Authors:
Stefan Lüdtke
;
Albert Hein
;
Frank Krüger
;
Sebastian Bader
and
Thomas Kirste
Affiliation:
University of Rostock, Germany
Keyword(s):
Sleep Detection, Actigraphy, Hidden Markov Model, Machine Learning, Dementia.
Abstract:
Actigraphy can be used to examine the sleep pattern of patients during the course of the day in their common
environment. However, conventional sleep detection algorithms may not be appropriate for real-world
daytime sleep detection, since they tend to overestimate the sleep duration and have only been validated for
nighttime sleep in a laboratory setting. Therefore, we evaluated the performance of a set of new sleep detection
algorithms based on machine learning methods in a real-world setting and compared them to two conventional
sleep detection algorithms (Cole’s algorithm and Sadeh’s algorithm). For that, we performed two studies with
(1) healthy young adults and (2) nursing home residents with Alzheimer’s dementia. The conventional algorithms
performed poorly for these real-world data sets, because they are imbalanced with respect to sensitivity
and specificity. A more balanced Hidden Markov Model-based algorithm surpassed the conventional algorithms
for both data sets. Using th
is algorithm leads to an improved accuracy of 4.1 percent points (pp) and
23.5 pp, respectively, compared to the conventional algorithms. The Youden-Index improved by 7.3 and 7.7,
respectively. Overall, for a real-world setting, the HMM-based algorithm achieved a performance similar to
conventional algorithms in a laboratory environment.
(More)