Design of 3D Convolution Neural Network-Based Diagnostic System for Mild Cognitive Impaired People Using Wearable Sensor Gait Signals

Abstract

As the life expectancy of humans increases, the elderly population is also increasing, increasing the prevalence of degenerative diseases, a social disease problem in the elderly population. Dementia is the most common degenerative disease in the elderly population and does not suddenly develop under normal cognitive function, but rather undergo a gradual decline in cognitive function. If symptoms worsen due to the decline of cognitive functions, including memory, daily life can be disrupted. Therefore, it is important to slow down the deterioration of the disease through early diagnosis and early treatment before dementia occurs, or to help improve the symptoms through continuous observation in real life. In particular, in the cognitive function extension linking normal to dementia, the medium-level mild cognitive impairment (MCI) is not up to the diagnosis criteria of dementia, but it is difficult to identify it early because it is a transitional stage in which cognitive function is degraded compared to normal. Recent deep learning methods are widely used to perform multidimensional data analysis, recognize images, and classify time series. In the field of diagnosing Alzheimer’s disease through walking analysis using wearable sensors, many studies are being conducted on classical machine learning methods, and diagnostic research methods using deep learning are lacking. Moreover, there is no research on classifying MCI patients, a precursor to Alzheimer’s disease, using walking data and deep learning measured by wearable sensors. In this paper, we try to solve the problem of diagnosing MCI patients and normal people through walking data extracted from acceleration/ gyroscope wearables sensor by applying the 3D-CNN model to reduce the opportunity cost for parameter extraction and to reflect time series characteristics. To further concentrate the fine differences between normal and patient, we also propose a diagnostic model with improved 3D-CBAM by extending the CBAM applied to the existing 2D-Attention to the time axis. It shows that simple walking without intervention of medical knowledge and no medical procedure makes it worthwhile to detect MCI patients early and facilitate monitoring, and to apply deep learning for the first time to MCI diagnostic research using wearable sensor data.