Parkinson's Disease Motor Symptoms in Machine Learning: A Review

Abstract

This paper reviews related work and state-of-the-art publications for recognizing motor symptoms of Parkinson's Disease (PD). It presents research efforts that were undertaken to inform on how well traditional machine learning algorithms can handle this task. In particular, four PD related motor symptoms are highlighted (i.e. tremor, bradykinesia, freezing of gait and dyskinesia) and their details summarized. Thus the primary objective of this research is to provide a literary foundation for development and improvement of algorithms for detecting PD related motor symptoms.

Classification and characterisation of movement patterns during levodopa therapy for parkinson's disease

“…This is most pronounced for dyskinesia, but is also true for tremor. This is an interesting result because a lot of existing classification work on both these movement types has focused on the frequency domain [1,9,14]. For tremor, a frequency analysis seems natural, since it is defined as a regular rhythmic oscillation of a body part, and the frequency ranges for di↵erent kinds of tremor are relatively well characterised.…”

“…Previous work on applying machine learning and data mining techniques to movement data collected from PD patients is reviewed in [1]. Particularly relevant are previous studies that have looked at predicting specific motor symptoms.…”

Wearable sensors for gesture analysis in smart healthcare applications

“…Examination of literature reveals a large number of techniques for the automatic detection of PwPD motor symptoms. However, generally machine learning algorithms were used to detect a single motor symptom as described in [11]. Typically, patients are likely to experience multiple symptoms, thus increasing the chance of false positive and false negative.…”

“…These results not only endorse the ability of sensHandV1 to assess the biomechanical parameters accurately from the upper limb exercises, but also suggest that these biomechanical parameters have the potential to discriminate the PwPD on a clinical scale. In the previous studies [11], the lack of representative measures was an important limitation in the objective measuring system, as most of the multi-dimensional motions are presented in the form of separate physical parameters. At the same time, none of the individual parameters would be representative to evaluate the motor performance of the upper and lower limbs.…”

Metric learning for Parkinsonian identification from IMU gait measurements

“…Machine learning has been successfully used for the diagnosis of individual forms of dementia related Parkinson's [10], but also early Alzheimer's [11]. ML disease progression approaches have also been explored to rate the severity [12] in PD (based on the UPDRS scale), for example via postural sway analysis employing support vector machine (SVM) classification [13] or via longitudinal measurements combined with random forest [3] regression.…”