Utilizing Heterogeneous Feature Dimensionality Reduction and Stacking Model for Hematoma Expansion Diagnosis

Hematoma expansion (HE) in hemorrhagic stroke (HS) patients has been showed to be an important indicator for clinical monitoring and prognostic control of stroke patients. However, on the one hand, manual diagnosis exist shortcomings such as inefficiency and inaccuracy. On the other hand, in the data mining of intelligent diagnosis, people usually downsize the data from a single source, through the coarse-grained method of single feature selection, which makes a large amount of useful information not being fully utilized. On top of that, there is a lack of technique to merge the effects of various state-of-the-art models. Consequently, this study develops a novel integrated learning HE prediction system based on multi-source data that automates the determination of whether a HS patient would have the risk of HE with higher precision and interpretability. The key contribution of this paper is to accurately, efficiently and robustly predict the HE by multi-stage feature engineering with heterogeneous integrated features as inputs for stacking integrated learning. In addition, the unbalanced dataset is optimized by employing the Synthetic Minority Over-sampling Technique (SMOTE) algorithm. The model constructed in this paper based on the ten selected features performs well in predicting HE, with 89.74% accuracy and 95.38% AUC on the test set.