Long short term memory based patient-dependent model for FOG detection in Parkinson's disease

• Published in: Pattern recognition letters 2020-03, Vol.131, p.23-29
• Main Authors: Ashour, Amira S., El-Attar, Amira, Dey, Nilanjan, El-Kader, Hatem Abd, Abd El-Naby, Mostafa M.
• Format: Article
• Language: English
• Subjects: Accelerometer sensor; Accuracy; Ankle; Classification; Deep learning; Freezing of gait; Gait; Learning algorithms; Long short term memory deep learning model; Long short-term memory; Machine learning; Movement disorders; Neurodegenerative diseases; Parkinson's disease; Sensors; Short term; Signs and symptoms; Support vector machine; Support vector machines; Wearable sensors
• ISSN: 0167-8655; 1872-7344
• DOI: 10.1016/j.patrec.2019.11.036

•Monitoring 'Parkinson's disease (PD) using a recorded signals’ dataset from on-body wearable sensors is a vital requirement.•Freezing of gait (FOG) in PD patients can be detected from the acceleration signals.•A deep learning using LSTM network-based patient-dependent model was adopted for FOG detection.•A comparison between the proposed model and support vector machine with linear kernel was reported.•The LSTM achieved superior performance of 83.38% accuracy, while the SVM achieved 79.48% accuracy. Deep learning has a great impact on healthcare for discovering hidden patterns in the clinical data to detect or predict the different diseases. This work proposed a monitoring procedure for Parkinson's disease (PD) using a recorded signals’ dataset from multiple wearable on body sensors placed at different positions on the leg, namely on knee, hip and ankle. Different symptoms of PD patients can be detected from the acceleration signals, where the Freezing of Gait (FOG) is considered the main sign. Typically, FOG is patient-dependent that varies in severity and incidence from patient to another. In this work, a deep learning model, namely the Long Short Term Memory (LSTM) network-based patient-dependent model was adopted for FOG detection. A comparison between the proposed model and the traditional machine learning methods, including the linear support vector machine (SVM) was conducted using the signals of the three sensors. The results established the superiority of the LSTM model, which achieved 83.38% in terms of the average accuracy in comparison with the SVM which achieved 79.48%. For example, in patient 2, the maximum accuracy achieved using the LSTM is 98.89%, while the corresponding maximum accuracy is 80% using the linear SVM. [Display omitted]