Neuro-Behavioral Convergence: A Multimodal Approach to Stroke Risk Assessment

Overview

Stroke is a critical global health issue, and many cases are preventable if risk factors are identified early. Current risk assessment approaches often separate behavioral and neurophysiological analyses, and physiological screening is expensive and less accessible. In this project, I performed a multimodal retrospective analysis integrating public health and electroencephalography (EEG) data to identify stroke-associated factors. I then translated these findings into a risk-awareness tool designed to expand prevention awareness beyond traditional clinical settings, particularly benefiting underserved populations. First, I used an interpretable machine learning (ML) approach to analyze high-dimensional CDC Behavioral Risk Factor Surveillance System (BRFSS) data using LASSO-regularized logistic regression to identify key risk factors while reducing overfitting. Second, I analyzed a clinical EEG dataset using Elastic Net regression to test whether brainwave features alone contained stroke-related signals after removing demographic and clinical confounders. Third, I developed a cross-platform application that displays personalized stroke risk and allows users to explore how lifestyle changes may affect their estimated risk. The BRFSS model demonstrated strong discrimination (AUC = 0.8052; sensitivity = 77.61%) with diabetes, smoking, general health, and mobility as the strongest contributors. The EEG ensemble classifier achieved an AUC of 0.8039 and 75.86% accuracy, indicating that neurophysiological patterns contain stroke-related information independent of behavioral data. Convergent performance across datasets suggests that behavioral and physiological signals capture complementary aspects of risk. The system was designed with a scalable, model-agnostic architecture. It supports future clinical risk assessment applications once larger clinical datasets become available.