Development of classification algorithms of human gait

Koh, Chee Hong (2022) Development of classification algorithms of human gait. Final Year Project, UTAR.

Preview

PDF Download (1622Kb) | Preview

Abstract

Gait analysis is essential for diagnosis, assessment, monitoring purpose, and prediction of gait disorder. However, the objective analysis method is less feasible in hospital environments for treatment purposes due to limited coverage of sources. Thus, this study aims to develop a classification algorithm that can effectively classify subjects with relatively simplified input data. This study employed several datasets acquired from PhysioNet containing subjects’ gait data of three classes. The training dataset contains at total of 48318 instances of three target classes (young healthy adults, old healthy adults, and Parkinson’s disease patients). Two classification algorithms were developed: Support Vector Machine (SVM) classification algorithm and Artifical Neural Network (ANN). Preprocessing was performed to the original dataset which includes data cleaning, data normalisation and new features generation. Next, fine-tuning on the manipulating hyperparameters was performed, and k-fold cross validation of k = 10 was used to obtain the average performance of the model. Results: The optimum confifuration of SVM model can generate an accuracy of 93.01% and F1 score of 92.58% with 43 minutes of computational time. On the contrary, the optimum configuration ANN classifier generates an accuracy of 90.56% and F1 score of 89.69% with 112 minutes computational time. In conclusion, comparing both of the proposed classification algorithms, the SVM classifier is more effectively than ANN classifier as overall for the gait dataset used in this study. In addtion, after compared with other state-of-the-arts of gait classification algorithms, our proposed classification algorithm produced comparable results with other state-of-arts using a smaller dataset with fewer training features.

Actions (login required)