A Validation Study of Freezing of Gait (FoG) Detection and Machine-Learning-Based FoG Prediction Using Estimated Gait Characteristics with a Wearable Accelerometer

• Published in: Sensors (Basel, Switzerland) Vol. 18; no. 10; p. 3287
• Main Authors: Aich, Satyabrata, Pradhan, Pyari Mohan, Park, Jinse, Sethi, Nitin, Vathsa, Vemula Sai Sri, Kim, Hee-Cheol
• Format: Journal Article
• Language: English
• Published: Switzerland MDPI 30.09.2018; MDPI AG
• Subjects: Accelerometry; Adult; Aged; Aged, 80 and over; feature extraction; freezing of gait; Gait - physiology; gait parameters; Humans; Machine Learning; mean error rate; Middle Aged; Parkinson Disease - diagnosis; Parkinson Disease - physiopathology; prediction; wearable accelerometer; Wearable Electronic Devices; feature extraction; wearable accelerometer; mean error rate; prediction; freezing of gait; machine learning; gait parameters
• ISSN: 1424-8220; 1424-8220
• DOI: 10.3390/s18103287

One of the most common symptoms observed among most of the Parkinson's disease patients that affects movement pattern and is also related to the risk of fall, is usually termed as "freezing of gait (FoG)". To allow systematic assessment of FoG, objective quantification of gait parameters and automatic detection of FoG are needed. This will help in personalizing the treatment. In this paper, the objectives of the study are (1) quantification of gait parameters in an objective manner by using the data collected from wearable accelerometers; (2) comparison of five estimated gait parameters from the proposed algorithm with their counterparts obtained from the 3D motion capture system in terms of mean error rate and Pearson's correlation coefficient (PCC); (3) automatic discrimination of FoG patients from no FoG patients using machine learning techniques. It was found that the five gait parameters have a high level of agreement with PCC ranging from 0.961 to 0.984. The mean error rate between the estimated gait parameters from accelerometer-based approach and 3D motion capture system was found to be less than 10%. The performances of the classifiers are compared on the basis of accuracy. The best result was accomplished with the SVM classifier with an accuracy of approximately 88%. The proposed approach shows enough evidence that makes it applicable in a real-life scenario where the wearable accelerometer-based system would be recommended to assess and monitor the FoG.