Interpretable machine learning and mendelian randomization identify risk factors for lower extremity arterial embolism and thrombosis.

BACKGROUND

Lower extremity arterial embolism and thrombosis lead to significant morbidity, but their risk factors are not fully characterized.

OBJECTIVES

To identify risk factors and develop an interpretable machine learning (ML) model for predicting lower extremity arterial embolism and thrombosis, with validation using mendelian randomization (MR).

METHODS

In this retrospective case-control study, data were collected from patients with lower extremity arterial embolism and thrombosis treated at our department of vascular surgery between January 2018 and November 2025. Predictors were selected using LASSO regression, and 11 ML models were developed using the selected variables. The optimal model was interpreted and implemented as a web-based prediction tool. Clinical utility and model calibration were assessed using decision curve analysis and calibration curves. Key predictors were further assessed using MR and multivariable logistic regression.

RESULTS

XGBoost achieved the highest discrimination, with an AUC of 0.951 (95% CI 0.925-0.972) in the test set. MR analyses indicated that genetically predicted cerebrovascular disease (CD) (OR 1.773; 95% CI 1.043-3.015; P = 0.035) and gamma-glutamyl transferase (GGT) (OR 1.327; 95% CI 1.105-1.595; P = 0.003) were risk factors, whereas mean platelet volume (MPV) was protective (OR 0.879; 95% CI 0.778-0.993; P = 0.038). Multivariable logistic regression confirmed CD (OR 10.19; 95% CI 4.36-23.82; P < 0.001) and MPV (OR 0.63; 95% CI 0.52-0.76; P < 0.001) as independent predictors.

CONCLUSIONS

Interpretable ML combined with MR identified a history of CD and lower MPV as factors associated with risk of lower extremity arterial embolism and thrombosis.