Background/Aims: Effective treatment for intracerebral hemorrhage (ICH) has remained elusive. This may be partly due to outcome heterogeneity resulting from the variation in patient and hematoma characteristics, which confounds therapeutic effects. Better outcome prediction with more detailed clinical and radiological analyses may aid individualized treatment strategies in ICH. Using machine learning (ML), we investigated whether applying more detailed clinical and radiological characteristics will enhance outcome prediction in ICH.
Methods: We developed a random forest ML model for 6-month outcome prediction using 80% of data as training data from 533 ICH patients of the University of Hong Kong stroke registry, who presented from 2011-2018. The remaining 20% were used for validation. The 6-month outcome was categorized as good (mRS 0-2), poor (mRS 3-5), and death. The detailed outcome prediction model consisted of additional characteristics, including individual components of Glasgow Coma Scale (GCS), worse limb power, specific ICH location, laterality, and Graeb score; and was compared with a standard ML model derived from components of the ICH score.
Results: The detailed ML model has precision and recall rates of 78% and 88% for good outcome; 71% and 71% for poor; and 85% and 73% for death. Compared to the standard model, the detailed model had a significantly higher accuracy (78% vs. 68%, p=0.015). Using recursive feature elimination, the most predictive features included hematoma volume, age, limb power of the affected side, and individual components of the GCS.
Conclusion: ML analysis using detailed clinical and radiological characteristics can enhance outcome prediction in ICH.