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

Area of Science:

• Neurology
• Medical Informatics
• Machine Learning

Background:

• Accurate prediction of subarachnoid hemorrhage (SAH) severity is vital for clinical decision-making and patient management.
• Imbalanced data presents a significant challenge in classifying SAH severity using the Modified Rankin Scale (MRS).

Purpose of the Study:

• To develop and evaluate a multi-stage machine learning framework for improved SAH severity prediction.
• To address data imbalance issues in SAH classification using a three-stage approach.

Main Methods:

• A three-stage classification framework was implemented, starting with binary classification (Good vs. Poor Outcome).
• Feature selection was performed using Random Forest, identifying the top 20 predictive features.
• Thirteen machine learning models were evaluated at each stage, with top performers selected for optimization. The dataset included 535 samples across seven MRS levels, validated with 5-fold cross-validation.

Main Results:

• The initial binary classification stage achieved approximately 90% accuracy using the Extra Trees model.
• The Random Forest model demonstrated 88% accuracy in classifying the 'Good Outcome' group in the second stage.
• The Random Forest model achieved 86% accuracy for the 'Poor Outcome' group in the third stage.

Conclusions:

• The proposed multi-stage classification technique offers a promising solution for predicting SAH severity, particularly for imbalanced datasets.
• This approach enhances prediction accuracy and holds potential for practical application in clinical settings.
• Further research is recommended for model tuning to optimize efficacy in real-world healthcare environments.