Bridging Interpretability and Performance: Enhanced Machine Learning-Based Prediction of Hematoma Expansion Post-Stroke via Comprehensive Feature Selection

Abstract

Monitoring and controlling the occurrence of hematoma expansion events after a stroke is a primary clinical focus. The introduction of machine learning (ML) techniques offers intelligent decision support for physicians in this domain. However, for doctors without an ML background, the behavior of a hematoma expansion predictor seems opaque, similar to a 'black box.' Moreover, the vast and diverse set of features typically present in medical data acts as a double-edged sword: while encapsulating rich information with potential value, it also includes redundant details that offer little to predictive utility. Comprehensive feature selection is crucial, but many current state-of-the-art hematoma expansion prediction studies based on ML often overlook this step. In this paper, we propose a methodology tailored for comprehensive feature selection across diverse and abundant medical data features and rigorously evaluate ML models. Through experiments on a real-world post-stroke hematoma expansion prediction dataset, we demonstrate the efficacy of our approach in enhancing the performance of ML predictors. Visualization of the associated feature selection process and results further bolsters physicians' understanding of the model's decision-making basis, thereby strengthening its interpretability.