Multi-class subarachnoid hemorrhage severity prediction: addressing challenges in predicting rare outcomes.

Abstract

Accurately predicting the severity of subarachnoid hemorrhage (SAH) is critical for informing clinical decisions and improving patient outcomes. This study addresses the challenges of imbalanced data in SAH severity classification by employing the Modified Rankin Scale (MRS) within a three-stage classification framework. We utilize a three-stage approach to effectively categorize SAH severity. In the first stage, we performed binary classification, grouping SAH severity into "Good Outcome" (class 0), which includes MRS levels 0, 1, 2, and 3, and "Poor Outcome" (class 1), encompassing levels 4, 5, and 6. Feature selection was done using a Random Forest algorithm to identify the top 20 features for the SAH severity prediction. We evaluated thirteen machine learning models at each stage, selecting the top-performing classifiers to optimize results. The dataset comprised 535 samples across seven MRS severity levels and was validated using 5-fold cross-validation and diverse subgroups to ensure robust model performance across various scenarios. Binary classification in the first stage achieved approximately 90% accuracy with Extra Trees. In the second stage, targeting the "Good Outcome" group, the Random Forest model reached 88% accuracy, while in the third stage, it achieved 86% accuracy for the "Poor Outcome" group. By increasing accuracy across unbalanced classes and emphasizing its potential for practical use, the multi-stage technique presents a promising solution for predicting the severity of SAH. Future research will concentrate on additional tuning to improve the model's efficacy in actual healthcare environments.