Introduction: Classifying gait patterns into several categories for gait analysis of children with spastic cerebral palsy(CP) is helpful for clinical decision making; managing muscle spasticity, surgical lengthening of the shortened muscles, and applying orthosis to improve gait patterns. However, the gait pattern classification is still conducted with subjective judgement and the personal experience of healthcare providers.
The aim of this study was to propose a deep learning model for classifying gait patterns in children with CP based on 3D motion capture data and to compare the performance with machine learning algorithms and clinical classification.
Method: We collected the data from 267 children with bilateral spastic CP to train the machine learning and deep learning models and tested those models in 91 children. The four gait patterns (True equinus, Jump knee, Apparent equinus, and Crouch) were classified according to the widely used definition by Rodda et al.
From 3D motion capture data of Helen Hayes marker set, the custom datasets were reconstructed with joint kinematics (hip, knee and ankle) on sagittal plane and marker coordinates data (sacrum, anterior superior iliac spine, thigh, tibia, ankle, and toe) for deep learning analysis. We also extracted the seven features from joint kinematic data (maximum hip flexion/extension, maximum knee flexion/extension, maximum knee flexion at stance phase, maximum ankle dorsiflexion/plantar-flexion) for machine learning analysis.
The deep learning model was constructed to train time-series data using Conv1D, LeakyReLU, Cross entropy loss, Adam, StepLR to form DCNN. Then, the seven machine learning algorithms including Multi Layer Perceptron(MLP), Discriminant Analysis(DA), Decision Tree(DT), Random Forests(RF), K-Nearest Neighbors(KNN), Support Vector Machine(SVM), and Naive Bayes(NB) were applied to compare the performance with deep learning approach. The clinical classification was performed by one experienced physician and one trainee physician.
Result: We obtained the accuracy with DCNN(71.6%) and MLP(61.1%), DA(54.4%), DT(52.2%), RF(62.2%), KNN(62.2%), SVM(56.7%), NB(62.2%). The clinical classification by observing gait pattern showed 75.8% accuracy in experienced physician and 56.0% accuracy in trainee physician for same test data. Especially in Apparent equinus(76.7%) and Crouch(75.0%), the DCNN showed higher accuracy even compared to the classification by an experienced physician: Apparent equinus(61.8%) and Crouch(63.8%).
Conclusion: Our deep learning model showed the higher classification accuracy compared to machine learning algorithms and clinical classification by a trainee physician. However, the machine learning algorithms did not show the satisfactory accuracy reported in a previous study.
The future studies are needed to optimize the deep learning model and make additional train data for better performance. |
