The objective of this study is to investigate the effects of feature selection methods on the performance of machine learning methods for quantifying motor symptoms of Parkinson’s disease (PD) patients. Different feature selection methods including step-wise regression, Lasso regression and Principal Component Analysis (PCA) were applied on 88 spatiotemporal features that were extracted from motion sensors during hand rotation tests. The selected features were then used in support vector machines (SVM), decision trees (DT), linear regression, and random forests models to calculate a so-called treatment-response index (TRIS). The validity, testretest reliability and sensitivity to treatment were assessed for each combination (feature selection method plus machine learning method). There were improvements in correlation coefficients and root mean squared error (RMSE) for all the machine learning methods, except DTs, when using the selected features from step-wise regression inputs. Using step-wise regression and SVM was found to have better sensitivity to treatment and higher correlation to clinical ratings on the Unified PD Rating Scale as compared to the combination of PCA and SVM. When assessing the ability of the machine learning methods to discriminate between tests performed by PD patients and healthy controls the results were mixed. These results suggest that the choice of feature selection methods is crucial when working with data-driven modelling. Based on our findings the step-wise regression can be considered as the method with the best performance.