Trajectory of influencing factors for post-thrombolysis early neurological deterioration and long-term functional recovery: insights from parallel explainable machine learning models.
Authors
Affiliations (4)
Affiliations (4)
- Department of Neurology, The First Affiliated Hospital of Soochow University, Suzhou 215006, China.
- Department of Public Health and Medicinal Administration, Faculty of Medicine, University of Macau, E12 Building, Avenida da Universidade, Taipa, Macau SAR 999078, China.
- Department of Neurology, The First Affiliated Hospital of Soochow University, Suzhou 215006, China. Electronic address: [email protected].
- Department of Neurology, The First Affiliated Hospital of Soochow University, Suzhou 215006, China. Electronic address: [email protected].
Abstract
Ischemic brain injury after intravenous thrombolysis (IVT) produces highly variable early deterioration and later disability, yet most prognostic models treat risk as a static, admission-only construct. This leaves the temporal reorganization of influencing factors poorly defined. In 609 consecutive patients treated with IVT alone within 4.5 h of last known well, we built three parallel explainable machine-learning models for 24-hour early neurological deterioration (END), poor function at 1 week, and poor function at 3 months. Stage-specific predictor sets followed the information actually available at each decision point, while an identical analytical pipeline (multi-method feature selection, Bayesian optimization, and synthetic minority oversampling) kept the comparisons methodologically aligned. Cross-validated areas under the receiver operating characteristic curve were 0.742, 0.831 and 0.914. SHapley Additive exPlanations (SHAP) analysis showed a coherent trajectory after ischemic injury: metabolic load and early infarct burden (Alberta Stroke Program Early CT Score) dominated the hyperacute window; baseline National Institutes of Health Stroke Scale score and intracranial hemorrhage rose in the subacute window; discharge disposition, age and large-vessel occlusion dominated chronic outcome. A glucose × stroke-severity interaction further indicated that metabolic stress is not uniform across severity strata. These results provide a computational description of how clinically measurable factors re-weight over time after reperfusion, and they support phase-specific management of post-ischemic recovery.