Dynamic prediction of mortality after traumatic brain injury using a machine learning algorithm

Abstract

Intensive care for patients with traumatic brain injury (TBI) aims to optimize intracranial pressure (ICP) and cerebral perfusion pressure (CPP). The transformation of ICP and CPP time-series data into a dynamic prediction model could aid clinicians to make more data-driven treatment decisions. We retrained and externally validated a machine learning model to dynamically predict the risk of mortality in patients with TBI. Retraining was done in 686 patients with 62,000 h of data and validation was done in two international cohorts including 638 patients with 60,000 h of data. The area under the receiver operating characteristic curve increased with time to 0.79 and 0.73 and the precision recall curve increased with time to 0.57 and 0.64 in the Swedish and American validation cohorts, respectively. The rate of false positives decreased to ≤2.5%. The algorithm provides dynamic mortality predictions during intensive care that improved with increasing data and may have a role as a clinical decision support tool.

Key Findings on Computed Tomography of the Head that Predict Death or the Need for Neurosurgical Intervention From Traumatic Brain Injury

“…In recent years, machine learning has been used to produce models aimed at predicting clinical outcomes, including after traumatic brain injury. [10][11][12] Furthermore, biomarkers, such as glial fibrillary acidic protein (GFAP), ubiquitin C-terminal hydrolase L1 (UCH-L1), and S-100B, have been explored as ways to identify patients who may experience poor outcomes after brain injury, but these tend to focus on long-term outcomes rather than assisting with immediate decision-making regarding management. 13,14 Similarly, machine learning models have been used to make predictions in the intensive care and long-term settings.…”

“…13,14 Similarly, machine learning models have been used to make predictions in the intensive care and long-term settings. 11,12 In the acute setting, Moyer et al were able to create a machine learning model using Traumabase national data on 2159 adult patients with GCS ≤ 12 to predict need for emergency neurosurgery within 24 hours after moderate or severe brain injury. Their model utilized predictors including GCS score, vital signs, intubation, pupillary abnormalities, administration of osmotherapy, mechanism of injury, and labs.…”

Comparison of machine learning models to predict long-term outcomes after severe traumatic brain injury

“…17 Other approaches, include using ML models with feedback from physicians or models that incorporate additional clinical information over time, may further improve performance. 40,41 Future modeling approaches should combine all three approaches, leveraging the strengths of cranial imaging, physician expertise, and additional clinical information over time, to build accurate models that capture the complexities of patients with severe TBI.…”

Development and validation of an interpretable Markov-embedded multilabel model for predicting risks of multiple postoperative complications among surgical inpatients: a multicenter prospective cohort study

“…Models and clinicians play different roles in this process. The model learns quickly from tens of thousands of surgical records to accelerate clinical cognition and provide a standardized procedural, data-driven, objective, and quantitative decision-making basis 46 . By contrast, clinicians draw from years of experience and provide flexible, experience-based, and subjective assessments.…”