Reproducing and improving one-dimensional convolutional neural networks for arterial blood pressure-based cardiac output estimation

Cardiac output (CO) is an essential indicator of patient hemodynamic status. Monitoring of CO in the intensive care unit has been shown to improve perioperative outcomes by supporting patient fluid management. Arterial blood pressure-based cardiac output (APCO) estimation devices are minimally invasive compared to CO estimation using (transpulmonary) thermodilution, like the PiCCO system or the highly invasive gold standard method using the pulmonary artery catheter (PAC). However, inaccuracy in APCO device estimations during hemodynamically unstable periods, especially in vasodilatory situations, hamper their application in the critical care setting. An approach to improve APCO estimation involves utilizing a one-dimensional convolutional neural network (1D-CNN) to predict stroke volume (SV) from arterial blood pressure (ABP) and patient demographics. Previously published work demonstrated that by pre-training models on SV data from commercial APCO devices and adjusting them with transfer learning using SV data from the PAC, 1D-CNNs have superior performance over the in-use FloTrac APCO device. Preliminary results in the current study showed that by altering model training hyperparameters, model performance was improved further, significantly lowering the absolute error in PAC SV predictions of the original settings model from 13.9 (SD 11.6) mL to 11.7 (SD 11.0) mL for the new settings model (p < 0.001). This result shows promise in further improvement of deep learning-based APCO algorithms and the estimation of CO from ABP in the critical care setting.