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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Intensive Care Medicine

Background:

• Cerebrovascular autoregulation (CA) is crucial for maintaining stable cerebral blood flow (CBF) despite fluctuating cerebral perfusion pressure (CPP).
• Traumatic brain injury (TBI) can disrupt CA, necessitating accurate monitoring.
• Current methods like the pressure-reactivity index (PRx) have limitations in dynamically assessing CA states.

Purpose of the Study:

• To develop and evaluate a novel deep representation learning model for dynamic CA state monitoring using arterial blood pressure (ABP) and intracranial pressure (ICP) time series.
• To understand the model's behavior across different CA states (active vs. inactive).
• To create feature-based classification models that outperform PRx in discriminating between active and inactive CA states.

Main Methods:

• A deep representation learning model was employed to analyze concurrent ABP and ICP time series.
• Cross-spectral analysis was used to evaluate the model's latent space and reconstruction error outputs.
• Various classifiers were trained using features derived from the deep learning model and compared against PRx threshold-based classification.

Main Results:

• Distinct frequency components were identified as triggers for increased ABP and ICP reconstruction errors during inactive CA states.
• The best feature-based classification model significantly outperformed PRx.
• Classification performance improved from a median of 0.14 precision and 0.62 recall to 0.77 precision and 0.87 recall.

Conclusions:

• The developed deep learning model provides valuable insights into CA dynamics and accurately classifies active and inactive CA states.
• Enhanced detection of inactive CA states holds potential for improving patient outcomes following TBI.
• This approach offers a more sensitive method for monitoring CA compared to existing techniques.