Identification of activities of daily living in tremorous patients using inertial sensors

• Published in: Expert systems with applications 2017-10, Vol.83, p.40-48
• Main Authors: Serrano, J. Ignacio, Lambrecht, Stefan, del Castillo, M. Dolores, Romero, Juan P., Benito-León, Julián, Rocon, Eduardo
• Format: Article
• Language: English
• Subjects: Artificial intelligence; Biosensors; Daily activity identification; Essential tremor; Inertial sensing devices; Inertial sensors; Machine learning; Methodology; Movement; Parkinson; Parkinson's disease; Patients; Pattern recognition; Remote monitoring; Remote sensors; Time series; Time series classification
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2017.04.032

•Tremor and voluntary movement of the upper limb are separated from IMUs signals.•Seven daily living tasks are classified with an accuracy of 86%.•Grain (precise or gross) of daily living tasks is identified with 79% accuracy.•Direction (distal/ proximal) of daily living tasks is identified with 89% accuracy.•Tremor kinematics can be coupled with the kind of task where it appears. Much attention has been given to the use of inertial sensors for remote monitoring of individuals suffering from neurological pathologies. However, the focus has been mostly on the detection of symptoms like tremor or dyskinesia, not on identifying specific activities carried out by subjects. The objective of this study was to develop an automated segmentation and recognition methodology, from inertial sensor data, to identify tasks and motor patterns during activities of daily living. This will enable clinicians to contextualize the symptoms of these diseases and improve their treatment. We designed and tested a methodology to automatically label continuous upper-limb activity from IMUs (Inertial Measurement Units). Three classification problems are considered: the task itself, the precision level required by the task with respect to movement (fine or gross) and the trajectory of the task (distal or proximal). These problems were identified by tremor experts as clinically important aspects to be monitored in tremor patients while performing daily activities. The proposed methodology reveals the relation between the functional context or activity and the patient's on-going tremor to clinicians. Overall task identification rate was 86%. Task precision and task trajectory are classified with 79% and 89% accuracy, respectively. Aligning the semantic nature of the activity with the tremor location and intensity can also provide novel and relevant information for clinical monitoring of tremor and help clinicians and researchers as a useful tool to develop new therapies or strategies and novel anti-tremor medications to combat context dependent tremor. The present study describes the development of a comprehensive methodology based on machine learning techniques to segment and detect activities of daily living in people with tremor using inertial sensors, which aims at facilitating detailed interpretation of tremor movements by neurologists.