Predicting Poor Sleep in People Living with Dementia: An Analysis of In‐Home Activities and Physiology Measurements During the Daytime

Abstract

AbstractBackgroundThe use of Internet of Things (IoT) technologies can help with remote monitoring of People Living With Dementia (PLWD) at their homes, ensuring effective and prompt provision of healthcare services. This study provides an analysis of data collected as part of a long‐term research project initiated by the UK Dementia Research Institute’s Care Research and Technology Centre, which aims to understand and slow down the progression of dementia. The goal of this work is to examine the relationship between daytime living status (i.e. activities and psychiatric measurements) and nighttime sleep quality, as poor sleep has been linked to the development of dementia. Early prediction of poor nighttime sleep through monitoring daytime activities may offer an opportunity for intervention, leading to an improvement in sleep quality. In order to achieve this goal, we have developed a classification model to predict poor sleep based on daytime living status.MethodThe activity data is from sensors placed throughout each participant’s home, tracking their movements between 6:00 AM and 7:00 PM each day. The physiological measurements are also taken daily, primarily in the daytime. Sleep is considered poor if more than 50∖% of the time is spent in the “Awake” state out of the total sleep duration (the sum of four sleep states: “Awake”, “Deep”, “Light”, “REM”). Machine learning classification models are applied to predict whether nighttime sleep is poor or not. Also, feature importance analysis can be conducted to identify activity patterns that contribute to poor sleep.ResultThe dataset includes 91 participants with dementia and the collection period spans from June 29th, 2021 to December 31st, 2022. A Gradient Boosting tree classifier was trained and evaluated using 10‐fold cross‐validation. 7% of the data are positive samples. The average sensitivity was 81.1% with a standard deviation of 6.6%, and the average specificity was 78.3% with a standard deviation of 2.3%.ConclusionThe study suggests that further improvement can be made by analysing time‐dependency within the data and utilising more advanced models. The combination of IoT and machine learning shows the potential to greatly benefit healthcare services for PLWD on a daily basis.