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

We present two Gated Recurrent Unit (GRU) neural networks for estimating the arterial partial pressure of carbon dioxide (PaCO2) in mechanically ventilated neonates, using time series data of the end-tidal partial pressure of carbon dioxide (PetCO2) and noninvasive respiratory and vital measurements. To train and validate these models, we use secondary data of preterm lambs. 1,182 arterial blood gas analyses (ABGs) from 58 animals are included. We compare our approaches to two clinical methods: (i) using PetCO2 as a surrogate for PaCO2 and (ii) adjusting PetCO2 with a constant offset based on invasive ABG analysis. Estimation errors are evaluated by comparing model outputs with PaCO2 measurements from ABGs via Bland-Altman analyses. Estimation by PetCO2 and the GRU model using a univariate time series of PetCO2 results the lowest accuracy. The invasive PetCO2+Offset method shows improvement, while the highest accuracy is achieved with a GRU model that incorporates a noninvasive multivariate time series of PetCO2 along with vital and respiratory measurements.