A multimodal vision dataset for nursing action recognition and quality assessment in NICU

This paper presents the Nursing Action in Multimodal Vision (NAMV) dataset, a large-scale collection of synchronized multimodal recordings capturing nursing procedures in a simulated neonatal intensive care unit (NICU). The dataset focuses specifically on hand-based interactions during bedside care, designed to support clinically meaningful research in workflow recognition and procedural quality assessment. In accordance with established clinical protocols, seven high-frequency nursing procedures are systematically decomposed into 19 sub-actions, each annotated with frame-level temporal boundaries that span the preparation, execution, and completion phases. Data were acquired using an Orbbec Femto Bolt RGB-D sensor, which integrates an RGB camera, a depth camera, and active infrared illumination; corresponding per-frame 3D point clouds are also provided. Notably, this release includes RGB images, depth maps, infrared imagery, and 3D point clouds. To the best of our knowledge, NAMV is the first multimodal dataset for NICU nursing workflows to integrate multiple sensing modalities with expert-annotated quality ratings at both the sub-action and event levels. This resource provides a valuable foundation for advancing research in multimodal action recognition, temporal segmentation, and skill proficiency modeling, focusing on the standardized operational behaviors of nurses in the NICU. Its immediate applications are in nursing operation quality assessment, skill proficiency modeling, and simulation-based training, with the potential to support future research on nurse-infant interaction.