Machine learning models for predicting 48-hour mortality in acute intracerebral hemorrhagic stroke

BACKGROUNDIdentifying patients with intracerebral hemorrhagic (ICH) at high risk of mortality is crucial for timely intervention. Machine learning (ML) offers novel methodologies for precise predictive models for ICH. Therefore, the aim of this study was to develop an ML-based predictive model for 48-hour mortality in patients with acute hemorrhagic stroke. METHODSA cross-sectional study was conducted using secondary data from 657 patients diagnosed with acute ICH. Demographic, clinical, laboratory, and radiological variables were extracted from medical records. Data preprocessing included cleaning, normalization, and class balancing using the Synthetic Minority Oversampling Technique (SMOTE). Three supervised algorithms—Random Forest, Decision Tree, and Gaussian Naïve Bayes—were developed and evaluated using stratified 5-fold cross-validation. Model performance was assessed using accuracy, sensitivity, specificity, precision, recall, F1-score, and AUC. RESULTSRandom Forest achieved the best overall performance for predicting 48-hour mortality, with an accuracy of 84.77%, F1-score of 84.63%, and AUC of 80.51, outperforming Decision Tree (AUC 61.12) and Gaussian Naïve Bayes (AUC 82.94). Random Forest most accurately identified >48-hour survival, with high sensitivity (93.5%) and PPV (92.9%), while Naïve Bayes provided the most reliable positive classification for this category (PPV 99.0; specificity 94.2%). For ≤24-hour mortality, Naïve Bayes showed the best detection performance (sensitivity 85.4%; NPV 98.7%). CONCLUSIONSMachine learning, particularly the Random Forest algorithm, enables reliable prediction of 48-hour mortality in patients with acute ICH using basic clinical and radiological data available at admission. The model offers practical potential for early risk stratification in emergency and critical care settings.