A machine learning-based risk prediction model for Hospitalized patients with deep vein thrombosis

Area of Science:

• Medical Informatics
• Clinical Prediction Modeling
• Thrombosis Research

Background:

• Deep vein thrombosis (DVT) is a prevalent condition associated with significant morbidity if not detected early.
• Existing machine learning (ML) models for early DVT risk prediction are limited in clinical applicability.
• There is a need for robust, clinically viable ML tools for timely DVT identification.

Purpose of the Study:

• To develop and validate an ML model for early DVT risk prediction using accessible clinical and laboratory data.
• To identify key predictors of DVT through model explainability techniques.
• To enhance early detection and management of patients at risk for DVT.

Main Methods:

• Retrospective analysis of 231 patients' clinical data (January 2017 - June 2024).
• Feature selection using LASSO regression, identifying seven predictors including D-dimer, hemoglobin, and platelet count.
• Training and evaluation of five ML algorithms (XGBoost, CatBoost, Random Forest, Logistic Regression, SVM) using 70/30 train/test split and 5-fold cross-validation.
• Model performance assessed via AUC, accuracy, recall, and F1 score, with interpretation using SHAP analysis.

Main Results:

• LASSO identified seven significant predictors: hemoglobin, platelet count, leukocyte count, fibrinogen, prothrombin time, D-dimer (DD), and glucose.
• The Random Forest (RF) model demonstrated superior performance with a test-set AUC of 0.874.
• SHAP analysis confirmed D-dimer as the most influential predictor, elucidating the impact of other variables.

Conclusions:

• An internally validated ML model using seven routine variables was developed for early DVT risk prediction.
• The Random Forest model achieved the best performance, highlighting D-dimer as the primary risk indicator.
• This model holds potential for earlier DVT identification and intervention, pending external validation.