Noninvasive Arterial CO₂ Estimation for Neonatal Ventilation Using Gated Recurrent Neural Networks in the Preterm...

Area of Science:

• Biomedical Engineering
• Computational Neuroscience
• Neonatal Physiology

Background:

• Accurate monitoring of arterial partial pressure of carbon dioxide (PaCO2) is critical for managing mechanically ventilated neonates.
• Current clinical methods, including using end-tidal partial pressure of carbon dioxide (PetCO2) as a surrogate or with an offset, have limitations in precision.

Purpose of the Study:

• To develop and validate advanced neural network models for noninvasive PaCO2 estimation in preterm lambs.
• To compare the accuracy of these models against established clinical methods.

Main Methods:

• Two Gated Recurrent Unit (GRU) neural network models were designed using time-series data.
• Model 1 utilized univariate PetCO2 data; Model 2 incorporated multivariate noninvasive respiratory and vital measurements.
• Data from 1,182 arterial blood gas analyses in 58 preterm lambs were used for training and validation.

Main Results:

• The GRU model using only PetCO2 showed lower accuracy than clinical methods.
• The PetCO2+Offset method demonstrated improved estimation accuracy.
• The multivariate GRU model incorporating vital and respiratory measurements achieved the highest accuracy in PaCO2 estimation.

Conclusions:

• Advanced GRU neural network models can accurately estimate PaCO2 in mechanically ventilated neonates.
• Multivariate noninvasive data significantly enhances the accuracy of PaCO2 estimation compared to univariate or simple offset methods.
• These findings suggest a potential for improved noninvasive monitoring strategies in neonatal intensive care.