Development and temporal validation of a clinical prediction score for deep vein thrombosis following femoral neck fracture surgery: a retrospective cohort study with decision curve analysis.

backgroundDeep vein thrombosis (DVT) is a frequent complication after femoral neck fracture surgery, with reported incidence rates of 6-19% depending on detection methods. Existing risk assessment tools, such as the Caprini score, show limited discrimination (AUC 0.62-0.75) in orthopedic trauma populations. We aimed to develop and temporally validate a clinical prediction score specifically designed for DVT risk stratification in patients undergoing femoral neck fracture surgery.

methodsThis retrospective cohort study included 2,241 consecutive patients undergoing femoral neck fracture surgery at a single Serbian center (January 2018-June 2024). The cohort was split temporally into a development set (n = 1,829; 2018-2023) and a temporal validation set (n = 412; January-June 2024). The primary endpoint was DVT diagnosed after surgery on scheduled bilateral duplex ultrasonography (24-48 h and day 7) or symptom-triggered imaging during 90-day follow-up. A multivariable logistic model with 12 prespecified predictors was developed and internally validated with 1,000 bootstrap resamples. Performance was assessed by AUC, calibration slope, Brier score, and decision curve analysis.

resultsDVT occurred in 269 patients (12.0%). The DVT Risk Score (DVTS) incorporated 12 predictors, including age > 75 years, D-dimer > 500 ng/mL, albumin < 35 g/L, and ASA score ≥ 3. The optimism-corrected development AUC of the shrinkage-adjusted logistic model was 0.80 (95% CI: 0.76-0.84), and its temporal-validation AUC was 0.82 (95% CI: 0.77-0.87). The simplified integer DVTS had an apparent AUC of 0.81 in development and 0.81 (95% CI: 0.76-0.86) in temporal validation. Calibration slope was 0.94 (95% CI: 0.89-0.99) in validation; Brier scores were 0.087 (development) and 0.084 (validation). At the DVTS ≥ 5 threshold in the validation cohort, sensitivity was 85.4% and specificity 41.8%, with a negative predictive value of 95.6%; a higher threshold (≥ 9) yielded specificity of 81.6%. Three risk categories showed DVT incidence of 3.3% (low), 11.7% (moderate), and 29.2% (high) (p < 0.001 for trend).

conclusionThe shrinkage-adjusted logistic model maintained good discrimination and acceptable calibration in temporal validation, and the simplified integer DVTS retained good discrimination (AUC 0.81) and stratified patients into clinically interpretable risk categories. External validation and prospective impact evaluation are required before clinical implementation.