Machine learning can detect the presence of Mild cognitive impairment in patients affected by Parkinson’s Disease

Abstract

Parkinson Disease (PD) consists in a progressive, neurodegenerative disorder whose clinically characteristic is a combination of several motor and non-motor symptoms. Recently, the construct of Mild Cognitive Impairment (MCI), originally conceptualized to identify the pre-dementia state in Alzheimer Disease, has been employed in PD to describe a frame of cognitive decline without impaired functional activity. The aim of this study was to differentiate PD patients with and without MCI using quantitative gait variables through a machine learning approach. Thus, 45 PD patients underwent gait analysis and spatial-temporal parameters were acquired in three different conditions (normal gait, motor dual-task and cognitive dual-task). While the demographic and clinical features of PD patients with and without MCI were compared through a statistical analysis, the features of each gait condition were given as input to decision tree (DT), random forests (RF) and k nearest neighbour (KNN) to detect the presence of MCI. Then, some evaluation metrics were computed. DT achieved the highest accuracy (86.8%) using motor dual-task features, and the best sensitivity (88.2%), using gait task features as well as KNN (88.2% of sensitivity). KNN obtained the highest AUCROC (0.900) with the cognitive dual-task. DT with motor and cognitive dual-tasks and KNN with cognitive dual-task achieved the highest sensitivity (85.3%). Averaging the metrics, the cognitive dual-task showed the highest mean accuracy and specificity while the best mean sensitivity was obtained by the gait task. This paper proved that gait analysis and machine learning can be used to detect the presence in MCI in PD patients.