Comparison of Context-based Imputation for Missing Intracranial Pressure Signal

Intracranial pressure (ICP) signal recordings are often riddled with artifacts. While multiple methods have been proposed to automatically detect such artifacts, they often lack tools to handle detected inconsistencies. Simple removal of the invalid fragment may be the safest option but some derivative metrics require computation over long time windows and cannot accept missing samples within the calculation horizon. We aimed to test the feasibility of using deep learning models to accurately impute missing data in the ICP signal.

Different signal imputation methods were used in the context of ICP monitoring in neurocrital care patients. Standard statistical approach called seasonal autoregressive integrated moving average (SARIMA) modeling was used as a baseline and compared with novel neural network models: bidirectional recurrent network (BRITS), variational autoencoder (VAE), and self-attention model (SAITS). The models aimed to reconstruct missing data using clean signals before and after the gap and were assessed based on maximum length of the missing fragment, quality of the reconstruction, and applicability in online processing.

While SARIMA allows for capturing of the periodicity of the data, the models do not take into account the signal after the artifact, leading to misalignment. The SAITS model outperforms SARIMA and the other models in all considered metrics.

With the use of proper modern machine learning techniques, it is possible to fill in short missing parts to improve the reliability of the ICP signal. These methods, however, are not suited for longer gaps.