Evolutionary type-2 fuzzy blood gas models for artificially ventilated patients in ICU

This paper proposes a new modelling and optimization architecture for improving the prediction accuracy of arterial blood gases (ABG) in the SOPAVent model (Simulation of Patients under Artificial Ventilation). The three ABG parameters monitored by SOPAVent are the partial arterial pressure of oxygen (PaO 2 ), the partial arterial pressure of carbon-dioxide (PaCO 2 ) and the acid-base measurement (pH). SOPAVent normally produces the initial ABG predictions and also the ABG predictions after any changes in ventilator settings are made. Two of SOPAVent's sub-models, namely the relative dead-space (Kd) and the carbon-dioxide production (VCO 2 ) were elicited using interval type-2 fuzzy logic system. These models were then tuned using a new particle swarm optimization (nPSO) algorithm, via a single objective optimization approach. The new SOPAVent model was then validated using real patient data from the Sheffield Royal Hallamshire Hospital (UK). The performance of the new SOPAVent model was then compared with its previous version, where Kd and VCO 2 were modeled using a neural-fuzzy system (ANFIS). For the initial ABG predictions, significant improvements were observed in the mean absolute error (MAE) and correlation coefficient (R) for PaCO 2 and pH. When the ventilator settings were changed, significant improvements were observed for the prediction of pH and other improvements were also observed for the prediction of PaCO 2 .