Machine Learning-Based Automatic Classification of Video Recorded Neonatal Manipulations and Associated Physiological...

Area of Science:

• Neonatal Intensive Care Unit (NICU) research
• Machine learning applications in healthcare
• Physiological monitoring in neonates

Background:

• Neonatal care involves frequent manipulations that can affect infant well-being.
• Objective monitoring of these manipulations and their physiological impact is crucial.
• Current methods for tracking neonatal interventions are often manual and subjective.

Purpose of the Study:

• To develop and validate a machine learning system for automatic recognition of neonatal manipulations.
• To assess the impact of different neonatal manipulations on physiological parameters like heart rate and oxygen saturation.
• To investigate the feasibility of using deep learning for objective documentation of NICU care.

Main Methods:

• A retrospective observational study utilizing video and physiological data (HR, SpO2) from ten neonates in NICUs.
• Development of a deep learning system combining Inception-v3 CNN and LSTM for manipulation classification.
• Validation using a leave-one-out strategy with 8-second video clips of manipulation activity.

Main Results:

• The machine learning system achieved 95% accuracy in training and 85% in validation.
• Diaper changes significantly altered HR and SpO2 in preterm infants (<32 weeks gestation).
• Patting and tube feeding were associated with significant HR changes in more mature infants (≥32 weeks gestation).

Conclusions:

• The presented deep learning system accurately classifies and documents neonatal manipulations.
• NICU care interventions demonstrably impact neonatal physiological parameters.
• This technology offers potential for objective assessment and improved understanding of neonatal care practices.