DEVELOPMENT OF CLINICAL AND MACHINE LEARNING MODELS FOR PREDICTING CATHETER-ASSOCIATED VENOUS THROMBOSIS IN HOSPITALIZED CHILDREN

Year: 2026

Type: article

Abstract: Background Venous catheters (VC) are a leading cause of venous thromboembolism in hospitalized children. Existing risk scores rely on baseline characteristics and inadequately capture acute clinical and procedural factors influencing Venous Catheters Thrombosis (VCT). Aims To develop and internally validate a clinical risk score and, in parallel, a machine learning (ML)-based model for VCT in a multicenter hospitalized pediatric cohort, comparing their performance and interpretability. Methods A prospective multicenter cohort registry (7 centers) which included 854 children with VC. All ultrasound screening for VCT during hospitalization were performed in defined times. Clinical and baseline characteristics were assessed. Data were split into derivation/training (70%) and validation/test (30%) sets. For the clinical score, predictors were selected by multivariable logistic regression, with variable weights and cutoffs optimized using the Grey Wolf Optimizer metaheuristic; performance was assessed by bootstrap resampling. In parallel, ten clinically relevant variables were evaluated using Light Gradient Boosting Machine (LightGBM), Random Forest, and Decision Tree algorithms, with metrics estimated via bootstrap 95% confidence intervals (CI) and interpretability assessed using SHapley Additive exPlanations (SHAP). Results Patients developing VCT (23.8%, n = 203) were significantly younger (median 18 vs. 53.5 months) and exhibited higher baseline comorbidities, including respiratory (35.0% vs. 20.7%), congenital (12.8% vs. 6.0%), and cardiovascular diseases (10.3% vs. 4.6%), with 56.6% requiring ICU admission. VCT cases presented higher rates of internal jugular access (73.4% vs. 41.9%), right upper limb placement (70.4% vs. 58.7%), and central vascular catheters (87.2% vs. 57.3%). Anticoagulation therapy was administered to 20.7% of VTE patients versus 5.4% of non-VCT controls (n = 651). The clinical score incorporates systemic infection (+3), major surgery (+2), right-sided placement (+1), catheter diameter >4 Fr (+1), >3 attempts (+1), and platelets >350 × 10⁹/L (+1), HCM ≤26.5 pg (+1), showing strong discrimination (C-index derivation 0.76; validation 0.79). A ≥4-point threshold balanced performance (NPV 0.87–0.91; VTE rate 48–52% vs. 13–15% in low-risk), whereas a ≥5-point threshold prioritized specificity (>0.99) and PPV (0.90–0.95). Among machine learning models, LightGBM performed best (C-index 0.74, sensitivity 0.60, specificity 0.83), with SHAP confirming systemic infection and major surgery as top predictors. Discussion ML and clinical scoring models converged on identical biological predictors for VCT, reinforcing their clinical validity. The clinical score enables rapid bedside risk stratification, whereas the ML model offers higher sensitivity by accounting for complex nonlinear variable interactions. Conclusion This novel pediatric VCT risk score demonstrates robust performance using readily available variables. A ≥4-point cutoff offers optimal utility for risk stratification, while a ≥5-point cutoff identifies a subgroup at extremely high thrombotic risk. The high incidence of asymptomatic VCT reinforce further research into monitoring and intervention strategies for high-risk pediatric patients. Approved by Ethics Committee of UNICAMP (CAAE:63131922.8.0000.5404). Supported by FAPESP (#2021/11963-0).

Source: Hematology Transfusion and Cell Therapy

DOI: 10.1016/j.htct.2026.108081

Language: English

Topic: Venous Thromboembolism Diagnosis and Management

Subfield: Internal Medicine

Field: Medicine

Domain: Health Sciences

SDG: Good health and well-being

Open Access status: diamond

OpenAlex ID: W7223100577

Development of Clinical and Machine Learning Models for Predicting Catheter-associated Venous Thrombosis in Hospitalized Children