Abstract 056: Machine Learning for Predicting Hemorrhagic Transformation in Ischemic Stroke: A Systematic Review

Introduction Hemorrhagic transformation (HT) is a major complication following ischemic stroke, especially in patients treated with reperfusion therapies. Accurate prediction of HT is essential to optimize clinical decisions, reduce iatrogenic risk, and guide individualized therapy. Machine learning (ML) models offer a promising avenue to integrate clinical, imaging, and biochemical variables for HT risk stratification. This systematic review evaluates the performance, methodologies, and clinical utility of ML algorithms developed to predict HT in acute ischemic stroke. Methods We conducted a comprehensive search of MEDLINE, Embase, IEEE Xplore, and Scopus from inception to May 2025, following PRISMA guidelines. Eligible studies developed or validated ML models for HT prediction post‐ischemic stroke using clinical or imaging features. Data extraction included algorithm type, feature categories, dataset size, validation approach, performance metrics (AUC, sensitivity, specificity), and model transparency. Meta‐analytical synthesis of area under the curve (AUC) values was conducted using R (metafor package), and performance visualization was generated in Python. Results Twenty‐four studies involving over 18,000 patients were included. Most studies were retrospective (n=21), and only 4 applied external validation. Imaging (NCCT, MRI DWI/FLAIR), clinical variables (NIHSS, glucose, atrial fibrillation), and treatment status (IV tPA, thrombectomy) were common features. Model types included logistic regression (n=6), random forest (n=5), XGBoost (n=4), convolutional neural networks (CNNs; n=3), and hybrid models integrating imaging and clinical data (n=6). The pooled AUCs ranged from 0.78 for logistic regression to 0.91 for hybrid CNN+clinical models. CNN‐based models using DWI/PWI data outperformed those using only NCCT. XGBoost consistently showed high sensitivity (mean: 0.83) but had lower interpretability. Only 3 studies conducted SHAP or LIME‐based explainability assessments. Among studies with external validation, performance decreased by ∼0.04 AUC on average. Calibration and decision‐curve analyses were rarely reported (n=2), limiting real‐world applicability. Conclusion ML models demonstrate promising accuracy in predicting HT following ischemic stroke, especially when integrating multimodal imaging with clinical parameters. Hybrid architectures and CNN‐based approaches show the strongest predictive performance. However, most models lack external validation and interpretability assessments, hindering clinical adoption. Future research must emphasize generalizability through multicenter prospective validation, explainability to support clinician trust, and standardized reporting per TRIPOD‐AI and MI‐CLAIM frameworks. ML‐based HT prediction can enhance patient selection for reperfusion and anticoagulation strategies, but its true clinical utility hinges on responsible and transparent deployment. image