Noninvasive estimation of PaCO from volumetric capnography in animals with injured lungs: an Artificial Intelligence approach.

To investigate the feasibility of non-invasively estimating the arterial partial pressure of carbon dioxide (PaCO) using a computational Adaptive Neuro-Fuzzy Inference System (ANFIS) model fed by noninvasive volumetric capnography (VCap) parameters. In 14 lung-lavaged pigs, we continuously measured PaCO with an optical intravascular catheter and VCap on a breath-by-breath basis. Animals were mechanically ventilated with fixed settings and subjected to 0 to 22 cmHO of positive end-expiratory pressure steps. The resultant 8599 pairs of data points - one PaCO value matched with twelve Vcap and ventilatory parameters derived in one breath - fed the ANFIS model. The data was separated into 7370 data points for training the model (85%) and 1229 for testing (15%). The ANFIS analysis was repeated 10 independent times, randomly mixing the total data points. Bland-Altman plot (accuracy and precision), root mean square error (quality of prediction) and four-quadrant and polar plots concordance indexes (trending ability) between reference and estimated PaCO were analyzed. The Bland-Altman plot performed in 10 independent tested ANFIS models showed a mean bias between reference and estimated PaCO of 0.03 ± 0.03 mmHg, with limits of agreement of 2.25 ± 0.42 mmHg, and a root mean square error of 1.15 ± 0.06 mmHg. A good trending ability was confirmed by four quadrant and polar plots concordance indexes of 95.5% and 94.3%, respectively. In an animal lung injury model, the Adaptive Neuro-Fuzzy Inference System model fed by noninvasive volumetric capnography parameters can estimate PaCO with high accuracy, acceptable precision, and good trending ability.