Hilbert-Schmidt Independence Criterion Lasso Feature Selection in Parkinson’s Disease Detection System

Abstract

Parkinson's disease is a neurological disorder which interferes human activities. Early detection is needed to facilitate treatment before the symptoms get worse. Earlier detection used vocal voice as a comparison with normal subject. However, detection using vocal voice still has weaknesses in detection system. Vocal voice contains a lot of information that isn't necessarily relevant for a detection system. Previous studies proposed a feature selection method on detection system. However, the proposed method can't handle variation in the amount of data. These variations include an imbalance sample to features and classes. In answering these problems, the Hilbert-Schmidt Independence Criterion Lasso (HSIC Lasso) feature selection method is used which has feature transformation capabilities that can produce more relevant features. In addition, detection system uses Synthetic Minority Oversampling Technique (SMOTE) method to balance data and several classification methods such as knearest neighbors, support vector machine, and multilayer perceptron to obtain best predictive model. HSIC Lasso produces 18 of 45 features with an accuracy of 88.34% on a small sample and 50 of 754 features with an accuracy of 96.16% on a large sample. From this result, when compared with previous studies, HSIC Lasso is more suitable on balanced data with more samples and features.