Development of an Active Cerebrovascular Autoregulation Model Using Representation Learning: A Proof of Concept Study With Experimental Data

It remains a challenge to monitor cerebrovascular autoregulation (CA) reliably and dynamically in an intensive care unit. The objective was to build a proof-of-concept active CA model exploiting advances in representation learning and the full complexity of the arterial blood pressure (ABP) and intracranial pressure (ICP) signal and outperform the pressure reactivity index (PRx). A porcine cranial window CA data set (n = 20) was used. ABP and ICP signals were preprocessed and downsampled to 20 Hz. Quadriphasic CA state labels were assigned to each piglet's CA curve and projected on their preprocessed ABP and ICP time series. Windowed ABP and ICP segments of 300 seconds, reflecting active CA, were used to optimize a neural network to reconstruct its own input. Reconstruction error of ABP and ICP were compared between active CA and inactive CA, and assessed together with PRx over quadriphasic CA states. The study confirmed that the optimized model achieved stellar reconstruction quality of ABP and ICP segments that derived from active CA while reconstruction quality deteriorated for segments that came from inactive CA. ABP and ICP reconstruction errors steadily increased concurrently with cerebral blood flow deviation from baseline. A significant interaction between variable and CA state showed that the model captured the differential behavior of CA with increasing vs decreasing cerebral perfusion pressures and offered improved discriminative ability regarding PRx. The present work showed that an active CA model can be built using advanced representation learning and the full complexity of 300-second ABP and ICP segments. On assessment in an experimental data set, relevant CA state information was present in both lower and higher frequencies of ABP and ICP. Improved discriminative ability between CA states was attained regarding PRx, which focuses only on slow-wave ABP and ICP information.