Tremor-related feature engineering for machine learning based Parkinson’s disease diagnostics

Abstract

Growing research interest has arisen towards the possibility to automatically discriminate between the patients with neurodegenerative disease and healthy controls based on the information extracted from the digital drawing tests.In this paper, we propose novel higher-order derivative based, angular-type and integral-like features extracted from the Archimedean spiral drawing tests for machine learning based Parkinson’s disease diagnostics. The proposed features describe micro-changes in the handwriting trajectory, which are hard or impossible to detect with visual observation. However, they may hold valuable information in terms of tremor-like symptom analysis.Two datasets are considered in this study: DraWritePD (acquired by the authors) and PaHaW (well known from the literature). A filter (Fisher’s score) and wrapper (Recursive Feature Elimination) methods were used for feature selection. Six classifiers were trained and evaluated in a nested cross-validated loop to discriminate between healthy controls and Parkinson’s patients.A nested wrapper-type feature selection method combined with the ensemble classifiers predicted a disease with an accuracy of 84.33%, sensitivity of 70.00% and specificity of 93.20% (DraWritePD), and accuracy of 73.71%, sensitivity of 75.00% and specificity of 71.43% (PaHaW). The non-nested feature selection showed an over-optimistically high performance for both datasets: an accuracy of 92.16% (DraWritePD) and 84.86% (PaHaW).The proposed novel tremor-related features were among the best performing predictors in the case of both datasets. Furthermore, the results indicate that the nested feature selection procedure plays a significant part in the classification performance.