Bottle-Feeding Intervention Detection in the NICU.

Area of Science:

• Medical technology
• Artificial intelligence in healthcare
• Neonatal care

Background:

• Video monitoring in Neonatal Intensive Care Units (NICUs) offers potential for enhanced patient care.
• Semi-automated charting of clinical events, like bottle feeding, can improve accuracy and efficiency.
• Current retrospective charting methods have limitations compared to contemporaneous recording.

Purpose of the Study:

• To develop a video-based system for detecting bottle-feeding events in NICU patients.
• To improve oral feeding assessments and patient growth monitoring.
• To address data scarcity in NICU datasets using transfer learning and data expansion.

Main Methods:

• Utilized transfer learning with a pretrained VGG-16 model for image classification.
• Implemented a data expansion technique using similar-feature images from public sources.
• Analyzed network activation maps to visualize domain gap and impact of data expansion.

Main Results:

• Achieved an increase in detection sensitivity of over 18% after data expansion.
• Demonstrated that data expansion reduces the domain gap between source (ImageNet) and target (NICU) datasets.
• Network activation map analysis confirmed improved knowledge transfer proficiency.

Conclusions:

• Video-based detection of bottle feeding is feasible and beneficial for NICU patient monitoring.
• Data expansion is an effective strategy to improve the performance of AI models in specialized domains like NICU.
• This approach can lead to more accurate clinical event charting and support feeding assessments.