Machine learning prediction of the adverse outcome for nontraumatic subarachnoid hemorrhage patients

Abstract

Objective: Subarachnoid hemorrhage (SAH) is often devastating with increased early mortality, particularly in those with presumed delayed cerebral ischemia (DCI). The ability to accurately predict survival for SAH patients during the hospital course would provide valuable information for healthcare providers, patients, and families. This study aims to utilize electronic health record (EHR) data and machine learning approaches to predict the adverse outcome for nontraumatic SAH adult patients.

Methods: The cohort included nontraumatic SAH patients treated with vasopressors for presumed DCI from a large EHR database, the Cerner Health Facts^® EMR database (2000-2014). The outcome of interest was the adverse outcome, defined as death in hospital or discharged to hospice. Machine learning-based models were developed and primarily assessed by area under the receiver operating characteristic curve (AUC).

Results: A total of 2467 nontraumatic SAH patients (64% female; median age [interquartile range]: 56 [47-66]) who were treated with vasopressors for presumed DCI were included in the study. 934 (38%) patients died or were discharged to hospice. The model achieved an AUC of 0.88 (95% CI, 0.84-0.92) with only the initial 24 h EHR data, and 0.94 (95% CI, 0.92-0.96) after the next 24 h.

Interpretation: EHR data and machine learning models can accurately predict the risk of the adverse outcome for critically ill nontraumatic SAH patients. It is possible to use EHR data and machine learning techniques to help with clinical decision-making.

Figure 1

Relation of the observational window and prediction window in the two prediction scenarios. T ₀ and T_e denoted the visit start and end time.

Figure 2

Variables and corresponding odds ratios in the elastic net regularized logistic regression model for predicting the risk of the adverse outcome with the first 24 h EHR data for nontraumatic SAH adult patients. Clipping of aneurysm is categorized as three levels: perform clipping, not perform clipping but performed other procedures, none of the procedures is performed based on the EHR data. In the final predictive model, perform clipping, and not perform clipping but performed other procedures are included.

Figure 3

Top two panels are Venn diagrams for variables in prediction Scenario 1(left) and Scenario 2 (right). A total of 15 and 94 predictors were commonly included in the final predictive models in Scenario 1 and 2, respectively. Their detailed predictor names and corresponding odds ratios were shown in supplemental materials (Fig. S6–S7). The bottom panel is the odds ratios of 8 predictors that were included in all six prediction models for the two scenarios.