Machine learning-based prediction of peripherally inserted central catheter-related thrombosis in hematological malignancies: development and validation of a dynamic model in a multicenter cohort.

The incidence of symptomatic PICC-related thrombosis in the study cohort was 5.8%.

• 232 symptomatic PICC-RT events were identified among 4015 included patients.
• The study screened 5420 patients and ultimately included adult patients with hematological malignancies who underwent PICC insertion across five participating centers.
• The cohort was partitioned into a training set (n=2810) and an independent testing set (n=1205).

The XGBoost model significantly outperformed logistic regression in predicting PICC-related thrombosis on the independent testing set.

• XGBoost achieved an AUC of 0.862 (95% CI: 0.825–0.899) versus logistic regression AUC of 0.774 (95% CI: 0.730–0.818).
• The difference was statistically significant (P < 0.001).
• Three algorithms were compared: logistic regression (LR), random forest (RF), and XGBoost, evaluated using AUC, calibration curves, and decision curve analysis (DCA).

Nine core predictors of PICC-related thrombosis were identified via LASSO regression.

• Independent predictors included history of VTE, use of triple-lumen catheters, immunomodulatory drug (IMiD) use, and peak values of D-dimer and the neutrophil-to-lymphocyte ratio (NLR).
• LASSO regression was used for variable selection from the broader set of clinical and laboratory features.
• The model integrated both static clinical data and dynamic biomarkers to provide individualized risk prediction.

SHAP analysis revealed a distinct threshold effect for peak D-dimer, where thrombosis risk accelerated sharply once levels exceeded 1.5 mg/L.

• SHapley Additive exPlanations (SHAP) were utilized for model interpretation.
• The threshold effect for D-dimer at 1.5 mg/L was identified as a clinically meaningful inflection point for risk escalation.
• SHAP analysis provided individualized feature contribution estimates, supporting clinical interpretability of the model.

The XGBoost model demonstrated a high negative predictive value of 98.6% alongside a positive predictive value of 18.8%.

• The authors attributed the pattern of high NPV and lower PPV to the low baseline incidence of PICC-RT (5.8%).
• The model was described as providing 'a balanced predictive profile given the low baseline incidence.'
• This performance profile was framed as enabling safe identification of low-risk individuals while flagging high-risk patients for enhanced surveillance.

Traditional risk assessment models were characterized as static and lacking hematology-specific predictors, motivating the development of a dynamic ML model.

• The authors noted that 'traditional risk assessment models are often static and lack hematology-specific predictors, resulting in suboptimal performance.'
• The study design incorporated both static clinical features and dynamic longitudinal biomarkers (peak D-dimer, peak NLR) to address this limitation.
• This was a multicenter retrospective study across five participating centers.

An open-access web-based risk calculator was developed to facilitate clinical application of the XGBoost model.

• The calculator is freely available at https://xxcc1114.shinyapps.io/PICC_Calculator/
• The tool was designed to enable identification of high-risk patients who may benefit from 'enhanced surveillance or targeted thromboprophylaxis.'
• The web-based format was intended to support real-world clinical deployment of the model.