Incremental Anomaly Detection with Guarantee in the Internet of Medical Things

Abstract

The Internet of Medical Things (IoMT), aided by learning-enabled components, is becoming increasingly important in health monitoring. However, the IoMT-based system must be highly reliable since it directly interacts with the patients. One critical function for facilitating reliable IoMT is anomaly detection, which involves sending alerts when a medical device’s usage pattern deviates from normal behavior. Due to the safety-critical nature of IoMT, the anomaly detectors are expected to have consistently high accuracy and low error, ideally being bounded with a guarantee. Besides, since the IoMT-based system is non-stationary, the anomaly detector and the performance guarantee should adapt to the evolving data distributions. To tackle these challenges, we propose a framework for incremental anomaly detection in IoMT with a Probably Approximately Correct (PAC)-based two-sided guarantee, guided by a human-in-the-loop design to accommodate shifts in anomaly distributions. As a result, our framework can improve detection performance and provide a tight guarantee on False Alarm Rate (FAR) and Miss Alarm Rate (MAR). We demonstrate the effectiveness of our design using synthetic data and the real-world IoMT monitoring platform VitalCore.