Development and validation of a risk assessment model for inpatient venous thrombosis in people with cancer

Abstract

Abstract Introduction Venous thromboembolism (VTE) acquired during hospitalization (hospital-acquired VTE) causes increased morbidity and mortality. Risk assessment models (RAM) for hospital-acquired VTE have not been widely assessed in people with cancer or have performed poorly. Our objective was to develop and validate a RAM for hospital-acquired VTE specifically for patients admitted to medical services with active cancer. Methods Medical admissions to 6 academic health systems from January 1, 2016, to December 31, 2022, were included. Admissions from 4 health systems in Michigan, Vermont, Minnesota, and Pennsylvania were used for development while admissions from 2 health systems in North Carolina and Texas were used for validation. Inclusion criteria were age ≥18 years, admission to a medical service, and cancer present at admission (defined as the presence of an ICD-10 code for cancer on the discharge summary with present on admission flag=yes). We excluded those with VTE at admission or hospitalization <1 midnight. Candidate risk factors for model development were: (1) ascertainable within 24 hours of admission; (2) objectively measured and reproducible; (3) available for >90% of hospital admissions; and (4) available within the electronic health record. The primary outcome was hospital-acquired VTE, defined using a previously validated computable phenotype (). We developed RAMs using two different approaches. In the first, we assessed the performance of a previously developed and validated RAM (), refit with the addition of a variable categorizing cancer into specific cancer types. Other variables included were history of VTE, serum creatinine >2 mg/dL, hemoglobin <13.6 (men) or <12.0 (women) g/dL, red cell distribution width ≥ 14.7%, serum sodium <136 mmol/L, malnutrition, and anticoagulation level within 24h of admission (none, prophylactic, full). For the second approach, we used Bayesian LASSO logistic regression to develop a de-novo RAM using a set of 65 candidate predictor variables. Only predictors with a t-statistic ≥1.4 were included in the final model. In both approaches, maximum level of anticoagulation at admission was also included in the model to address potential confounding by indication. Model fit was assessed using the area under the receiver operating curve (AUC) and calibration slope. Results The development and validation cohorts consisted of 123,750 and 58,124 inpatients with active cancer, respectively. Of those, 709 (0.6%) in the development cohort developed hospital-acquired VTE (36.5% pulmonary embolism [PE], 39.9% lower extremity deep vein thrombosis [DVT], and 23.6% upper extremity DVT), and 298 (0.5%) in the validation cohort developed hospital-acquired VTE (29.2% PE, 29.5% lower extremity DVT, 41.3% upper extremity DVT). The development cohort had a mean (SD) age of 65.6 (14.6) years, were 46% female, and was 79% White race and 14% Black race. The validation cohort had a mean age of 64.6 (15.6) years, were 48% female, and was 54% White race and 25% Black race. The most common cancer types were hematologic cancers excluding aggressive lymphoma (22.4%), lung cancer (19.5%), genitourinary cancers (16.0%), breast cancer (10.2%) and aggressive lymphomas (9.2%), with other cancer types individually <9%. For approach 1 (modifying the existing RAM), the AUC and calibration slope were 0.66 and 0.67 in the development cohort and 0.60 and 0.60 in the validation cohort, respectively. For approach 2 (developing a de-novo RAM), the risk factors retained in the model were (odds ratio; 95% credible intervals): prior VTE (2.95; 2.31,3.72), malnutrition (1.52; 1.23, 1.87), hematologic malignancy excluding aggressive lymphomas (1.42; 1.13, 1.78), lung cancer (1.39; 1.08, 1.79), rheumatologic disease (1.52; 1.01, 2.21), age ≥75 years (0.57; 0.43, 0.76), and COPD (OR 0.71; 0.5, 0.97). The AUC and calibration slope were 0.68 and 1.07 in the development cohort and 0.64 and 0.84 in the validation cohort, respectively. ConclusionsVTE affecting medical inpatients with cancer is predictable using a novel RAM, which performed better than a non-cancer specific model. This RAM, along with a validated RAM to predict bleeding risk (PMID40683552), will allow for individualized prevention strategies to minimize both VTE and bleeding risk.