Predicting ischemic stroke patients' prognosis changes using machine learning in a nationwide stroke registry.

Area of Science:

• Neurology
• Medical Informatics

Background:

• Accurate prognosis prediction for ischemic stroke patients post-discharge is vital for effective long-term care planning.
• Previous machine learning (ML) models showed promise but struggled to identify specific clinical features influencing prognosis changes.

Purpose of the Study:

• To assess and compare different prediction models for estimating stroke patient prognosis changes over time.
• To identify key clinical factors associated with prognosis changes for improved patient recovery plans.

Main Methods:

• Utilized a large national stroke registry database.
• Compared three prediction models: logistic regression, clinical scores, and XGBoost (ML).
• Evaluated model performance in predicting 3-month prognosis changes.

Main Results:

• The XGBoost model significantly outperformed logistic regression and clinical scores, achieving an AUROC of 0.929.
• XGBoost maintained high precision using only the 20 most relevant clinical features.
• Identified specific clinical features strongly correlating with prognosis changes.

Conclusions:

• XGBoost is a superior model for predicting ischemic stroke prognosis changes.
• Selected clinical features effectively predict post-discharge prognosis, aiding physician decision-making.
• This approach facilitates optimized patient recovery strategies.