Unobtrusive detection of frailty in older adults

Abstract

Sensor technologies have gained attention as an effective means to monitor physical and mental wellbeing of elderly. In this study, we examined the possibility of using passive in-home sensors to detect frailty in older adults based on their day-to-day in-home living pattern. The sensor-based elderly monitoring system consists of PIR motion sensors and a door contact sensor attached to the main door. A set of pre-defined features associated with elderly’s day-to-day living patterns were derived based on sensor data of 46 elderly gathered over two different time periods. A series of feature vectors depicting different behavioral aspects were derived to train and test three machine learning algorithms; Logistic Regression, Linear Discriminant Analysis and Naïve Bayes. The best prediction scores yielded by seven features, namely, daytime napping, time in the bedroom, night-time sleep, kitchen activity level, kitchen use duration, in-home transitions and away duration. These features produced an area under the ROC curve of 98%, 79% and 93%, for Logistic Regression, Linear Discriminant Analysis and Naïve Bayes algorithms respectively. The findings of this study provide implications on how a non-intrusive sensor-based monitoring system comprised of a minimum set of sensors coupled with predictive analytics can be used to detect frail elderly.