Missing data resilient decision-making for healthcare IoT through personalization: A case study on maternal health

Abstract

<p>Remote health monitoring is an effective method to enable tracking of 
\nat-risk patients outside of conventional clinical settings, providing 
\nearly-detection of diseases and preventive care as well as diminishing 
\nhealthcare costs. Internet-of-Things (IoT) technology facilitates 
\ndevelopments of such monitoring systems although significant challenges 
\nneed to be addressed in the real-world trials. Missing data is a 
\nprevalent issue in these systems, as data acquisition may be interrupted
\n from time to time in long-term monitoring scenarios. This issue causes 
\ninconsistent and incomplete data and subsequently could lead to failure 
\nin decision making. Analysis of missing data has been tackled in several
\n studies. However, these techniques are inadequate for real-time health 
\nmonitoring as they neglect the variability of the missing data. This 
\nissue is significant when the vital signs are being missed since they 
\ndepend on different factors such as physical activities and surrounding 
\nenvironment. Therefore, a holistic approach to customize missing data in
\n real-time health monitoring systems is required, considering a wide 
\nrange of parameters while minimizing the bias of estimates. In this 
\npaper, we propose a personalized missing data resilient decision-making 
\napproach to deliver health decisions 24/7 despite missing values. The 
\napproach leverages various data resources in IoT-based systems to impute
\n missing values and provide an acceptable result. We validate our 
\napproach via a real human subject trial on maternity health, in which 20
\n pregnant women were remotely monitored for 7 months. In this setup, a 
\nreal-time health application is considered, where maternal health status
\n is estimated utilizing maternal heart rate. The accuracy of the 
\nproposed approach is evaluated, in comparison to existing methods. The 
\nproposed approach results in more accurate estimates especially when the
\n missing window is large.<br /></p>