Depth-Based Intervention Detection in the Neonatal Intensive Care Unit Using Vision Transformers.

Area of Science:

• Medical Technology
• Computer Vision
• Neonatal Care

Background:

• Depth cameras offer noncontact, privacy-preserving patient monitoring in Neonatal Intensive Care Units (NICUs).
• Clinical interventions disrupt continuous video monitoring, necessitating manual annotation for system development.
• Automated detection of these events is crucial for efficient system development and future real-time analysis.

Purpose of the Study:

• To develop an automated method for detecting clinical interventions using only depth camera data.
• To investigate the impact of various depth data encoding methods and perspective transformations on detection accuracy.
• To evaluate the performance of a vision transformer (ViT) model for intervention detection in the NICU.

Main Methods:

• A vision transformer (ViT) model was developed using real-world depth data from NICU patients.
• Depth data encoding techniques, including HHA (Horizontal disparity, Height above ground, and Angle with gravity), were explored.
• Perspective transform was applied to address nonoptimal camera placements.

Main Results:

• The best-performing ViT model utilized approximately 85 million trainable parameters.
• The model achieved a sensitivity of 85.6%, precision of 89.8%, and an F1-Score of 87.6%.
• The optimal configuration incorporated both perspective transform and HHA encoding.

Conclusions:

• An effective, depth-data-only method for detecting clinical interventions in the NICU was successfully developed.
• The automated detection significantly reduces the need for manual annotation, streamlining system development.
• This approach holds promise for real-time and retrospective analysis of intervention events in neonatal care.