Machine Learning Based Approach for Anomaly Detection in Healthcare IoT Systems with Variational Autoencoders for Data Integrity and Security Enhancement

Abstract

IoT solutions improve the quality of patient care in the healthcare system by allowing patients to connect to devices that monitor and evaluate essential health data. However, the massive amounts of data generated carry serious security threats, which are self-inflicted in detecting abnormal system behaviours that can indicate a possible breach or system failure. Usually, anomaly detection is done by employing heuristic algorithms on labeled datasets. However, in real-world healthcare applications, several privacy barriers in data dump hinder obtaining labeled datasets. This research proposes a new model that uses spatial-temporal variational autoencoders (VAEs) as an unsupervised model for the detection of anomalies in healthcare IoT. The model works by detecting anomalies in unlabeled data that the model capacities make use of incomplete patterns, thereby detecting modes of anomalies that can be related to security, integrity issues. The VAE model integrates various IoT devices and environmental studies by generating a figure of the normal data set and treating its deviations as anomalies. Tests performed on simulated healthcare System IoT data recorded 95% accuracy, 95% precision, 75% recall, and 98% specificity to give an F1 score of 83.82% and AUC of 92%, hence demonstrating efficient and accurate detection of anomalies.