Automatic Gait Classification Patterns In Spastic Hemiplegia

Clinical gait analysis and the interpretation of related records are a powerful tool to aid clinicians in the diagnosis, treatment and prognosis of human gait disabilities. The aim of this study is to investigate kinematic, kinetic, and electromyographic (EMG) data from child patients with spastic hemiplegia (SH) in order to discover useful patterns in human gait. Data mining techniques and classification algorithms were used to explore data from 278 SH patients. We studied different techniques for selection of attributes in order to get the best classification scores. For kinematics data, the dimension of the initial attribute space was 1033, which was reduced to 78 using the Ranker and FilteredAttributeEval algorithms. For kinetics data, the best combination of attributes was determined by SubsetSizeForward Selection and CfsSubEval with a reduction of attribute space size from 931 to 25. Decision-tree based learning algorithms, in particular the logistic model tree based on logistic regression and J48, produced the best scores for correct SH gait classification (89.393% for kinetics, 89.394% for kinematics, and 97.183% for EMG). To evaluate the effectiveness of combined feature selection methods with the classifiers, quantitative measures of model quality were used (kappa statistic, measures of sensitivity and specificity, verisimilitude rates, and ROC curves). Comparison of these results to a qualitative assessment from physicians showed a success rate of 100% for results from kinematics and EMG data, while for kinetics data the success rate was 60%. The patterns resulting from automatic data analysis of gait records have been integrated into an end-user application in order to support medical decision-making.