Machine Learning-based Gait Analysis for Recognition of Amnestic Mild Cognitive Impairment and Alzheimer's Disease

Abstract

Background The prevalence of age-related cognitive impairment, including dementia, has significantly increased with population aging. It has been shown that cognitive function is associated with gait status. Previously, researchers used statistical analysis methods instead of machine learning methods to study the gait of amnestic mild cognitive impairment (aMCI) and Alzheimer's disease (AD) . Objective To develop a model to identify aMCI and AD based on gait status using machine learning methods, explore gait markers differentiating between aMCI and AD, and to assess their possible values as aided tools in diagnosing aMCI and AD. Methods We recruited 102 cases from the Rehabilitation Hospital Affiliated to National Research Center for Rehabilitation Technical Aids, the First People's Hospital of Foshan, and Affiliated Zhongshan Hospital of Dalian University from December 2018 to December 2020, and included 98 of them according to the screening criteria, including 55 patients with aMCI, 10 patients with AD, and 33 healthy controls (HC) . The gait parameters of the participants were collected during performing single-task (free walking) , dual-task (counting backwards in sevens) and another dual-task (counting backwards from 100) using a wearable device. Random forest (RF) algorithm and gradient boosting decision tree (GBDT) algorithm were separately used to establish a model to compare the effect of two algorithms in recognizing three groups, with 10 gait parameters as predictive variables and the physical status (healthy, aMCI, AD) as response variables. Then important features were chosen using a machine learning algorithm combined with recursive feature elimination (RFE) . Results No statistically significant differences were found among the three groups in terms of sex ratio, average age, height, body weight or shoe size (P>0.05) , while the differences in terms of average MMSE score and MoCA score were statistically significant (P<0.05) . In the free walking test, aMCI group and AD group had shorter average stride length and smaller average heel-to-ground angle (HtA) than HC group (P<0.05) . AD group had slower average gait speed and smaller average toe-off angle (ToA) than both HC group and aMCI group (P<0.05) . In performing the dual-task of counting backwards in sevens, compared with HC group, aMCI group and AD group had slower average gait speed and smaller average ToA and HtA (P<0.05) . AD group had longer average stance phase than HC group (P<0.05) . AD group had average smaller ToA than aMCI group (P<0.05) .In performing the dual-task of counting backwards from 100, AD group had slower average gait speed and smaller average HtA and ToA than both HC group and aMCI group (P<0.05) . Moreover, AD group had shorter average stride length than HC group (P<0.05) . The average HtA in aMCI group was smaller than that in HC group (P<0.05) . Using the GBDT-RFE method, we found important gait features in distinguishing between aMCI and AD to be the stride length, ToA and HtA, and the model using the RF algorithm performed better in identifying aMCI and AD, with an accuracy as high as 87.69%. Conclusion Stride length, ToA and HtA are important gait markers to identify aMCI and AD. These findings could help clinicians diagnose aMCI and AD in the future.

Keywords

• cognitive dysfunction|amnestic mild cognitive impairment|alzheimer&apos;s disease|gait analysis|random forest|gradient boosting decision tree