Development of Venous Thromboembolism Risk Prediction Models Based on Whole Blood Gene Expression Profiling Using 20 Machine Learning Algorithms: Comprehensive Analysis Study

TL;DR

This study developed VTE risk prediction models using blood gene expression data and 20 machine learning algorithms, with nine models showing strong diagnostic performance in validation.

Contribution

The novel use of whole blood gene expression profiling with 20 ML algorithms to build and validate VTE prediction models.

Findings

Nine machine learning models achieved an area under the curve greater than 0.75 in external validation.

Most models maintained high specificity in external validation cohorts.

Combining these models with D-dimer could improve VTE diagnosis.

Abstract

There is a lack of venous thromboembolism (VTE) risk prediction models based on gene expression information.

This study aimed to construct a VTE prediction model based on whole blood gene expression profiling, by performing a comprehensive analysis of 20 machine learning (ML) algorithms.

Two transcriptome datasets containing patients with VTE and healthy controls were obtained by searching the Gene Expression Omnibus database and used as the training and validation sets, respectively. Feature selection for model construction was performed on the training set using the least absolute shrinkage and selection operator and random forest, followed by the selection of the intersection of the chosen features. Subsequently, recursive feature elimination was applied to further refine the selected features. The selected features underwent model construction using 20 ML algorithms. The performance of the models was evaluated using various methods such as receiver operating characteristic and confusion matrix. The validation set was used for external model validation.

The final results demonstrated that all algorithm models, except for k-nearest neighbor, exhibited good performance in VTE prediction. External validation data indicated that 9 algorithm models had an area under the curve greater than 0.75. The confusion matrix analysis revealed that the algorithm models maintained high specificity in the external validation cohort.

This study used 20 ML algorithms to construct VTE prediction models based on whole blood gene expression information, with 9 of these models demonstrating good diagnostic performance in external validation cohorts. The above models, when used in conjunction with D-dimer, may provide more valuable references for VTE diagnosis.

Figures