Development of Venous Thromboembolism Risk Prediction Models Based on Whole Blood Gene Expression Profiling Using 20...

Area of Science:

• Genomics
• Bioinformatics
• Machine Learning

Background:

• Venous thromboembolism (VTE) lacks robust risk prediction models utilizing gene expression data.
• Current diagnostic methods for VTE require improvement in predictive accuracy and efficiency.

Purpose of the Study:

• To develop and validate a predictive model for VTE risk using whole blood gene expression profiling.
• To comprehensively analyze the performance of 20 distinct machine learning algorithms for VTE prediction.

Main Methods:

• Utilized two transcriptome datasets from the Gene Expression Omnibus (GEO) database for training and validation.
• Employed feature selection techniques including LASSO, Random Forest, and Recursive Feature Elimination.
• Constructed and evaluated 20 machine learning models, assessing performance via ROC curves and confusion matrices.

Main Results:

• Nine machine learning models achieved an Area Under the Curve (AUC) greater than 0.75 in external validation.
• All models, except k-nearest neighbor, demonstrated good performance in VTE prediction.
• High specificity was maintained across algorithm models in the external validation cohort.

Conclusions:

• Developed 9 machine learning models with significant diagnostic performance for VTE prediction based on gene expression.
• These gene expression-based models, potentially combined with D-dimer, offer valuable new tools for VTE diagnosis.