Human–object interaction (HOI) understanding in pediatric wards is essential for monitoring abnormal behaviors and preventing fall-related accidents. Unlike generic HOI scenarios, clinical safety monitoring requires fine-grained reasoning over human pose to identify subtle but high-risk actions such as leaning outside the bed or climbing over the rails. Existing transformer-based HOI detectors primarily rely on bounding-box and appearance cues, making them insufficient for modeling detailed body configurations. We propose a pose-aware HOI (PA-HOI) detection framework that implicitly embeds skeleton structure into the interaction reasoning process. The core contribution is a pose-aware graph-attention module that introduces a global pose context node and performs structured message passing across interaction proposals, enabling the model to capture discriminative limb orientations and body–object spatial dependencies. We evaluate our method on a private pediatric-ward dataset containing nine abnormal behaviors and ten environmental entities. The proposed approach achieves a mean Average Precision of 85.78%, substantially outperforming other contemporary HOI frameworks. These results demonstrate the effectiveness of integrating skeleton-aware structural information for reliable clinical safety monitoring.