Public Health.

Area of Science:

• Epidemiology
• Environmental Health
• Machine Learning in Medicine

Background:

• Stroke is a primary cause of vascular dementia.
• Traditional stroke prediction models focus on individual risk factors.
• Environmental exposures may impact cerebrovascular health, necessitating novel prediction approaches.

Purpose of the Study:

• To develop and compare machine learning models for stroke risk prediction.
• To evaluate the predictive performance of an environmental risk model (Env-RiskModel) against a standard risk factor model (Standard-RiskModel).

Main Methods:

• Utilized data from the REasons for Geographic and Racial Differences in Stroke (REGARDS) study (n=24,251).
• Developed five random survival forest models incorporating demographics, neighborhood characteristics (greenspace, social vulnerability, disaster risk), and ambient air pollution (Ozone).
• Compared model performance using C-index and time-dependent Area Under the Receiver Operating Characteristic (AUROC) curves.

Main Results:

• Over 12.1 years, 6.0% experienced stroke.
• Models incorporating environmental factors showed marginal improvements over demographics alone.
• The Env-RiskModel (C-index=66.1%) and Standard-RiskModel (C-index=65.7%) demonstrated comparable predictive performance.

Conclusions:

• Environmental factors did not significantly enhance stroke prediction beyond established individual risk factors.
• Environmental exposures combined with demographics offer a viable alternative for stroke prediction when clinical data is scarce.
• Further research is warranted to explore additional environmental variables for stroke risk assessment.