Premium
Automated prescreening of MCI through deep learning models based on wearable inertial sensors data
Author(s) -
Lee Hyeonil,
Shahzad Ahsan,
Kim Kiseon
Publication year - 2021
Publication title -
alzheimer's and dementia
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 6.713
H-Index - 118
eISSN - 1552-5279
pISSN - 1552-5260
DOI - 10.1002/alz.052744
Subject(s) - wearable computer , physical medicine and rehabilitation , gait , cognitive impairment , dementia , artificial intelligence , deep learning , gait analysis , computer science , medicine , cognition , psychology , neuroscience , disease , pathology , embedded system
Background During mild cognitive impairment (MCI), a precursor of dementia, often the mild symptoms go unnoticed and are assumed as normal aging signs by elderly people, resulting in late diagnosis and progression to dementia. Recently, several studies investigated the potential of behavioral analysis (ADL assessment, speech analysis etc.) based biomarkers for automated early screening of MCI in home settings. In this work, different deep learning models with walking data have been examined and a 2D‐CNN‐GRU model for automatic diagnosis of MCI based on wearable inertial sensors gait data is proposed. Method Sixty subjects (30 Cognitively Normal‐CN and 30 MCI) recruited from National Research Center of Dementia Database, Gwangju, South Korea, underwent comprehensive medical assessments by Chosun hospital doctors. The gait data of each participant was recorded using Shimmer3 inertial sensors under dual‐task settings. The subjects wore 7 shimmer3 sensors at five different body locations, i.e. waist, each thigh, each shank and each foot and performed 5 trials of dual‐task walking. For each experiment trials, the subject walked straight 10m and returned 10m while naming animals, food, fruits, colors and plants. Out of 10 meters, the central 6m stable walking signals were extracted and used for post processing. A total of eight features utilized for deep learning models’ training were, angular velocity along each axis of tri‐axial gyroscope sensor, 3‐axis acceleration data, and the signal vector magnitude of both sensors. In order to learn the gait pattern and time series characteristics for each step, the proposed model was built using six Convolutional Neural Networks‐(CNNs) as input for three Gated Recurrent Units‐(GRUs). Finally, the Leave One subject Out (LOO) Cross‐Validation (CV) was used to report the model's performance. Result As compared to individual representative deep learning models, the proposed combined model 2D‐CNN‐GRU proved to be more effective in MCI detection by learning discriminating walk signal characteristics and achieved Sensitivity 83.33%, Accuracy 73.33 % and Specificity 63.34% (Refer Table below). Conclusion This research work is the pioneer to investigate deep learning approaches for MCI diagnostic research using inertial sensor data and has the potential to facilitate early detection of MCI in home settings.