The study focused on developing and evaluating a method to analyze neonatal care procedures in NICUs using a multi-modal dataset of video recordings and physiological data. The dataset included an extension of recordings collected using the NEO device with a camera module, capturing routine care procedures. The study aimed to minimize interference with clinical workflow by orienting the camera towards the neonates during care activities. Data was collected from urban and rural NICUs in India, with 27 neonates included in the study.
Video recordings were manually inspected and annotated for care manipulation activities, resulting in 650 identified activities. Deep learning techniques, specifically the SlowFast model and ActionFormer, were used to analyze the videos and detect these activities. Physiological data, such as heart rates and SpO2 levels, were integrated with video analysis to quantify stressors on neonates during care procedures. The study also explored the quantification of stress using the Neonatal Infant Stressor Scale (NISS) for diaper changes.
A 5-fold cross-validation was conducted to evaluate the method’s performance in temporal activity localization and quantification of physiological responses. Precision-recall curves were used to assess activity detection, while paired t-tests were employed to compare results between algorithmic predictions and human annotations. The study demonstrated the effectiveness of deep learning methods in analyzing neonatal care procedures and quantifying stress responses, providing valuable insights for improving care practices in NICUs.
Source link
Source link: https://www.nature.com/articles/s41746-024-01164-y