On the classification of tremor signals into dyskinesia, Parkinsonian tremor, and Essential tremor by using machine learning techniques. (March 2022)

• Record Type: Journal Article
• Authors: Ferreira, Gabriel A.S.; Teixeira, João Lucas S.; Rosso, Ana Lucia Z.; de Sá, Antonio Mauricio F.L. Miranda
• Is Part Of: Biomedical signal processing and control. Volume 73(2022)
• Publication Date: 2022-03
• Subjects: Dyskinesia -- Essential Tremor -- Parkinsonian tremor -- Classification -- Parkinson's disease; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103430
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Parkinson's disease (PD) is a neurological disorder that leads to motor symptoms. Some of them are the Parkinsonian tremor (PT) and dyskinesia (DS). Essential Tremor (ET) is a monosymptomatic disorder, usually confused with PT. Five machine learning algorithms were used for distinguishing the three disorders. Decision Tree and Random Forest performed best in the data set used. Abstract: Parkinson's disease is a neurological disorder that leads to motor symptoms, some of them being the typical Parkinsonian tremor (PT) and dyskinesia, which is characterized by random movements of the limbs that appear as medication side effect. Moreover, the Essential Tremor (ET) is a monosymptomatic disorder, usually confused with PT, due to frequency content overlap of both signals. Unified Parkinson's Disease Rating Scale is used for the clinical assessment of PD. It considers the forms filled out with patients' symptoms information and their physical evaluation. This scale is subjective due to the inter-rater variability, justifying the aim of the present work of developing a methodology to distinguish these three disorders. Thus, five machine learning algorithms - K-Nearest Neighbours (K-NN), Decision Trees (DT), Random Forest (RF), Naïve Bayes (NB) and Support Vector Machines (SVM) - were chosen for classification purposes. First, variables were extracted from the collected signals (mean, standard deviation, and amplitude peak in time; dominant and second dominant frequencies; first and second spectral peaks; correlation peak and instant time of it). Next, a Principal Component Analysis was carried out for reducing the data set to three components that explained 95% of the data variance, which were then used as inputs for the classification models. DT and RF showed highest accuracy (=1), followed by SVM (=0.9394) using the Gaussian kernel function, whereas K-NN and NB showed the lowest one (=0.8788). Considering also precision, recall, and F1-score, DT and RF were found to be the most appropriate models for this problem.