Prediction of intracranial pressure crises after severe traumatic brain injury using machine learning algorithms

Avoiding intracranial hypertension after traumatic brain injury (TBI) is a foundation of neurocritical care, to minimize secondary brain injury related to elevated intracranial pressure (ICP). However, this approach at best is reactive to episodes of intracranial hypertension, allowing for periods of elevated ICP before therapies can be initiated. Accurate prediction of ICP crises before they occur would permit clinicians to implement preventive strategies, minimize total time with ICP above threshold, and potentially avoid secondary injury. The objective of this study was to develop an algorithm capable of predicting the onset of ICP crises with sufficient lead time to enable application of preventative therapies. Thirty-six patients admitted to a level I trauma center with severe TBI (Glasgow Coma Scale score < 8) between April 2015 and January 2019 who underwent continuous intraparenchymal ICP monitor placement were retrospectively identified. Continuous ICP data were extracted from each monitoring period (range 4-96 hours of monitoring). An ICP crisis was treated as a binary outcome, defined as ICP > 22 mm Hg for at least 75% of the data within a 5-minute interval. ICP data preceding each ICP crisis were grouped into four total data sets of 1- and 2-hour epochs, each with 10- to 20-minute lead-time intervals before an ICP crisis. Crisis and noncrisis events were identified from continuous time-series data and randomly split into 70% for training and 30% for testing, from a subset of 30 patients. Machine learning algorithms were trained to predict ICP crises, including light gradient boosting, extreme gradient boosting, and random forest. Accuracy and area under the receiver operating characteristic curve (AUC) were measured to compare performance. The most predictive algorithm was optimized using feature selection and hyperparameter tuning to avoid overfitting, and then tested on a validation subset of 5 patients. Precision, recall, F1 score, and accuracy were measured. The random forest model demonstrated the highest accuracy (range 0.82-0.88) and AUC (range 0.86-0.88) across all four data sets. Further validation testing revealed high precision (0.76), relatively low recall (0.46), and overall strong predictive performance (F1 score 0.57, accuracy 0.86) for ICP crises. Decision curve analysis showed that the model provided net benefit at probability thresholds above 0.1 and below 0.9. The presented model can provide accurate and timely forecasts of I