• deep learning
• neural networks
• patient monitoring
• privacy-preserving

Patient monitoring in hospital settings is crucial for assessing well-being, detecting anomalies, and supporting clinical decision-making. Tasks such as action recognition and video classification enable automated analysis of patient movements and behaviors, improving healthcare outcomes. However, the collection and use of visual data raise significant privacy concerns, as sensitive attributes could lead to re-identification risks. To address this challenge, we introduce a framework for the automatic anonymization of patient monitoring images, ensuring privacy while preserving the original data distribution. Our approach leverages neural networks to synthesize anonymized videos that maintain essential visual features, such as pose and action, while obfuscating privacy-sensitive attributes. Unlike traditional anonymization techniques, which often degrade data utility, our model generates realistic, privacy-preserving images with seamless integration into the original scene. To evaluate our method, we introduce a diverse dataset encompassing various poses and actions, tested in both healthcare and non-healthcare conditions. Experimental results demonstrate that our approach effectively anonymizes images while retaining critical information for downstream deep learning applications.