Machine learning-based prediction of in-hospital deep vein thrombosis in patients with acute ischemic stroke: a multicenter study.

Among 21,459 patients with acute ischemic stroke, 1,324 (6.17%) developed in-hospital deep vein thrombosis.

• Data were drawn from a publicly available multicenter retrospective dataset
• The primary outcome was imaging-confirmed in-hospital DVT
• Participants were randomly divided into a training set (70%) and a held-out test set (30%)
• DVT prevalence of 6.17% reflects a class imbalance addressed in part by SMOTE sensitivity analyses

In the full 19-variable predictor set analysis, RANGER achieved the highest AUC of 0.976 in the held-out test set.

• Eight machine learning algorithms were trained on both a 19-variable full predictor set and an 8-variable simplified predictor set
• Hyperparameters were optimized using repeated 5-fold cross-validation with 2 repeats
• Model performance was evaluated using AUC, AUPRC, Brier score, calibration, decision curve analysis, and additional classification metrics
• RANGER refers to a random forest-based algorithm

Among models using the simplified 8-variable predictor set, XGBoost achieved the highest AUC (0.917) and sensitivity (0.852).

• The simplified predictor set contained 8 variables selected via LASSO regression, the Boruta algorithm, variance inflation factor assessment, and clinical judgment
• XGBoost prioritized sensitivity at 0.852, making it favorable for minimizing missed DVT cases
• All 8 simplified predictor-set models were incorporated into an online prediction platform
• The simplified set was designed to facilitate practical clinical implementation

SVM demonstrated the most favorable overall performance profile among simplified predictor-set models, with the highest AUPRC (0.605), lowest Brier score (0.038), highest positive predictive value (0.440), and highest F1-score (0.549).

• AUPRC of 0.605 is particularly informative given the class imbalance (6.17% DVT prevalence)
• A Brier score of 0.038 indicates good calibration relative to the other models
• Positive predictive value of 0.440 and F1-score of 0.549 reflect a balance between precision and recall
• SVM was identified as showing 'the most favorable overall performance profile' among simplified predictor-set models

D-dimer was the most influential predictor of in-hospital DVT, and model performance was attenuated after its exclusion.

• Sensitivity analyses specifically excluded D-dimer to assess model robustness
• Model performance was described as 'attenuated after exclusion of D-dimer'
• D-dimer was identified as 'the most influential predictor' across models
• This finding suggests D-dimer contributes substantially to discriminative ability but also raises questions about availability in all clinical settings

SMOTE (synthetic minority oversampling technique) analysis applied within folds showed slightly lower discrimination and greater calibration discrepancies compared to the primary analysis.

• SMOTE was applied within-fold to address class imbalance as a sensitivity analysis
• Results showed 'slightly lower discrimination and greater calibration discrepancies'
• This suggests oversampling did not improve and may have slightly worsened model performance in this dataset
• The primary analysis without SMOTE was therefore considered the main model set

Feature selection employed a multi-method approach combining LASSO regression, the Boruta algorithm, variance inflation factor assessment, and clinical judgment to derive the simplified 8-variable predictor set from a larger candidate pool.

• Least absolute shrinkage and selection operator (LASSO) regression was used for penalized variable selection
• The Boruta algorithm is a wrapper method that identifies all relevant features
• Variance inflation factor assessment was used to address multicollinearity
• Clinical judgment was incorporated as a fourth filter, resulting in a final set of 8 predictors from an initial 19-variable full set

The authors concluded that independent external validation and prospective clinical-impact assessment are required before routine clinical implementation of these models.

• The study was a secondary analysis of a retrospective dataset, limiting causal inference
• The online prediction platform incorporating all 8 simplified predictor-set models was described as a step toward clinical utility
• The multicenter design (21,459 patients) provides a broad sample but internal validation alone was performed
• The authors explicitly stated models are not yet ready for 'routine clinical implementation'