Machine learning algorithms in spatiotemporal gait analysis can identify patients with Parkinson’s disease

Abstract

Changes to spatiotemporal gait metrics in gait-altering conditions are characteristic of the pathology. This data can be interpreted by machine learning (ML) models which have recently emerged as an adjunct to clinical medicine. However, the literature is undecided regarding its utility in diagnosing pathological gait and is heterogeneous in its approach to applying ML techniques. This study aims to address these gaps in knowledge. This was a prospective observational study involving 32 patients with Parkinson’s disease and 88 ‘normative’ subjects. Spatiotemporal gait metrics were gathered from all subjects using the MetaMotionC inertial measurement unit and data obtained were used to train and evaluate the performance of 10 machine learning models. Principal component analysis and Genetic Algorithm were amongst the feature selection techniques used. Classification models included Logistic Regression, Support Vector Machine, Naïve – Bayes, Random Forest, and Artificial Neural Networks. ML algorithms can accurately distinguish pathological gait in Parkinson’s disease from that of normative controls. Two models which used the Random Forest classifier with Principal Component analysis and Genetic Algorithm feature selection techniques separately, were 100% accurate in its predictions and had an F₁ score of 1. A third model using principal component analysis and Artificial neural networks was equally as successful (100% accuracy, F₁ = 1). We conclude that ML algorithms can accurately distinguish pathological gait from normative controls in Parkinson’s Disease. Random Forest classifiers, with Genetic Algorithm feature selection are the preferred ML techniques for this purpose as they produce the highest performing model.