Adaptive Anomaly Detection in Cardiovascular Time Series Through Deep Reinforcement and Active Learning

Abstract

In this research, a novel framework is proposed for detecting anomalies in time series and medical imaging data, leveraging Deep Reinforcement Learning (DRL) and Active Learning (AL) methodologies. This approach integrates DRL’s dynamic learning capabilities with AL’s efficiency in handling label scarcity, specifically for a cardiovascular disease dataset. This dataset encompasses various features, including demographic (age, gender), anthropometric (height, weight), clinical (blood pressure, cholesterol, glucose levels), and lifestyle factors (smoking, alcohol intake, physical activity), alongside the binary target variable indicating the presence of cardiovascular disease. The proposed framework employs RLAD (Reinforcement Learning for Anomaly Detection) to identify anomalies in the cardiovascular dataset. The RL component is designed to adaptively learn from the continuous action space, efficiently detecting outliers in the clinical and lifestyle features that could indicate potential cardiovascular anomalies. Concurrently, the AL component selectively queries the most informative data points to enhance the labeling process, addressing the challenge of limited labeled data. The results of the method show the efficacy of the proposed approach in accurately identifying anomalies, outperforming traditional methods through methods of precision, recall, and F1-score. This hybrid DRL-AL framework not only improves convergence rates but also adapts effectively to the evolving nature of cardiovascular health data. It highlights the potential of advanced machine learning techniques in enhancing the early identification and diagnosis of cardiovascular diseases, paving the way for improved patient outcomes and healthcare strategies.