736 REAL-TIME MOBILITY DATA FROM WEARABLE DEVICES PREDICTING FALLS RISK IN OLDER ADULTS. (8th March 2022)

• Record Type: Journal Article
• Authors: Hewage, K; Lee, S; Fosker, S; Hardcastle, T; Ali, K
• Is Part Of: Age and ageing. Volume 51(2022)Supplement 1
• Publication Date: 2022-03-08
• Subjects: Aging -- Periodicals; Geriatrics -- Periodicals; 618.97
• DOI: 10.1093/ageing/afac035.736
• Languages: English
• ISSNs: 0002-0729

Abstract

Abstract: Introduction: Falls are a common cause of injuries in people over 65 years, usually resulting in life-threatening or life-altering sequalae for affected individuals. Falls risk assessment typically requires a face-to-face evaluation delivered by a trained professional in a clinical setting, thus limiting their application for many older people with falls. Therefore, there is a need for a remote monitoring system (RMS) that captures mobility data to assess comprehensively an older individual's true falls risk undertaking their everyday activities in their usual environment. Methods: A pilot study was conducted with 16 participants over 65 years classified as high (n = 9) or low falls risk (n = 7) based on their fall's history. A RMS was developed, which combines inertial measurement unit (IMU) data from a wearable device and online real-time Machine Learning (ML) to generate a two-stage fall risk assessment. In stage one, a ML activity motion classifier identified raw IMU data from participants as either, walking, resting or transitional movements (turning or sitting to standing). In stage two, previously classified motion segments were analysed by a ML fall-risk classifier, providing an objective risk assessment based on gait characteristics during everyday motion. Falls incidence was recorded prospectively at one year follow-up. Results: The motion classifier achieved an 94.8% accuracy in differentiating between resting, walking and transitional movements. The fall risk classifier initially trained with retrospective fall history data achieved an 87.5% accuracy with 88.9% sensitivity and specificity, which is comparable to other accurate classifiers in published literature. The fall risk classifier retrained with prospective one year follow-up data achieved a 73% accuracy which was superior to the 67% accuracy of the Timed Up and Go test for falls prediction. Conclusion: Our study shows the potential of a machine-learning remote monitoring system using information from mobility data to classify an older person's falls' risk.