Deep learning to quantify care manipulation activities in neonatal intensive care units

Area of Science:

• Neonatal care
• Neuroscience
• Artificial Intelligence

Background:

• Early-life stress poses significant risks for neurodevelopmental impairments.
• Monitoring neonatal stress in NICUs is critical for infant well-being.

Purpose of the Study:

• To quantify duration, frequency, and physiological responses of care manipulation activities in neonates.
• To develop and evaluate a deep learning method for automated stress monitoring.

Main Methods:

• Utilized 289 hours of bedside video and physiological data from 27 neonates across 2 NICUs.
• Developed a deep learning model to detect manipulation activities, estimate duration/frequency, and integrate physiological signals.
• Validated the method against human annotations and assessed its effectiveness in estimating physiological responses.

Main Results:

• Deep learning method achieved statistically equivalent results to human annotations for activity duration and frequency (13.8% relative error tolerance).
• The method effectively estimated short-term physiological responses to care activities.
• Successfully detected activities with significant physiological deviations and quantified neonatal infant stressor scale scores.

Conclusions:

• The developed deep learning approach offers a reliable tool for objective neonatal stress monitoring in NICUs.
• This technology can aid in assessing the impact of care activities on infant stress levels.
• Potential for improving early detection and management of neurodevelopmental risks associated with neonatal stress.