Modeling of Intensive Care Risk Factors for Spreading Depolarizations in Severe Traumatic Brain Injury

Abstract

Spreading depolarizations (SDs) are a mechanism of secondary injury associated with poor outcomes, occurring in 60% of patients who undergo surgery following severe brain trauma. They are triggered by functional metabolic failure in peri-lesional tissue and are associated with several variables that are routinely managed in neurocritical care, such as low blood pressure, tissue hypoxia, and fever. To understand SD risk factors and develop prediction models, here we performed a retrospective analysis of an observational study of 138 patients who underwent electrocorticographic (ECoG) monitoring of SDs. Timestamps of SD occurrence were aligned to hourly nursing chart data, including mean arterial pressure (MAP), intracranial pressure (ICP), core temperature, brain tissue oxygenation (P_brO₂), peripheral oxygen saturation (S_aO₂), fraction of inhaled oxygen (F_iO₂), and heart rate. Blood gas and biochemistry variables of pH, arterial partial pressure of oxygen (P_aO₂) and carbon dioxide (P_aCO₂), and plasma glucose were also assessed. A total of 13,315 h were available from 137 patients, 82 of whom had 2700 SDs in 7067 total hours of monitoring. In univariate analysis of patients with SDs, hours with SDs were associated with significantly lower MAP, brain oxygenation (P_brO₂), heart rate, and P_aCO₂, as well as higher pH, glucose, and P_aO₂. These relationships were even more pronounced for SDs classified as isoelectric. In multivariate analysis, hours with SDs were independently associated with lower MAP, heart rate, and P_aCO₂, as well as higher temperature, S_aO₂, and glucose. Using a 90% to 10% random split of the data for training and testing, respectively, multivariate regression models showed that several systemic variables were independently associated with SD occurrence with moderate performance (area under the curve [AUC] ∼0.72). When a random-effects variable was included to account for unknown within-patient variables, performance improved to AUC = 0.84 with sensitivity = 62.9% and specificity = 88.1%. As a final step, we developed models using the full dataset, including all 137 patients with and without SDs, using the hourly predictors in addition to known fixed-factor, within-patient variables previously established as SD predictors (Rotterdam head CT, pre-hospital hypotension, and Morris-Marshall grade of traumatic subarachnoid hemorrhage). Superior results were achieved with a Poisson model that yielded AUC = 0.916, sensitivity = 66.1%, and specificity = 87.9% in predicting hours in which SDs occurred. These results confirm the importance of key systemic variables as SD risk factors and support the notion that SD occurrence can be influenced by modifiable factors that adversely impact the balance of energy supply–demand in vulnerable tissue. The final prediction model, while requiring further validation and with undetermined generalizability, suggests a potential tool for use in neurocritical care to identify cerebral and systemic states that are associated with SDs and secondary injury. The model thus represents an important step toward more widespread application of results and insights obtained from invasive ECoG monitoring.