This project aims to develop and validate an interpretable machine learning model for predicting incident stroke, including overall, ischemic, hemorrhagic, and perioperative subtypes, by integrating traditional clinical risk factors, NMR metabolomics, and Olink proteomics data from the UK Biobank. The research questions are: Can a multi-omics model outperform traditional risk models in predicting overall and subtype-specific stroke? What are the distinct molecular signatures for different stroke subtypes, particularly perioperative stroke? How can these features be translated into a clinically useful risk score?
Objectives include: 1) Developing a multi-omics prediction model for overall and subtype-specific stroke (ischemic, hemorrhagic, perioperative) and comparing its performance with traditional and single-omics models; 2) Identifying key predictive features for each subtype using SHAP analysis; 3) Constructing a Stroke Multi-Omics Risk Score (SMORS) for comprehensive risk stratification; 4) Evaluating the absolute and relative risks across stratified groups.
The scientific rationale is that current stroke risk models lack precision and do not fully utilize emerging molecular biomarkers. Integrating multi-omics data can capture complex biological processes, improve prediction accuracy, and provide insights into the distinct pathophysiology of stroke subtypes-especially perioperative stroke, which involves unique mechanisms like inflammatory and hemodynamic stress-enabling personalized prevention and perioperative management strategies.
