Rapid prediction of secondary neurologic decline after traumatic brain injury: a data analytic approach

Abstract

Secondary neurologic decline (ND) after traumatic brain injury (TBI) is independently associated with outcome, but robust predictors of ND are lacking. In this retrospective analysis of consecutive isolated TBI admissions to the R. Adams Cowley Shock Trauma Center between November 2015 and June 2018, we aimed to develop a triage decision support tool to quantify risk for early ND. Three machine learning models based on clinical, physiologic, or combined characteristics from the first hour of hospital resuscitation were created. Among 905 TBI cases, 165 (18%) experienced one or more ND events (130 clinical, 51 neurosurgical, and 54 radiographic) within 48 h of presentation. In the prediction of ND, the clinical plus physiologic data model performed similarly to the physiologic only model, with concordance indices of 0.85 (0.824-0.877) and 0.84 (0.812-0.868), respectively. Both outperformed the clinical only model, which had a concordance index of 0.72 (0.688-0.759). This preliminary work suggests that a data-driven approach utilizing physiologic and basic clinical data from the first hour of resuscitation after TBI has the potential to serve as a decision support tool for clinicians seeking to identify patients at high or low risk for ND.

Figure 1

^aAll cases identified from trauma registry with Head AIS 1–5, Thoracic/Abdominal AIS $\le 1$, and toxicology screen negative for opiates and cocaine, from Nov 2015–Jun 2018. ^bCases included if $\ge$ 30% of first hour continuous physiologic data was available for analysis. TBI = traumatic brain injury; LOS = hospital length of stay; ND = neurologic decline.

Figure 2

A. Ven diagram displaying the prevalence of ND events by subtype. B. Temporal distribution of ND, with y-axis representing the number of patients experiencing ND during each hour, as denoted by the x-axis, color-coded by ND subtype.

Figure 3

Neurologic decline (ND) prediction model performance based on clinical (A), physiologic (B), and combined (C) predictor variables. Individual receiver operating characteristic (ROC) curves demonstrate model performance for predicting ND at specific times (in hours) from presentation, denoted by line color.

Figure 4

Time-dependent areas under the curve (AUC) of receiver operating characteristic (ROC) analysis (solid lines) displayed with 95% confidence intervals(CI, dotted lines) for ND prediction models based on clinical (A), physiologic (B), and combined (C) predictor variables. The time-independent concordance index (C-Index) for each model was 0.72 (95%CI: 0.69–0.76), 0.84 (95%CI: 0.81–0.87), and 0.85 (95%CI: 0.82–0.88), respectively.

Figure 5

For the clinical (A), physiologic (B), and combined (C) neurologic decline (ND) prediction models, each contributing feature is displayed on the y axis. The x axis shows each feature’s Shapley Additive Explanations (SHAP) values. A larger SHAP value denotes a higher log odds ratio that a variable’s value added to the prediction. Values are represented in color ranging from red to blue (high to low). The y axis from top to bottom ranks the variables’ importance, which is the mean of their absolute SHAP values.