Falls Prediction in Care Homes Using Digitally Recorded Residents’ Daily Activities

Conference: The European Conference on Aging & Gerontology (EGen2022)
Title: Falls Prediction in Care Homes Using Digitally Recorded Residents’ Daily Activities
Stream: Frailty
Presentation Type: Oral Presentation
Authors:
Robert Moskovitch, Ben Gurion University, Israel
Ofir Dvir, Ben Gurion University, Israel
Paul Wolfson, University College London, United Kingdom
Laurence Lovat, University College London, United Kingdom

Abstract:

A leading cause of unintentional injury-related death in older adults are falls. While falls among elderly are a well-documented phenomenon, it has been mainly investigated in clinical settings, due to lack of documentation. This study consists on data from over 1,769 care homes, including 149,250 residents across the United Kingdom. Carers continuously documented the residents’ daily activities using the Mobile Care Monitoring mobile app along three years. We introduce FallPry, a first fall prediction framework for elderly residents who live in care homes, which continuously predicts ahead, based on recent weeks. Due to the heterogeneity and temporal nature of the data, Temporal Abstraction and Time Intervals Related Patterns (TIRPs) discovery are employed. These patterns are used as predictors for the fall’s prediction. To identify the most predictive variables we introduce a novel Irregularly Measured Temporal Variable Selection method (IMTVS), that ranks the variables according to the frequency of each value in the various classes. In this talk we will describe our investigation on a large and unique database of care homes daily activities, and how they can predict falls in high accuracy. We will demonstrate the method, and our insights about the most predictive variables, and the most predictive temporal patterns (TIRPs), and a comparison with state of the art deep learning methods. The use of the top 40 variables selected by IMTVS performed better than selecting them randomly, and even better than using all the 88 variables, achieving 87% AUC.

---

Conference Comments & Feedback

Place a comment using your LinkedIn profile


Share this Presentation