Realtime Bidirectional Data Imputation of Multivariate Time Series for Anomaly Detection in Internet of Things Devices

Abstract

Remote health monitoring systems are gaining momentum. Indicators of underlying conditions are discoverable through patterns of ubiquitous patient health data which include continuous multivariate time series data (MTSD). Classification of data in resource constraint devices has many challenges and requires extensive analyses. In a time-critical domain, missing values can profoundly impact the accuracy of classification models. Additionally, due to the limited battery life of IoT devices and the time-sensitive nature of the medical domain, the trade-off between model accuracy, latency, and power consumption should be entertained as a balancing factor. 
In this paper, we develop a system which adapts a powerful imputation technique, known as bidirectional recurrent imputation for time series (BRITS), to be performed in real-time for MTSD. The model operates concurrently with anomaly detection algorithms to address resource consumption. To test the introduced system, we investigated a healthcare dataset of 4K patients, including 48 hours of MTSD, labelled appropriately on in-hospital survivability. A random forest algorithm executes the labelling task and returns classification performance. Data suggests that the healthcare dataset was imputed most accurately compared to other imputation techniques under a real-time condition. In future work, we plan to clarify the resource costs for anomaly detection in real systems. Our data indicate that real-time adaptation of BRITS improves classification under a time-critical situation.