Real-Time, Personalized Anomaly Detection in Streaming Data for Wearable Healthcare Devices

Ubiquitous deployment of low cost wearable healthcare devices and proactive monitoring of vital physiological data, are widely seen as a solution for the high costs and risks associated with personal healthcare. The healthcare data generated from these sensors cannot be manually analyzed for anomalies by clinicians due to its scale and therefore automated techniques has to be developed. Present approaches in literature depends on accurate detection of features from the acquired signal which is not always realistic due to noisy nature of the ambulatory physiological data obtained from the sensors. In addition, present anomaly detection approaches require manual training of the system for each patient, due to inherent variations in the morphology of physiological signal for each user. In this chapter, we will first introduce the system architecture for wearable health-care monitoring systems and present discussions on various components involved. Then we discuss on the complexities involved in realizing these methods and highlight key features. We then present our experiences in extracting the ECG segments in real-time and detecting any anomalies in the streams. Particularly, we apply real-time signal processing methods and heuristics to estimate the boundary limits of individual beats from the streaming ECG data. We discuss the importance of designing methods, which are blind to inherent variations among multiple patients and less dependent on the accuracy of the feature extraction. The proposed methods are tested on public database from physionet (QTDB) to validate the quality of results. We highlight and discuss all the significant results and conclude the chapter by proposing some open-ended research questions to be addressed in the near future.