A LightGBM machine learning model incorporating five readily available coagulation and inflammatory markers (D-dimer, CRP, APTT, INR, and FDP) accurately predicts in-hospital VTE in sepsis patients with T2DM, with robust discrimination, calibration, and clinical utility across internal and external validation cohorts.
Key Findings
Results
Among 648 sepsis patients with T2DM, 17.7% developed in-hospital venous thromboembolism.
115 out of 648 sepsis patients with T2DM admitted to Shaanxi Provincial People's Hospital between January 2020 and October 2024 developed VTE.
The study used a retrospective cohort design.
An internal validation cohort of 303 patients and an independent external validation cohort of 260 patients from a separate tertiary care center (Xi'an No. 3 Hospital) were also included.
The external validation cohort was collected between November 2023 and December 2025.
Results
The Boruta algorithm combined with forward sequential feature selection identified five key predictors of VTE in this population.
The five identified predictors were: D-dimer, C-reactive protein (CRP), activated partial thromboplastin time (APTT), international normalized ratio (INR), and fibrin degradation products (FDP).
Feature selection combined the Boruta algorithm with forward sequential feature selection.
All five predictors are described as 'readily available coagulation and inflammatory markers.'
Six machine learning algorithms were evaluated before feature selection: Logistic Regression, SVM, MLP, LightGBM, XGBoost, and Random Forest.
Results
LightGBM achieved the best overall predictive performance among six machine learning algorithms evaluated.
LightGBM achieved an AUC of 0.9025 in the test set.
AUC was 0.8048 in the internal validation cohort.
AUC was 0.8902 in the independent external validation cohort.
The six algorithms compared were Logistic Regression, SVM, MLP, LightGBM, XGBoost, and Random Forest.
Model performance was evaluated by discrimination (AUC), calibration (calibration plots and Brier scores), and clinical utility (decision curve analysis).
Results
Calibration curves demonstrated good agreement between predicted and observed VTE outcomes for the LightGBM model.
Calibration was assessed using calibration plots and Brier scores.
Good agreement between predicted and observed outcomes was demonstrated across cohorts.
Decision curve analysis confirmed positive net clinical benefit across a wide range of threshold probabilities.
These results indicate both statistical and clinical utility of the model.
Results
SHAP analysis identified elevated D-dimer, CRP, INR, and FDP, together with shortened APTT, as the strongest drivers of VTE risk.
SHAP (SHapley Additive exPlanations) analysis was used to enhance model interpretability.
Elevated D-dimer, CRP, INR, and FDP were associated with increased VTE risk.
Shortened APTT (rather than elevated) was identified as a risk driver, consistent with a procoagulant state.
SHAP-based interpretability was described as enhancing transparency and potentially facilitating personalized risk stratification and informed thromboprophylaxis decisions.
Results
Sensitivity analyses confirmed the robustness of the LightGBM model across multiple analytical scenarios.
Three sensitivity analyses were performed: excluding D-dimer-driven cases, removing the D-dimer feature entirely, and restricting the analysis to symptomatic VTE only.
AUC range across all sensitivity analyses was 0.84–0.89.
Results were described as 'consistent' across all sensitivity analyses.
These analyses were conducted to address potential concerns about circularity or ascertainment bias related to D-dimer.
What This Means
This research suggests that machine learning can accurately predict which hospitalized patients with both sepsis and type 2 diabetes are at high risk of developing dangerous blood clots (venous thromboembolism, or VTE). The researchers studied 648 patients and found that about 1 in 6 (17.7%) developed VTE during their hospital stay — a serious complication that can cause pulmonary embolism or deep vein thrombosis. By testing six different machine learning approaches, they identified LightGBM as the best-performing model, which used just five standard blood test results — D-dimer, C-reactive protein (CRP), APTT, INR, and fibrin degradation products (FDP) — to predict VTE risk with high accuracy (AUC ~0.90 in the test set, and 0.80–0.89 in separate validation groups).
An important feature of this study is its use of SHAP analysis, a technique that explains why the model makes a particular prediction for each patient, rather than acting as a 'black box.' This transparency is clinically valuable because it shows that higher levels of D-dimer, CRP, INR, and FDP — along with a shorter APTT (indicating a tendency toward clotting) — were the main factors driving high VTE risk predictions. The model was also tested on an entirely separate hospital's patient population, and it continued to perform well, suggesting it could generalize beyond the institution where it was developed.
This research suggests that combining routine coagulation and inflammation blood tests with interpretable machine learning could help clinicians identify sepsis patients with diabetes who are most at risk for blood clots early in their hospital stay. Such a tool could potentially support more personalized decisions about when to use blood-thinning medications (thromboprophylaxis) in this high-risk patient group. The study was retrospective and conducted at hospitals in one region of China, so further prospective studies in diverse populations would be needed before broader clinical adoption.
Ma L, Yu X, Xin Q. (2026). Interpretable machine learning models to predict venous thromboembolism in sepsis patients with type 2 diabetes mellitus.. Frontiers in endocrinology. https://doi.org/10.3389/fendo.2026.1958305