Utilizing Machine Learning Techniques to Predict Negative Remodeling in Uncomplicated Type B Intramural Hematoma

Background: To evaluate the effectiveness of machine learning (ML) techniques in predicting negative remodeling in uncomplicated Stanford type B intramural hematoma (IMHB) during the acute phase. Methods: Univariate logistic regression and 8 ML models were used for prediction. The least absolute shrinkage and selection operator (LASSO) method was used to minimize regression coefficients and reduce the likelihood of overfitting. Hyperparameter tuning was achieved using randomized and grid searches. Model performance was evaluated using the area under the curve (AUC), F1 score, and Brier score. Shapley additive explanations (SHAP) values were calculated, allowing us to meticulously rank the significance of the input features and elucidate the outcomes of the prediction models. Results: One hundred fifty-four cases of uncomplicated IMHB diagnosed between January 2020 and January 2023 were included. The LASSO method identified 7 features that were significantly associated with negative remodeling in uncomplicated IMHB as follows: lymphocytes, white blood cells, neutrophil-to-lymphocyte ratio, eosinophils, monocytes, hypertension, and statins. The CatBoost model (which handles categorical features, simplifies data preprocessing, mitigates overfitting, and provides feature importance tools for better model interpretability) outperformed the other models in terms of the AUC (0.969), F1 score (0.94), and Brier score (0.0625). The SHAP method revealed that in the CatBoost model, the top 3 features associated with negative remodeling in uncomplicated IMHB patients were monocyte count (1.250), lymphocyte count (0.296), and eosinophil count (0.249) on admission. Conclusion: The CatBoost model effectively predicts negative remodeling in uncomplicated IMHB patients during the acute phase by integrating clinical features (monocyte, lymphocyte, and eosinophil counts), representing a tool to improve patient outcomes.