Public Health.

Area of Science:

• Neurology
• Data Science
• Gerontology

Background:

• Stroke is a significant risk factor for Alzheimer's disease and related dementias (ADRD), with a tripled risk in the first year post-stroke.
• Limited research exists on predicting post-stroke dementia using machine learning with real-world clinical data.
• Electronic health records (EHR) offer a valuable resource for developing predictive models.

Purpose of the Study:

• To evaluate various predictive modeling approaches for identifying ADRD risk within one year after stroke.
• To leverage electronic health records (EHR) for post-stroke dementia prediction.
• To compare the performance of machine learning algorithms against traditional statistical methods.

Main Methods:

• Extracted EHR data from the TriNetX Network for adult patients experiencing their first stroke in 2018.
• Developed and compared six prediction models: logistic regression (LR-Back, LR-LASSO, LR-Ridge), Random Forest (RF), LightGBM, and XGBoost.
• Incorporated demographic data, vital signs, lab results, comorbidities, and medication history; evaluated performance using accuracy and AUC-ROC.

Main Results:

• 8% of 55,888 stroke patients developed ADRD within one year.
• XGBoost and LightGBM models demonstrated the highest predictive performance, with accuracy exceeding 92% and AUCs around 0.83.
• Demographic factors, comorbidities, and medication use were significant predictors in ADRD cases.

Conclusions:

• Machine learning and traditional statistical models can effectively predict one-year post-stroke dementia risk using EHR data.
• These predictive models can aid in identifying high-risk individuals for targeted monitoring and early intervention.
• Further research can refine these models for clinical application in stroke survivor care.