Accurate prediction of neurologic changes in critically ill infants using pose AI

Area of Science:

• Medical technology
• Artificial intelligence
• Neonatal neurology

Background:

• Infant neurological status is critical but often assessed subjectively.
• Continuous, objective monitoring is needed in the Neonatal Intensive Care Unit (NICU).
• Existing methods like electroencephalograms (EEG) can be invasive and provide intermittent data.

Purpose of the Study:

• To evaluate the feasibility of using computer vision-based pose estimation (pose AI) for neuro-telemetry in neonates.
• To determine if pose AI can predict neurological changes and diagnoses from video data alone.
• To develop a scalable, minimally invasive method for continuous neurological assessment in the NICU.

Main Methods:

• Collected 4,705 hours of video data from 115 infants in the NICU, linked to EEG.
• Trained a deep learning pose algorithm to accurately predict infant anatomical landmarks.
• Developed classifiers using pose AI landmarks to predict sedation and cerebral dysfunction.

Main Results:

• The pose AI algorithm demonstrated high accuracy in predicting anatomical landmarks (ROC-AUCs 0.83-0.94).
• Pose AI successfully predicted infant sedation levels (ROC-AUCs 0.87-0.91).
• Pose AI accurately predicted cerebral dysfunction (ROC-AUCs 0.76-0.91), correlating with EEG diagnoses.

Conclusions:

• Deep learning with pose AI is feasible for neuro-telemetry in the NICU.
• Pose AI can predict key neurological states and diagnoses from video data, potentially replacing or supplementing EEG.
• This approach offers a scalable, minimally invasive solution for continuous infant neurological monitoring.