analysis approach to the evaluation of preclinical treatment benefits in GRMD dogs.

Keywords: Accelerometry; Animal model; Discriminant analysis; GRMD; Gait assessment; Muscular dystrophy; Treatment evaluation.

Fig. 1

LDA Model 1: analysis of gait accelerometry parameters in healthy and GRMD dogs a Box and whisker diagrams of canonical variable F1 coordinates calculated by linear discriminant analysis of gait accelerometry parameters in healthy (green) and GRMD (red) dogs with genotype as the dependent variable. b Factor loading chart. SF, stride frequency; Reg, regularity; TP, total power; CCP/TP, cranio-caudal power normalized to TP; DVP/TP, dorso-ventral power normalized to TP; MLP/TP, medio-lateral power normalized to TP; SL/HW, stride length normalized to height at withers

Fig. 2

LDA Model 2: analysis of gait accelerometry parameters in healthy and GRMD dogs. a Linear discriminant analysis of gait accelerometry parameters for healthy and GRMD dogs with genotype and age (in months) as dependent variables. Individual measurements (dots) and groups (centroids and 95% confidence ellipses) are positioned on the plane using their values for the two first canonical variables, F1 and F2. Green and red colours correspond to healthy and GRMD dogs, respectively. For clarity, only the groups of younger and older animals were indicated for each genotypes. Arrows illustrate the evolution of class centroids according to age. The percentage variance explained by each canonical variable is indicated in parentheses. b Factor loading chart of F1 and F2 canonical variables. SF, stride frequency; Reg, regularity; TP, total power; CCP/TP, cranio-caudal power normalized to TP; DVP/TP, dorso-ventral power normalized to TP; MLP/TP, medio-lateral power normalized to TP; SL/HW, stride length normalized to height at withers

Fig. 3

LDA Model 3: analysis of gait accelerometry parameters in healthy and GRMD dogs. a Linear discriminant analysis plot of gait accelerometry parameters for healthy and GRMD dogs with genotypes and age in months as dependent variables, and age in days as an additional explanatory variable. Individual measurements (dots) and groups (centroids and 95% confidence ellipses) are positioned on the plane using their values for the two first canonical variables, F1 and F2. Green and red colours correspond to healthy and GRMD dogs, respectively. For clarity, only the groups of younger and older animals were indicated for each genotypes. Arrows illustrate the evolution of class centroids according to age. The percentage variance explained by each canonical variable is indicated in parentheses. b Factor loading chart of F1 and F2 canonical variables. SF, stride frequency; Reg, regularity; TP, total power; CCP/TP, cranio-caudal power normalized to TP; DVP/TP, dorso-ventral power normalized to TP; MLP/TP, medio-lateral power normalized to TP; SL/HW, stride length normalized to height at withers

Fig. 4

Gait phenotype separation: blind testing of LDA Model 3. To validate our proposed methodology, we used a blind dataset pertaining to 81 dogs and investigated the capability of Model 3 to predict genotype based on gait assessment. Each row corresponds to an individual dog and each column to a specific time point (age). Colours denote predicted phenotype based on the results of gait analysis. Green, red, and yellow correspond, respectively, to gait patterns that resemble that of healthy dogs (p > 95%), are strictly different to that of healthy dogs but similar to that of GRMD dogs (p > 95%), and are strictly different to the gait patterns of both healthy (p < 5%) and GRMD dogs (p < 5%). Grey colour indicates that measurements were not available. Cells of the penultimate column, entitled Predicted, are coloured to reflect the predominant phenotype (Healthy or GRMD) predicted by the model for each dog. Thus, green and red cells indicate that the gait of the corresponding dog resembles that of a healthy and a GRMD dog, respectively. n.c. denotes an inconclusive prediction result. Cells in the rightmost column are coloured according to the actual genotype of the dog

Fig. 5

LDA Model 3 as a tool for evaluating the effects of immunosuppressive treatment on gait in GRMD dogs. Immunosuppressive treatment has beneficial effects on gait in GRMD dogs. Using Model 3, curves were generated by plotting centroids, and the associated 95% confidence intervals, corresponding to healthy, untreated GRMD, and immunosuppressant-treated GRMD dogs on F1 and F2 axes. Centroids corresponding to GRMD-immunosuppressed dogs are colour-coded to reflect their comparison with the healthy group: green, similar to healthy gait (p > 0.95); yellow, intermediate gait (0.95 > p > 0.05), red, similar to GRMD gait (p < 0.05). For a more comprehensive representation, the projection of the age in days axe on the factorial plan was calculated and added as the upper axe (see text)