ICU Heart Failure Patients Need Better VTE Risk Stratification
Critically ill patients with heart failure remain exposed to venous thromboembolism even when early thromboprophylaxis is already in place. The problem is not simply whether prophylaxis is prescribed, but whether bedside teams can identify which patients continue to carry enough residual thrombotic risk to justify closer surveillance. In this cohort, routinely available clinical data were used to build an interpretable prediction framework aimed at separating higher-risk patients from those less likely to develop in-hospital VTE.
The clinical logic is straightforward. Heart failure contributes to venous stasis, systemic inflammation, and hemodynamic instability, while critical illness adds immobility, mechanical ventilation, and frequent organ dysfunction. When those factors overlap in the ICU, prophylaxis lowers risk but does not erase it. A model that can rank patients by residual probability of thrombosis may help clinicians decide where to focus attention, especially when resources for imaging, monitoring, and escalation are limited.
Performance Across Internal and External Validation Cohorts
Investigators assembled nine prediction models from 8,825 patients in MIMIC-IV and validated them externally in 5,425 eICU records. Fifteen predictors emerged after combined least absolute shrinkage and selection operator regression and Boruta feature selection, a two-step process that narrows the variable set while preserving the strongest signals in the data. Among the nine models, XGBoost delivered the best discrimination.
Its ROC AUC reached 0.764 (95% CI, 0.723–0.805) in internal testing and 0.731 (95% CI, 0.676–0.786) in external validation. Those values do not suggest a perfect bedside oracle, but they do indicate measurable separation between patients who went on to experience VTE and those who did not. The external drop in performance is familiar in clinical prediction work and reflects the harder test of transportability across ICUs, documentation styles, and care patterns.
A single model cannot replace clinical judgment.
Still, the magnitude of discrimination is large enough to matter if the model is deployed as a triage aid rather than a stand-alone decision rule. In practice, a tool with this level of performance may be most useful for flagging patients whose risk is high enough to justify heightened vigilance, repeat reassessment, or lower thresholds for diagnostic evaluation when clinical suspicion rises.
A Handful of Predictors Carried Much of the Signal
SHAP analysis identified several leading contributors to model predictions: atrial tachycardia/fibrillation, respiratory failure, previous thromboembolism, antiplatelet therapy, and mechanical ventilation within 24 hours. These features fit the biology of ICU thrombosis risk. Atrial arrhythmia often marks cardiovascular instability and a more complicated hemodynamic course. Respiratory failure and early mechanical ventilation reflect severe acute illness, immobility, and the kind of systemic stress that amplifies thrombotic propensity.
Previous thromboembolism is an especially intuitive predictor because it signals an established tendency toward recurrence. The inclusion of antiplatelet therapy is also clinically interesting, not because antiplatelet agents substitute for anticoagulant prophylaxis, but because they may act as markers of underlying cardiovascular disease burden and treatment complexity. In a real ICU, these are not abstract variables; they are clues to the patient’s physiologic reserve and cumulative vascular risk.
The model’s interpretability matters because clinicians are often reluctant to act on algorithms that behave like black boxes. SHAP analysis gives the bedside team a way to see which variables are driving a particular risk estimate, making the output easier to trust and easier to challenge when the clinical picture does not fit.
Recalibration Improved Clinical Usefulness
External validation did more than confirm the model’s transportability problem. It also showed the value of recalibration. Exploratory five-fold cross-fitted intercept recalibration improved overall calibration and predictive accuracy, bringing predicted probabilities closer to observed event rates. In prediction research, this step can be as important as model selection because even a well-separated risk ranking loses clinical value if the absolute risk estimates are systematically too high or too low.
Decision-curve analysis supported potential net benefit at selected low-risk thresholds after recalibration. In plain clinical terms, the model looked more useful once its output was tuned to the external population. That is especially relevant for ICU workflows, where action thresholds are often low and decisions are made under uncertainty. A model that improves decision-making only after recalibration still has value, but it needs local adaptation before deployment.
What This Means for ICU Practice
The study positions interpretable machine learning as a practical adjunct to standard thrombosis prevention in critically ill heart failure patients. The target population is not one in which prophylaxis can be assumed to eliminate risk; it is one in which the remaining residual hazard is concentrated in a subset of patients with recognizable clinical features. A model built from routinely available variables can therefore support individualized surveillance rather than blanket escalation.
The bedside implication is not that every patient with heart failure should be screened more aggressively. It is that some patients, especially those with atrial tachycardia or fibrillation, respiratory failure, prior thromboembolism, or early mechanical ventilation, may warrant closer attention despite early thromboprophylaxis. ICU teams already weigh this type of information informally. A validated prediction model gives that process a structured form and may reduce reliance on memory alone.
Because the analysis relied on deidentified ICU databases, the work also reflects the strengths and limits of observational prediction research. The cohorts were large, the validation was external, and the variables were practical. Yet the study still stops short of proving that model-guided care reduces thrombosis, imaging, or downstream complications. The next question is whether a recalibrated, workflow-integrated tool can improve real-world decisions without increasing alarm fatigue or unnecessary intervention.