Clinical Report: Utilizing Explainable Machine Learning to Forecast VTE in Septic Shock
Overview
This study developed and validated an explainable machine learning framework to predict venous thromboembolism (VTE) in patients experiencing septic shock. The Random Forest algorithm demonstrated high accuracy, achieving an AUC of 0.9718 in external validation.
Background
Venous thromboembolism (VTE) is a significant cause of morbidity and mortality in critically ill patients, particularly those with septic shock. Current risk assessment tools are often inadequate, highlighting the need for more precise, individualized prediction methods. Machine learning offers a promising approach to enhance risk stratification by analyzing complex clinical data.
Data Highlights
Metric
Development Cohort
External Validation Cohort
VTE Incidence
17.74% (130/733)
N/A
AUC
N/A
0.9718
F1 Score
N/A
0.7917
Sensitivity
N/A
0.7037
Key Findings
The study included 733 patients in the development cohort and 257 in the external validation cohort.
Six predictors were identified: fibrin degradation products (FDP), prothrombin time (PT), white blood cells (WBC), activated partial thromboplastin time (APTT), D-dimer, and C-reactive protein (CRP).
The Random Forest algorithm outperformed other models in predicting VTE.
SHAP analysis provided insights into the contribution of thrombo-inflammatory markers to VTE risk.
The model aims to enhance clinical intuition and risk assessment for thromboprophylaxis.
Clinical Implications
The developed machine learning model utilizes routine clinical biomarkers for risk assessment.
Conclusion
The study established a predictive model for VTE in septic shock patients. Further validation is necessary to assess its real-world applicability.
