Developing and validating a machine learning model for predicting post-thrombolysis seizures in acute ischemic...

Area of Science:

• Neurology
• Medical Informatics
• Biostatistics

Background:

• Post-stroke seizures (PSS) are a common complication of ischemic brain injury, but risk factors remain poorly understood.
• Predicting PSS is challenging due to variable manifestation and complex underlying mechanisms.

Purpose of the Study:

• To develop and validate a machine learning (ML) model for predicting PSS risk in acute ischemic stroke (AIS) patients receiving thrombolysis.
• To identify key clinical and laboratory predictors of PSS to improve patient risk stratification and management.

Main Methods:

• Retrospective analysis of 332 AIS patients treated with thrombolysis, utilizing 21 clinical and laboratory variables.
• Development of seven ML models, including Random Forest (RF), with feature selection via expert consensus and Boruta algorithm.
• Performance evaluation using AUC, Brier score, accuracy, sensitivity, specificity, and SHAP analysis for feature interpretability.

Main Results:

• The Random Forest model demonstrated optimal performance with an AUC of 0.867.
• Key predictors identified were fasting blood glucose, serum sodium, serum calcium, and age.
• Lower serum electrolytes, elevated glucose, and younger age were associated with increased PSS risk.

Conclusions:

• The developed RF-based ML model effectively stratifies PSS risk in AIS patients treated with thrombolysis using accessible clinical data.
• SHAP analysis highlights fasting glucose, serum sodium/calcium, and age as crucial predictors, providing actionable insights for personalized care.
• The model, deployed as a web tool, can aid in early intervention strategies to reduce the burden of PSS.