UTUG: An unsupervised Timed Up and Go test for Parkinson's disease. (March 2023)

• Record Type: Journal Article
• Authors: da Rosa Tavares, João Elison; Ullrich, Martin; Roth, Nils; Kluge, Felix; Eskofier, Bjoern M.; Gaßner, Heiko; Klucken, Jochen; Gladow, Till; Marxreiter, Franz; da Costa, Cristiano André; da Rosa Righi, Rodrigo; Victória Barbosa, Jorge Luis
• Is Part Of: Biomedical signal processing and control. Volume 81(2023)
• Publication Date: 2023-03
• Subjects: 00-01 -- 99-00; Gait analysis -- Gait test -- IMU -- Machine learning -- TUG -- Wearable sensors; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2022.104394
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Inertial measurement units (IMU) are used diagnostically in the movement analysis of Parkinson's disease (PD) patients, allowing an objective way to assess biomechanical motion and gait parameters. The Timed Up and Go (TUG) is a standardized clinical gait test widely used in the monitoring of patient fall risk and disease progression. Gait tests performed at home have been applied as part of movement monitoring protocols, enabling a link to clinical supervised reference assessments. However, unsupervised gait tests in a real-world data context present challenges, mainly regarding the interaction between participants and the recording system. Therefore, we developed and evaluated a novel algorithmic pipeline called unsupervised TUG (uTUG). Our contribution is the automatic detection and decomposition of TUG tests into their subphases, performed at home with no clinician supervision. In contrast to related studies, we used only foot-mounted IMU with no additional markers or manual annotations, allowing the detection of TUG test frames for subsequent classification by machine learning Support Vector Machine (SVM), Random Forest (RF) and Naïve Bayes Classifier (NBC) algorithms. The evaluation comprised 96 daily recordings of real-world gait data and 81 clinical visits accumulating 300 real TUG test samples processed from 32 PD patients. A prefiltering sensitivity of 98.6%, followed by the precision of 90.6%, recall of 88.5%, and Fl-score of 89.6% for TUG test detection were achieved using RF for the automatic classification in continuous real-world gait data. Thus, uTUG simplifies the test for patients and avoids manual annotations for clinicians, automatically detecting TUG tests. Highlights: Automatic detection of unsupervised Timed Up and Go tests in a real-world context. Automatic decomposition of unsupervised Timed Up and Go tests into their subphases. Inertial measurement units support the movement analysis of Parkinson's disease patients. Foot-mounted Inertial Measurement Units with no additional markers or annotations. Abstraction of software interaction, helping the patient focus on the test execution.