Classification of Nursing Care Activities Using Smartwatches

Around 46.8 million people worldwide currently suffer from dementia. More than 7.7 million new cases occur every year. As the prevalence increases with age and our life expectancy trends more and more towards an older population, this can become a growing burden for care homes, the healthcare system and the public. Nursing homes are already suffering from a shortage of nursing staff. To make matters worse, nurses are overwhelmed by the increasing amount of care documentation that takes away their time to interact with and care for their patients. With this work, we therefore evaluate the performance of wearable sensors for automatic recognition and documentation of 10 common nursing tasks using machine learning. Our main goal is to reduce the documentation workload and allow the nursing staff to refocus on patient care by finding the most suitable, robust and accurate combination of sensors and machine learning algorithms for automatic documentation of nursing activities. We evaluated 12 machine learning algorithms using 3D motion data collected at the wrist. Our test data contains 10 activities and was generated in collaboration with two care communities using a Google Pixel Watch GQF4C. While there was no single best solution to classify all activity types, Fast Tree, Generalized Additive Models and Light Gradient Boosting Machine showed superior results compared to all other classifiers. Additional improvements could be achieved by dividing the activities into 3 main groups with best per group classifier types and implementing a quick model calibration per user.