Hybrid Feature Selection Approach for Stability Prediction for Intracerebral Hemorrhage Patients

Abstract

Intra Cerebral Hemorrhage (ICH) stability estimation is useful for enhancing diagnosis accuracy, selecting the best course of treatment, and clinically evaluating variations with healthy individuals. Due to their low predictive value, the clinical application of several ICH progression scoring systems is constrained. The dataset includes clinical parameters including age, Glasgow Coma Scale (GCS), and CT Angiography (CTA) spot that have been retrieved using geometric features from the segmented bleeding volume to predict ICH stability. To train and compare stability estimation, many cutting-edge machines learning approaches, including Multi-Layer Perceptron (MLP), Support Vector Machine (SVM), Gradient Boosting, and Random Forest, are applied. We identify the shapely values (SHAP) and describe the key components of the ICH risk scoring system in order to harmonize clinical judgement with model learning. 20 of the patients, were judged to be in critical condition. With a precision of 82.9% and accuracy of 78.3%, the stability of the ICH patients was predicted. The mean square error of regression for expansion rate of the hemorrhage was 0.46. According to the SHAP analysis, the most important factors defining the stability of the stroke lesion are the CTA spot sign, age, solidity, position, and length of the initial axis of the ICH volume. An ablation study was conducted to reach the conclusion that by predicting long-term results, the integration of significant geometric elements with clinical features can enhance the ICH progression rating.