Early Warning from Consecutive Blood Pressure and Heart Rate for Predicting Symptomatic Hemorrhagic Transformation Following Endovascular Thrombectomy for Acute Large Vessel Occlusive Ischemic Stroke
Background: Acute ischemic stroke (AIS) with large vessel occlusion (LVO) carries high morbidity despite endovascular thrombectomy (EVT).Symptomatic hemorrhagic transformation (sHT), a reperfusion-related complication, remains a critical concern. Elevated blood pressure variability (BPV) post-EVT correlates with sHT risk, but traditional BPV assessments lack real-time utility. We propose a machine learning model analyzing continuous blood pressure and heart rate to predict sHT early, enabling timely intervention and improved outcomes.Methods: We conducted a retrospective study including 615 patients with AIS who achieved successful recanalization after EVT across two hospitals. Blood pressure (BP) and heart rate (HR) measurements were recorded at 30-minute or 1-hour intervals during the first 24 hours post-EVT, and non-contrast head CT was performed to determine the presence of sHT. Patients were categorized into symptomatic HT (sHT) and non-sHT groups based on CT findings. We applied deep learning models, including LeNet and Generative Adversarial Networks (GANs), to characterize and compare BP/HR temporal patterns between groups. GAN-based augmentation was used to mitigate class imbalance by generating synthetic sHT cases, and a sliding-window strategy was implemented to support real-time prediction of sHT.Results: Using data from two hospitals, the deep learning framework identified distinct pre-sHT temporal signatures in systolic (SBP), diastolic (DBP), and HR, with SBP exhibiting the largest and most informative fluctuations prior to sHT onset. After GAN-based augmentation to address class imbalance, the model achieved strong discrimination for sHT prediction in both the training and held-out test evaluations. In the independent external validation cohort, the model correctly identified all 73 non-sHT cases and 25 of 28 sHT cases, yielding an AUC of 0.978, an accuracy of 97.0%, a specificity of 100%, and a sensitivity of 89.3%. With a 10-point sliding window, the model provided near–real-time forecasting of sHT in the next measurement interval with high accuracy. Notably, the optimal decision threshold was not identical across hospitals, consistent with inter-site differences in baseline haemodynamics, measurement frequency, and case mix. Therefore, the operating threshold was optimized separately for each hospital on the corresponding test/evaluation cohort using ROC analysis and Youden’s J statistic to balance sensitivity and specificity. In the overall test set, the model showed good generalization, achieving high accuracy in the non-sHT group and substantially improved performance for sHT following GAN-based balancing.Conclusion: Our findings show that deep learning models can anticipate symptomatic haemorrhagic transformation after EVT using early post-procedural blood pressure and heart-rate trajectories. In this framework, LeNet was used for prediction, while a GANs was applied separately to mitigate class imbalance through synthetic case generation. The resulting pipeline supports near–real-time risk estimation with high discriminative performance, offering a clinically actionable strategy for earlier identification of high-risk patients and timely intervention, with operating thresholds calibrated to the intended deployment setting.