Novel and Improved Stage Estimation in Parkinson's Disease using Clinical Scales and Machine Learning

TL;DR

This paper introduces a machine learning approach combining clinical scales to accurately stage Parkinson's disease, achieving high prediction accuracy and identifying key features relevant for diagnosis.

Contribution

It develops a novel PD staging method integrating MDS-UPDRS features and HY scale with machine learning models, improving accuracy over existing methods.

Findings

Support vector machine and AdaBoost models achieved up to 97.46% accuracy.

Key features for PD staging include bradykinesia, tremor, and facial expression.

Random forests identified the most important features for PD prediction.

Abstract

The stage and severity of Parkinson's disease (PD) is an important factor to consider for taking effective therapeutic decisions. Although the Movement Disorder Society-Unified Parkinson's Disease Rating Scale (MDS-UPDRS) provides an effective instrument evaluating the most pertinent features of PD, it does not allow PD staging. On the other hand, the Hoehn and Yahr (HY) scale which provides staging, does not evaluate many relevant features of PD. In this paper, we propose a novel and improved staging for PD using the MDS-UPDRS features and the HY scale, and developing prediction models to estimate the stage (normal, early or moderate) and severity of PD using machine learning techniques such as ordinal logistic regression (OLR), support vector machine (SVM), AdaBoost- and RUSBoost-based classifiers. Along with this, feature importance in PD is also estimated using Random forests. We observe that the predictive models of SVM, Adaboost-based ensemble, Random forests and probabilistic generative model performed well with the AdaBoost-based ensemble giving the highest accuracy of 97.46%. Body bradykinesia, tremor, facial expression (hypomimia), constancy of rest tremor and handwriting (micrographia) were observed to be the most important features in PD. It is inferred that MDS-UPDRS combined with classifiers can form effective tools to predict PD staging which can aid clinicians in the diagnostic process.