Identifying Modifiable Predictors of Patient Outcomes After Intracerebral Hemorrhage with Machine Learning

Andrew N. Hall*, Bradley Weaver, Eric Liotta, Matthew B. Maas, Roland Faigle, Daniel K. Mroczek, Andrew M. Naidech

*Corresponding author for this work

Research output: Contribution to journalArticle

Abstract

Background/Objective: Demonstrating a benefit of acute treatment to patients with intracerebral hemorrhage (ICH) requires identifying which patients have a potentially modifiable outcome, where treatment could favorably shift a patient’s expected outcome. A decision rule for which patients have a modifiable outcome could improve the targeting of treatments. We sought to determine which patients with ICH have a modifiable outcome. Methods: Patients with ICH were prospectively identified at two institutions. Data on hematoma volumes, medication histories, and other variables of interest were collected. ICH outcomes were evaluated using the modified Rankin Scale (mRS), assessed at 14 days and 3 months after ICH, with “good outcome” defined as 0–3 (independence or better) and “poor outcome” defined as 4–6 (dependence or worse). Supervised machine learning models identified the best predictors of good versus poor outcomes at Institution 1. Models were validated using repeated fivefold cross-validation as well as testing on the entirely independent sample at Institution 2. Model fit was assessed with area under the ROC curve (AUC). Results: Model performance at Institution 1 was strong for both 14-day (AUC of 0.79 [0.77, 0.81] for decision tree, 0.85 [0.84, 0.87] for random forest) and 3 month (AUC of 0.75 [0.73, 0.77] for decision tree, 0.82 [0.80, 0.84] for random forest) outcomes. Independent predictors of functional outcome selected by the algorithms as important included hematoma volume at hospital admission, hematoma expansion, intraventricular hemorrhage, overall ICH Score, and Glasgow Coma Scale. Hematoma expansion was the only potentially modifiable independent predictor of outcome and was compatible with “good” or “poor” outcome in a subset of patients with low hematoma volumes, good Glasgow Coma scale and premorbid modified Rankin Scale scores. Models trained on harmonized data also predicted patient outcomes well at Institution 2 using decision tree (AUC 0.69 [0.63, 0.75]) and random forests (AUC 0.78 [0.72, 0.84]). Conclusions: Patient outcomes are predictable to a high level in patients with ICH, and hematoma expansion is the sole-modifiable predictor of these outcomes across two outcome types and modeling approaches. According to decision tree analyses predicting outcome at 3 months, patients with a high Glasgow Coma Scale score, less than 44.5 mL hematoma volume at admission, and relatively low premorbid modified Rankin Score in particular have a modifiable outcome and appear to be candidates for future interventions to improve outcomes after ICH.

Original languageEnglish (US)
JournalNeurocritical Care
DOIs
StateAccepted/In press - 2020

Keywords

  • Cerebral hemorrhage
  • Machine learning
  • Stroke

ASJC Scopus subject areas

  • Clinical Neurology
  • Critical Care and Intensive Care Medicine

Fingerprint Dive into the research topics of 'Identifying Modifiable Predictors of Patient Outcomes After Intracerebral Hemorrhage with Machine Learning'. Together they form a unique fingerprint.

  • Cite this