Development and internal validation of a modified Caprini score for postoperative venous thromboembolism in gastrointestinal adenocarcinoma: a retrospective cohort study

Background The conventional Caprini score may have limited discriminatory ability in patients undergoing gastrointestinal cancer surgery because many are classified as high risk. This study examined factors associated with postoperative venous thromboembolism (VTE) and developed and internally validated a modified Caprini score. Methods This retrospective cohort study included 220 patients with gastrointestinal adenocarcinoma who underwent surgery between June 2023 and December 2024. To reduce double counting of risk information already captured by the conventional Caprini score, its logistic regression coefficient was fixed at 1, and incremental coefficients were estimated for tumor-node-metastasis stage III/IV, operative time ≥ 240 min, perioperative blood transfusion, and postoperative day 1 (POD1) D-dimer > 5.0 mg/L. Three missing or non-POD1 D-dimer values were median-imputed, and a complete-case sensitivity analysis was performed. Model performance was assessed using receiver operating characteristic analysis, the DeLong test, calibration intercept and slope, the Brier score, the index of prediction accuracy (IPA), and 1,000 bootstrap resamples. Results Postoperative VTE occurred in 49 patients (22.3%). After accounting for the conventional Caprini score, stage III/IV disease [odds ratio (OR), 3.587; P = 0.001], perioperative blood transfusion (OR, 2.358; P = 0.024), and POD1 D-dimer > 5.0 mg/L (OR, 2.937; P = 0.005) provided significant incremental information. Operative time ≥ 240 min had a positive but non-significant association (OR, 2.060; P = 0.076) and was retained in the score. The apparent area under the curve (AUC) was 0.726 for the modified score and 0.571 for the conventional score (DeLong Z = 3.983; P < 0.001). After bootstrap correction, the AUC was 0.711, calibration intercept was − 0.004, calibration slope was 0.942, Brier score was 0.150, and IPA was 13.7%. The grouped Hosmer-Lemeshow test indicated local miscalibration (χ²=10.025; df = 4; P = 0.040). At a cutoff of ≥ 7 points, sensitivity was 59.2% and specificity was 82.5%. Conclusions The modified Caprini score showed better discrimination than the conventional score in this cohort. However, the modest incremental predictive accuracy, grouped calibration deviation, instability of integer weights, and uncertain timing of POD1 D-dimer sampling relative to VTE diagnosis warrant cautious interpretation. External validation and prospective evaluation are required before clinical implementation.