Machine learning-based predictive model for post-stroke dementia.

Area of Science:

• Neurology
• Artificial Intelligence
• Medical Informatics

Background:

• Post-stroke dementia (PSD) is a frequent complication impacting stroke patient recovery and prognosis.
• Existing methods often overlook the complex interactions between demographic, comorbidity, and clinical factors in PSD development.
• Predicting PSD is crucial for improving patient outcomes and rehabilitation effectiveness.

Purpose of the Study:

• To investigate the effectiveness of machine learning (ML) approaches for predicting post-stroke dementia (PSD).
• To identify key predictive factors for PSD by analyzing interactions among various patient characteristics.

Main Methods:

• Feature selection was performed using Spearman correlation analysis and the Boruta algorithm, identifying 9 key features.
• Eight distinct machine learning models were developed and assessed, including logistic regression, elastic net, k-nearest neighbors, decision tree, extreme gradient boosting, support vector machine, random forest, and multilayer perceptron.

Main Results:

• The study included 539 stroke patients.
• Extreme gradient boosting and random forest models achieved the highest predictive performance, with AUC values of 0.7287 and 0.7285, respectively.
• Significant predictors for PSD included patient age, elevated high-sensitivity C-reactive protein levels, stroke laterality and location, and history of cerebral hemorrhage.

Conclusions:

• Machine learning models demonstrate strong potential for predicting the risk of post-stroke dementia.
• Extreme gradient boosting is particularly effective, offering a robust tool for early PSD risk assessment.