Interpretable model committee for monitoring and early prediction of intracranial pressure crises

Traumatic brain injury (TBI) often leads to elevated intracranial pressure (ICP), a dangerous complication associated with increased morbidity and mortality. This study introduces a novel artificial intelligence-based solution for predicting potentially life threatening ICP crises by considering high ICP burden (pressure–time dose), impaired cerebrovascular reactivity, and reduced cerebrospinal compensatory reserve. We propose a committee of models that uses explainable TabNet structures to predict ICP crisis events and identify the most important features driving the predictions to provide additional insights into ICP crisis development over time. The model’s performance was assessed in a large multi-center dataset of long-term high-resolution recordings of physiological signals (ICP and arterial blood pressure) collected in 749 TBI patients. The model achieved very high recall (0.94) with lower precision (0.34) in predicting rare critical events 30 min in advance based on 4 h of previous signal-derived metrics. Explainability analysis revealed that the predictions are primarily influenced by dynamic changes in cerebrospinal pressure–volume relationships which supports the view that patients with reduced cerebrospinal pressure buffering capacity may be at higher risk of presenting ICP elevation. In addition to a well-performing model for ICP crisis prediction in TBI validated in a large dataset, the proposed solution offers the potential for real-time monitoring of feature importance which could help inform the clinical decision-making process and enhance patient outcomes.