Miscalibration of venous thromboembolic risk assessment models across race and ethnicity in medical patients: A cross-model analysis, correction strategy and potential mediators

Abstract Background: Hospital-associated venous thromboembolism (VTE) affects 350,000 US patients annually and is the leading cause of preventable hospital death. Risk assessment models (RAMs) are clinical decision tools to guide decisions on VTE chemoprophylaxis. Previously, we presented a retrospective analysis suggesting a systematic risk misclassification in Black and Latino surgical patients using the Caprini RAM, which is mitigated with a race/ethnicity-based score correction (Nassar, ASH 2024). We hypothesized that Caprini and other standard RAMs share a similar systematic bias estimating VTE risk in Black and Latino medical patients. In further studies, we explored potential mediators of these biases through analysis of the structure and components of these RAMs. Methods: We conducted a retrospective cohort study of 40,672 adult medical inpatients admitted to a single urban academic medical center serving a racially and ethnically diverse population between Jan 2010 and Mar 2023. Collected data underwent quality checks for diagnostic coding consistency and clinical feature prevalence. The primary study outcome was VTE, defined by presence of a new diagnostic ICD-9/10 code. The primary study variables were self-reported race and ethnicity, with Black race and Latino ethnicity evaluated as distinct binary variables. RAMs analyzed included the Caprini, Padua, and Geneva RAMs. Risk scores for each RAM were calculated according to published criteria and underwent post hoc plausibility checks for total score and component distributions. Race/Ethnicity-Aware (REA) RAMs were generated by stepwise point addition/subtraction for Black race or Latino ethnicity. VTE prediction ability was compared between standard RAM vs REA-RAM scores with multivariable logistic regression, adjusting for race/ethnicity as a covariate. Model performance was evaluated using area under the ROC curve (AUC). Moderation was tested via multivariable logistic regression with component-wise interaction with race/ethnicity and backward stepwise approach to isolate significant effect modifiers. Results: Median age of cohort was 52 (IQR 37-65) yrs, 51% female, 52% Black (n=21,094), 23% Latino (n=9,334). Overall VTE incidence was 4.5%, 5.2% in Black patients, and 3.2% in Latino patients. RAM score distributions were similar for all RAMs by race/ethnicity. All RAMs underestimated VTE risk in Black patients in a consistent, robust manner (ORs: Caprini 1.32, Padua 1.35, Geneva 1.34; all P<0.001) and in their overestimation of VTE risk in Latino patients (ORs: Caprini 0.67, Padua 0.68, Geneva 0.67; all P<0.001). Optimal score adjustments for Black patients in the Caprini, Padua, and Geneva RAMs were +1, + 2, and +1 points, respectively. REA-RAM scores corrected VTE risk underestimation (ORs: Caprini 1.05, Padua 0.98, Geneva 1.03; all P>0.3) with a mean AUC of 0.003. For Latino patients, the optimal score adjustment for Caprini RAM was -2, with REA-RAM scores correcting VTE risk overestimation (OR: 1.04; P=0.5) with an AUC increase of 0.0009. Padua and Geneva score adjustments could not optimize performance of the RAMs in Latino medical patients. On moderation analysis of Caprini, recent heart failure showed effect modification among Black patients (interaction β=1.32, P=0.03) and Latino patients (interaction β=-0.35, P=0.02), suggesting an interaction effect. Conclusion: We found that in general, standard RAMs used to predict VTE in medical patients share the same systematic biases in Black and Latino patients, in alignment with our previously described findings with Caprini in surgical patients. Furthermore, score additions in Black patients are able to mitigate these biases to optimize VTE chemoprophylaxis strategies. Unfortunately, for Latino patients, score adjustments had limited utility in all RAMs except for Caprini, likely reflecting limited ability to refine marginal risk estimates when baseline VTE risk is already very low. Systematic risk overestimation in Latino patients thus increases risk of bleeding from unnecessary chemoprophylaxis. External validation of these results are underway at multiple additional academic sites. The mechanism of these observed VTE risk discrepancies remains unclear. Our preliminary moderation analysis suggests that this may relate to disparities in heart failure phenotype, documentation, or management. Further studies are needed to further confirm and elucidate the underlying mechanisms of differential VTE risk.