Comparison of Context-based Imputation for Missing Intracranial Pressure Signal.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Data Science

Background:

• Intracranial pressure (ICP) monitoring is crucial in neurocritical care.
• ICP signals frequently contain artifacts, complicating analysis and metric calculation.
• Existing artifact detection methods often lack robust data imputation capabilities.

Purpose of the Study:

• To evaluate the feasibility of deep learning models for imputing missing data in ICP signals.
• To compare the performance of novel neural network models against traditional statistical methods for ICP signal reconstruction.

Main Methods:

• Comparison of Seasonal Autoregressive Integrated Moving Average (SARIMA) with deep learning models: Bidirectional Recurrent Network (BRITS), Variational Autoencoder (VAE), and Self-Attention Imputation (SAITS).
• Models reconstructed missing ICP data using surrounding clean signal segments.
• Performance assessed by reconstruction quality, maximum gap length handled, and online processing applicability.

Main Results:

• SARIMA models captured data periodicity but failed to account for post-artifact signal trends, causing misalignment.
• The Self-Attention Imputation (SAITS) model demonstrated superior performance across all evaluated metrics compared to SARIMA and other deep learning approaches.
• Deep learning models effectively reconstructed missing data, particularly for shorter gaps.

Conclusions:

• Modern machine learning techniques, specifically SAITS, can reliably impute short missing segments in ICP signals.
• Enhanced ICP signal reliability is achievable through accurate data imputation.
• Current deep learning imputation methods are not suitable for addressing extended data gaps in ICP monitoring.