Background: Intraparenchymal hemorrhage (IPH) is one of the most morbid complications of patients presenting with acute leukemia 1-5. While studies have shown that ~5% of patients will experience intracerebral hemorrhage, with IPH being the most common, it is often difficult to identify those most at risk 2. Presentation of acute leukemia is heterogenous with clinical symptoms and laboratory data ranging tremendously. Initial treatment is focused on efficiently cytoreducing patients while trying to limit complications of treatment such as tumor lysis syndrome 3,4,6. Identifying the patients most at risk of IPH could lead to more personalized decision-making regarding cytoreduction, neurologic monitoring, and other clinical management making such as blood pressure and transfusion goals.

Methods: This study included 162 adult patients who presented to the University of Pennsylvania with acute myelogenous leukemia or acute lymphoblastic leukemia between 2021 and 2022. We gathered patient demographic, laboratory, imaging, and other clinical data from the electronic medical record. We completed a retrospective analysis utilizing Mann Whitney U test for continuous and chi-squared for categorical variables followed by multivariate regression to identify significant risk factors. We then utilized a Balanced Random Forest machine learning model to predict patients at-risk for IPH. There were 35 factors evaluated in the EMR for each patient including demographic features, past medical history, disease characteristics, and laboratory data.

Results: There were 162 patients included in this study with 9 patients (6%) experiencing IPH and 24 patients (15%) with any intracerebral hemorrhage. There were 123 patients (76%) with acute myelogenous leukemia (AML) and 39 patients (24%) with acute lymphoblastic leukemia (ALL). In this overall cohort, 91 (56%) of the patients were male and 71 (44%) of the patients were female. Significant risk factors included d-dimer (OR 1.05, p=0.02) and hemoglobin (OR 0.44, p=0.02). Other important factors included platelet count (OR 0.85, p=0.06), fibrinogen (OR 0.98, p=0.06), and uric acid (OR 1.2, p=0.11). In patient's with AML, those initially treated with low-intensity regimen (e.g., venetoclax and azacitadine) were at mild increased risk of hemorrhage (OR 1.11, p=0.11) compared to high-intensity regimen (e.g., cytarabine and anthracycline) (OR 0.93, p=0.28), though this was not significant. The Balanced Random Forest model performed with a testing accuracy of 91% and AUC 0.96. The model performed with a specificity of 0.90 and sensitivity of 1.0. The model was generalizable with average cross validation accuracy of 93% and standard deviation of 8%. The 5 most important features were similarly fibrinogen, uric acid, d-dimer, platelet count, and hemoglobin.

Conclusions: In this cohort, anemia and elevated d-dimer were significantly associated with increased risk of IPH. Thrombocytopenia, hypofibrinogenemia, and hyperuricemia were also correlated, though not significantly. This machine learning model accurately identified those at highest risk of intraparenchymal hemorrhage and noted similar risk factors as traditional statistical techniques.

This content is only available as a PDF.
2025
Sign in via your Institution