Representation Learning for Cerebrovascular Autoregulation: A Model Study With Experimental Data Classification

Cerebrovascular autoregulation (CA) maintains adequate cerebral blood flow (CBF) during changes in cerebral perfusion pressure (CPP) and can be dynamically altered after traumatic brain injury (TBI). A novel deep representation learning model was developed to dynamically monitor CA state using concurrent time series of arterial blood pressure (ABP) and intracranial pressure (ICP). The latter model showed a significant increase in ABP and ICP reconstruction error concurrent with increases or decreases in CBF. The aim of this study is to understand the model's behavior across CA states, and to develop classification models based on derived features that can outperform the pressure-reactivity index (PRx) in discriminating between active -and inactive CA states.

The latent space output of the deep learning model, and ABP and ICP reconstruction output time series were evaluated using cross-spectral analysis. Various classifiers were compared to discriminate between CA states using features derived from the novel deep representation learning model and were contrasted with PRx threshold-based classification.

The present model study elucidated that increased ABP or ICP reconstruction errors following inactive CA were induced by distinct frequency components. Consequently, the best feature-based model outperformed the currently used PRx signal, improving CA state classification from a median 0.14 precision and 0.62 recall to 0.77 precision and 0.87 recall.

This work sheds insight into the model's features and provides support for its ability to classify dynamic active and inactive CA states more accurately.

Improved detection of inactive CA could improve outcome after TBI.