Recognition of Gait Disorders using Deep Learning Approaches
| Authors | |
|---|---|
| Year of publication | 2022 |
| Type | Conference abstract |
| MU Faculty or unit | |
| Citation | |
| Description | Introduction. We anticipate that the rapid development of deep-learning technologies will significantly influence the process of gait analysis in the future. In this work, we investigated the suitability of current deep-learning classifiers to the problem of recognition of gait disorders. Research Question. What is the most suitable approach to distinguish between two different musculoskeletal disorders (Legg Calvé Perthes Disease – LCPD and Cerebral Palsy – CP) based on various data-transformation techniques and deep-learning classifiers applied to kinematic/kinetic data from instrumented gait analysis? Methods. A total number of 1,870 gait cycles for 80 LCPD and 82 CP patients were used for analysis. Gait cycles were randomly split into halves in a balanced way to determine the training and test sets of patients. Both kinematics and kinetics were used – a single gait cycle was described by 21 parameters for each side of the lower body (x/y/z angles for pelvis, hip, knee, and ankle; and x/y/z moments for hip, knee, and ankle). Such gait-cycle data were further transformed into various representations by normalizing the temporal dimension or swapping the left and right legs. Three types of deep-learning classifiers (auto-encoder, convolutional network, and recurrent network in other three configurations with GRU, LSTM, and bi-LSTM cells) were trained on such transformed representations to recognize the movement disorder on the level of individual gait cycles. To recognize the disorder on the level of patients, the majority-voting principle was applied to a set of gait cycles that reflect a person’s movement pattern. Results. The best results were achieved by the combination of temporal normalization of gait cycles, data augmentation by mirroring the representations of the left and right leg, and the application of the LSTM-based neural network classifier. In particular, auto-encoder, convolutional network, and GRU, LSTM and bi-LSTM configurations of a recurrent network recognized the patient’s disease with the accuracy of ~81%, ~92%, ~88%, ~94%, and ~94%, respectively. The classification accuracy decreased by 5 percentage points if it was simply evaluated on the level of individual gait cycles, i.e., without the application of the majority-voting principle on the level of patients. The accuracy further decreased by 3 percentage points if the training set was not augmented using the leg-mirroring approach. Discussion. The main contribution to existing approaches was the combination of both kinematic and kinetic gait parameters and the application of deep-learning classifiers, in contrast to traditionally used SVM or PCA methods. |
| Related projects: |