Detection of Freezing of Gait and Feature Extraction of Human Gait with Linear Discriminant Analysis

Sammanfattning: Linear discriminant analysis (LDA) is performed in order to detect freezing of gait (FOG) in gait acceleration data. The data is collected from 10 healthy test persons, imitating Parkinson's disease (PD) gait, with the cellphone application Medoclinic. Feature spaces are constructed and reduced with principal component analysis (PCA), on which LDA is performed. The best performing model constructed with the LDA technique has a precision and recall of 0.6071 and 0.6800, respectively. This model projects its feature space on the five first principal components. Extended versions of the algorithm written by Moore-Bächlin and later developed by Capecci et al., are constructed in order to detect the FOG phenomenon in gait acceleration. The best performing model is a simplified version of the algorithm written by Moore et al., with both a precision and recall of 0.6154. More characteristics of the gait data can be captured with LDA than what the algorithms of Moore-Bächlin and Capecci allow.