Development and internal validation of a nomogram for early prediction of deep vein thrombosis after liver transplantation: an exploratory machine-learning comparison.

BACKGROUND: Lower-extremity deep vein thrombosis (DVT) after liver transplantation (LT) may delay recovery and complicate thromboprophylaxis. Practical transplant-specific tools for early risk stratification remain limited. METHODS: Adult recipients who underwent LT between August 2023 and January 2026 were retrospectively screened. After exclusions, 131 recipients were divided by outcome-stratified random sampling into a training cohort (n = 99) and a hold-out subset (n = 32). The outcome was newly detected lower-extremity DVT identified by bilateral duplex color Doppler ultrasonography within the first 7 postoperative days. Postoperative laboratory variables were obtained from the first urgent blood test, generally within 1 h after completion of LT and before initiation of postoperative pharmacologic anticoagulation according to routine workflow. Least absolute shrinkage and selection operator regression was used for penalized feature selection, followed by multivariable logistic regression and nomogram construction. Random forest, support vector machine, LightGBM, and XGBoost were evaluated as exploratory comparators. Performance was assessed using discrimination, threshold-dependent classification measures, calibration, the Hosmer-Lemeshow test, decision curve analysis, stratified 5-fold cross-validation, and bootstrap optimism correction. RESULTS: DVT occurred in 26 of 131 recipients (19.8%). The logistic model included postoperative ln (D-dimer), anhepatic phase, postoperative activated partial thromboplastin time, and preoperative prothrombin time. Its AUC was 0.79 (95% CI, 0.67-0.90) in the training cohort and 0.81 (95% CI, 0.56-1.00) in the hold-out subset. The 5-fold cross-validated AUC was 0.76 (95% CI, 0.63-0.86), and the bootstrap optimism-corrected AUC was 0.76. Wide, overlapping confidence intervals and model-specific optimism did not support reliable algorithm ranking. CONCLUSION: A concise logistic regression-based nomogram using routinely available perioperative variables may support early individualized DVT risk estimation after LT. The machine-learning analyses should be regarded as exploratory. Larger prospective multicenter studies are required before clinical implementation.