Objectives

Hemorrhagic transformation (HT) following acute ischemic stroke (AIS) critically affects patient prognosis. This study aims to develop reliable machine learning (ML) models for predicting HT in AIS patients using magnetic resonance imaging (MRI).

Materials and Methods

543 AIS patients were enrolled (HT: 166; non-HT: 377) and randomly split into the training and testing set with a ratio of 80% to 20%. All underwent MRI, including diffusion- (DWI) and perfusion-weighted imaging (PWI), within 24 hours of symptom onset, and had follow-up CT or MRI within 14 days. Clinical baseline features and MRI parameters (volumes of ADC < 620 × 10⁻⁶ mm²/s, various Tmax thresholds, PWI-DWI mismatch area, and hypoperfusion intensity ratio [HIR]) were collected. Six ML algorithms—Logistic Regression, Support Vector Machine, Decision Tree, XGBoost, LightGBM, and Random Forest—were trained and optimized. Various fusion models were created, including soft voting, stacking, and integrating subgroup-based modeling into robust baseline models. The best-performing model was selected based on testing set performance.

Results

The Random Forest-based model performed best with a relatively high area under the curve (AUC) of 0.933 (95% CI: 0.883-0.973). The proposed subgroup-driven fusion model, by utilizing Decision Tree as stratification mechanism and subgroup-tailored baseline models (Decision Tree, LightGBM, Random Forest), achieved the highest AUC of 0.951 (95% CI: 0.906–0.988) with an accuracy of 93.6% (95% CI: 89.0-98.2) among all fusion models and exceeded all baseline models. This model is accessible at https://yike-wood.github.io/HT-Predict/.

Conclusion

The proposed subgroup-driven model, leveraging MRI-DWI and PWI, reliably predicts HT in AIS patients, potentially aiding radiologists and clinicians in timely HT risk assessment post-MRI.

Clinical relevance statement

This study describes the application of machine learning to predict hemorrhagic transformation after stroke from 543 patients based on their MRI parameters and baseline features. A proposed subgroup-driven fusion model has been developed with good performance and is available for clinical use with public access at https://yike-wood.github.io/HT-Predict/.