Machine Learning Approaches to Intracranial Pressure Prediction in Patients with Traumatic Brain Injury: A Systematic Review

Purpose: Intracranial pressure (ICP) monitoring is a gold standard monitoring modality for severe traumatic brain injury (TBI) patients. The capacity to predict ICP crises could further minimise the rate of secondary brain injury and improve the outcomes of TBI patients by facilitating timely intervention prior to a potential crisis. This systematic review sought (i) to identify the most efficacious approaches to the prediction of ICP crises within TBI patients, (ii) to access the clinical suitability of existing predictive models and (iii) to suggest potential areas for future research. Methods: Peer-reviewed primary diagnostic accuracy studies, assessing the performance of ICP crisis prediction methods within TBI patients, were included. The QUADAS-2 tool was used to evaluate the quality of the studies. Results: Three optimal solutions to predicting the ICP crisis were identified: a long short-term memory (LSTM) model, a Gaussian processes (GP) approach and a logistic regression model. These approaches performed with an area under the receiver operating characteristics curve (AUC-ROC) ranging from 0.86 to 0.95. Conclusions: The review highlights the existing disparity of the definition of an ICP crisis and what prediction horizon is the most clinically relevant. Moreover, this review draws attention to the existing lack of focus on the clinical intelligibility of algorithms, the measure of how algorithms improve patient care and how algorithms may raise ethical, legal or social concerns. The review was registered with the International Prospective Register of Systematic Reviews (PROSPERO) (ID: CRD42022314278).

Automated summary

This systematic review aimed to predict intracranial pressure (ICP) crises in patients with severe traumatic brain injury (TBI) in order to prevent secondary brain injury and improve patient outcomes. The review identified three effective approaches: long short-term memory (LSTM), Gaussian processes (GP), and logistic regression models, with area under the receiver operating characteristics curve (AUC-ROC) ranging from 0.86 to 0.95. The review also highlighted the lack of consensus on the definition of an ICP crisis, the most clinically relevant prediction horizon, and the clinical intelligibility, improvement of patient care, and ethical concerns of algorithms.