Abstract TP127: Application Of Machine Learning Algorithms For Prediction Of Hematoma Expansion In Spontaneous Intracerebral Hemorrhage

Abstract

Introduction: Hematoma expansion (HE) is a known prognostic indicator of spontaneous intracerebral hemorrhage (sICH). Although several scores exist for prediction of HE, universal adoption has been limited due to their lack of sensitivity and specificity. As machine learning (ML) algorithms have shown promise in the stroke field, here we examine the predictive accuracy of several ML algorithms for HE in sICH patients. Methods: We retrospectively analyzed demographic, clinical data, and radiographic signs of patients with sICH in our 2-hospital database. A total of 61 clinical, imaging, and treatment variables were included in the study. Nine ML models were applied: Adaptive Boost (AdaBoost), Bernoulli Naïve Bayes (BNB), Decision Tree (DT), Gaussian Naïve Bayes (GNB), Logistic Regression (LR), Multi-Layer Perceptron (MLP), Multinomial Naïve Bayes models (MNB), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). All models were trained to predict HE. Model accuracy was assessed using the area under characteristic curve (AUC). Results: Of the 301 patients with sICH, 63 developed HE (21.93%). Of the 9 models studied, MLP had the highest AUC score (0.93±0.042), followed by XGBoost (0.80±0.06). All models demonstrated moderate to high predictive accuracy (AUC 0.64-0.93) for HE. The top predictors in MLP were Baseline NIHSS score, HDL, aPPT, time from last known well to ER, initial hematoma volume, and island sign. MLP had moderate sensitivity of 0.46±0.17 and high specificity of 0.99±0.02. GNB, however, showed the highest sensitivity at 0.86±0.06 and a moderate specificity of 0.65±0.07. Five of the 9 models ranked time last known well to ER presentation as a predictor of HE. Conclusion: In our study, we found all ML models applied had moderate to high predictive accuracy for prediction of HE in sICH, with MLP having the highest accuracy of all models. Future studies examining the use of these algorithms are warranted.