An Rnn For Detecting Mild Cognitive Impairment Using Gait Data From A Single Waist-worn Imu

Mild cognitive impairment (MCI) can impact daily activities with a high risk of developing dementia. Gait analysis has been widely performed in the evaluation of neuro-degenerative conditions involving complex cognitive demands through the dual-task (DT) paradigm. Wearable inertial measurement units (IMU) are valuable for long-term monitoring and provide an effective method of quantifying movement mechanics. The recurrent neural network (RNN) method was proposed for IMU signals and has performed excellently in time series signal processing. Also, it has shown the potential to reduce the number of sensors (single waist-worn) in human motion analysis. The RNN-based model using gait biomarkers from inertial sensors, exhibits high accuracy and holds promise for MCI detection in older individuals. PURPOSE: The current study aimed to develop a new approach based on RNN techniques, namely Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models, for MCI early detection using a single waist-worn IMU data obtained during DT walking. METHODS: Thirty patients with MCI and 18 age-matched controls wore an IMU on the waist and walked on a 10-meter walkway while collecting 6-axis accelerometer and gyroscope data. Each patient performed normal and dual-task (serial sevens) walking tests. These IMU data were calculated through a low-pass filter and trained for proposed Uni-LSTM, Bi-LSTM, Uni-GRU, and Bi-GRU models. Data from 4538 gait cycles was used to train the models, and 10% of trials were used to test the model. Sensitivity and specificity between model-predicted data and experimental ground truth were used to quantify the prediction accuracy of each proposed model. RESULTS: The Bi-LSTM model performed high precision with a sensitivity of 80.67% and specificity of 90.33%, surpassing the Bi-GRU model (sensitivity: 77.23%, specificity: 88.24%), Uni-LSTM model (sensitivity: 48.56%, specificity: 88.34%), and Uni-GRU model (sensitivity: 27.32%, specificity: 99.58%). CONCLUSIONS: The proposed Bi-LSTM model employed for MCI pre-screening is based on inertial sensor-derived gait biomarkers with high prediction accuracy. This leads to excellent potential for real-life detection in older people with potential MCI.