A spatiotemporally validated machine-learning score for venous thromboembolism risk after neurosurgery

Technical Summary: A Spatiotemporally Validated Machine-Learning Score for VTE Risk After Neurosurgery

Problem Statement
Venous thromboembolism (VTE) is a life-threatening complication following neurosurgery, yet current risk stratification tools face significant limitations. The widely used Caprini risk assessment model often overestimates risk in neurosurgical populations because it assigns points for general surgical factors (e.g., prolonged surgery, restricted mobility) that are ubiquitous in this cohort, leading to a "ceiling effect" where most patients are categorized as high-risk. This reduces specificity and hinders individualized thromboprophylaxis. Furthermore, existing machine-learning (ML) models for VTE prediction often lack interpretability, require excessive clinical variables that increase assessment burden, or fail to undergo rigorous temporal and external validation, limiting their generalizability across different institutions and clinical environments.

Methodology
The study employed a retrospective prediction-model development and validation design adhering to TRIPOD and STROBE guidelines.

• Data Sources: The internal dataset comprised 3,644 adult neurosurgical patients from Huashan Hospital, Fudan University (August 2018–December 2021). To optimize computational efficiency while retaining all outcome events, a 1:4 case-control sampling strategy was used (731 VTE cases and 2,913 controls), artificially increasing the VTE prevalence to ~20% for model training. This cohort was chronologically split into a training set (n=2,627) and a temporal test set (n=1,017). An independent external validation was performed using the MIMIC-IV database (n=10,268), where 4.34% developed VTE.
• Feature Selection: Candidate predictors were limited to routinely available perioperative variables, explicitly excluding the Caprini score to avoid incorporation bias. Initial feature reduction distilled 25 candidates to 15 core features.
• Model Development: Four modeling architectures were evaluated using nested five-fold cross-validation: Random Forest, Extreme Gradient Boosting (XGBoost), LASSO-penalized logistic regression, and conventional logistic regression. XGBoost demonstrated the highest discrimination (AUC=0.790) in the internal test set.
• Interpretability and Simplification: SHapley Additive exPlanations (SHAP) and permutation importance were applied to the optimal XGBoost model to identify the most influential predictors. This process isolated four key variables: peak D-dimer, age, blood transfusion, and body mass index (BMI).
• Score Derivation: A simplified clinical score (0–18 points) was derived by assigning points proportional to the relative mean SHAP values of these four categorized predictors. Risk categories were defined as low (0–3), moderate (4–7), and high (≥8).
• Validation: The model underwent spatiotemporal validation, assessing discrimination (AUC), calibration, clinical utility (Decision Curve Analysis), and risk reclassification against the Caprini model in both the internal temporal test cohort and the external MIMIC-IV cohort.

Key Results

• Predictive Performance: The simplified four-variable score significantly outperformed the Caprini model in all validation settings.
  • Internal Test Cohort: AUC of 0.745 (95% CI, 0.707–0.782) for the new score vs. 0.662 for Caprini (P < 0.0001).
  • External Validation (MIMIC-IV): AUC of 0.726 (95% CI, 0.703–0.749) for the new score vs. 0.642 for Caprini (P < 0.0001).
  • The simplified score was statistically non-inferior to the full continuous logistic regression model, confirming that reducing the model to four variables preserved predictive integrity.
• Clinical Utility: Decision Curve Analysis (DCA) demonstrated that the new score provided a consistently greater net clinical benefit across all relevant threshold probabilities compared to the Caprini tool.
• Risk Reclassification: Sankey diagram analysis revealed that the new score systematically and safely downgraded a substantial proportion of patients whom the Caprini model had inaccurately overestimated as high-risk, reassigning them to moderate or low-risk categories. This suggests the new model can mitigate systemic over-alerting.
• Feature Importance: Peak D-dimer emerged as the dominant predictor, followed by age, blood transfusion, and BMI. Notably, blood type showed no significant association in this cohort, contrasting with some general literature.

Key Contributions

1. Development of a Parsimonious Score: The study successfully translated complex machine-learning findings into a concise, interpretable four-variable score (Peak D-dimer, Age, Blood Transfusion, BMI) that is feasible for bedside application.
2. Rigorous Spatiotemporal Validation: Unlike many prior studies, this model was validated not only temporally (using a distinct time period from the training data) but also geographically and institutionally (using the independent MIMIC-IV database), demonstrating robust transportability.
3. Addressing Overestimation: The study provides empirical evidence that the Caprini model overestimates VTE risk in neurosurgical patients and offers a validated alternative that improves specificity without sacrificing sensitivity.
4. Clinical Implementation Tools: The authors developed a user-friendly clinical nomogram and a dynamic web-based calculator to facilitate the adoption of this risk stratification tool.

Significance and Claims
The paper claims that this spatiotemporally validated four-variable score offers a superior method for VTE risk stratification in neurosurgery compared to conventional frameworks. By utilizing only routinely available early postoperative parameters, the tool aims to reduce the assessment burden on clinicians while enabling more precise, individualized thromboprophylaxis strategies. The authors propose a risk-stratified clinical pathway where low-risk patients receive mechanical prophylaxis and early mobilization, while moderate- and high-risk patients receive tailored pharmacological interventions. The study concludes that machine learning can substantially enhance VTE prediction accuracy while maintaining the feasibility required for routine clinical practice, thereby supporting safer management of this high-risk population.