Dynamic, Explainable Artificial Intelligence Framework for Sequential Prediction of Symptomatic Intracranial...

Symptomatic intracranial hemorrhage (sICH) remains a devastating complication undermining reperfusion therapy in acute ischemic stroke (AIS). Current prediction models rely on static baseline parameters, neglecting dynamic post-treatment physiological changes. We aimed to develop and validate a dynamic, explainable risk prediction framework integrating high-resolution temporal data.

In this prospective multicenter study, 2,341 AIS patients with anterior circulation large vessel occlusion receiving reperfusion therapy were consecutively enrolled (January 2023-June 2024). High-frequency vital signs and serial laboratory tests were collected over 72 h post-treatment. Time-series features (trend, volatility, stability) were engineered, and XGBoost-based dynamic models were constructed at five timepoints (6, 12, 24, 48, 72 h). Performance was evaluated in a held-out validation cohort (n = 576) against a static baseline model. SHAP was applied for interpretability.

Of 2,158 analyzed patients, 218 (10.1%) developed sICH. The dynamic model consistently outperformed the static model across all timepoints (AUC 0.83-0.90 vs. 0.71-0.75; all p < 0.001). SHAP identified three core risk drivers: hemodynamic volatility (24-h SBP standard deviation ranked highest), metabolic-inflammatory trajectories, and early tissue injury markers. Unsupervised clustering revealed three risk phenotypes: persistent low-risk (71.5%, sICH 0.8%), early rapid-riser (18.2%, 52.3%), and delayed-riser (10.3%, 38.1%).

This dynamic, explainable framework shifts sICH risk prediction from static snapshots to continuous, interpretable monitoring. Hemodynamic volatility emerged as the strongest modifiable predictor, offering clinically actionable insights for personalized post-procedural care in AIS.