Smart Insole-based Classification of Alzheimer's Disease using Multi-scale Relation Network

Abstract

Alzheimer's Disease (AD) is a progressive mental deterioration disease which occurs at an older age. Gait has been adopted as a potential diagnostic tool for early detecting AD since gait is closely linked to cognitive function. Recently, wearable sensor devices have been used to extract useful clinical data for gait assessment. However, these wearable devices usually require high adherence from the user as binding on the thigh, waist, or arm, which may not be convenient for elderly patients. The smart insole provides an unobtrusive way to perform gait monitoring since it is extremely comfortable, thin, and lightweight to wear. It is equipped with a pressure sensor, an acceleration sensor, and a gyro sensor. However, as far as we know, there is no diagnostic tool for early detecting AD using a smart insole device. Hence, in this work, we proposed a smart insole-based gait analysis model to classify patients with AD and mild cognitive impairment (MCI), which is an early stage of AD, from healthy controls using various sensors in a smart insole. We proposed 4 gait experiment paradigms including normal walking and dual-task walking designed for observing the correlation between the pressure level to the brain and the gait patterns of healthy control and MCI patients. To obtain datasets for training our model with a smart insole, five subjects conducted a walking simulation five times per each classes. Through this, we obtained the gait features of a total of 75 samples (25 healthy controls, 25 MCI patients, and 25 AD patients) and 65 samples(42 healthy controls, 10 MCI patients, and 13 AD patients) were obtained from real elderly subjects.. We use metric-based few-shot learning for our classification model since the size of our dataset is very small. In our experiments, we demonstrate the performance of our proposed model under 4 gait experiment paradigms. The results show that our proposed classification model performs well especially for a complex gait task which highly affects the gait pattern of MCI patients. It is also verified using a dataset measuring the gait of real elderly subjects. These results show that our proposed method with gait features from smart insole can be used as an early diagnosis model to quickly initiate early intervention and prevention strategies that slow the progress of dementia.