Implementation of Automated Thrombosis and Bleeding Risk Assessment in Cancer Patients Undergoing Systemic Therapy

Introduction: Despite improved individualized risk predictions for venous thromboembolism (VTE) in cancer patients, ambulatory thromboprophylaxis is rarely implemented due to the lack of automated risk calculator and the challenge to mitigate bleeding risk. We examined the utility of an automated algorithm to minimize bleeding risk and recognize VTE risk. Specifically, we assessed the performance of clinical trial exclusion criteria for patients at-risk of clinically relevant bleeding (CRB) and derived an improved risk model.

Methods: We performed a retrospective cohort study at Harris Health System from 1/2017 to 1/2023. New line of therapy (LOT), treatment date/type, cancer type, and staging were identified using Epic Beacon, the module where oncologists manage chemotherapy plans. Additional risk predictors for VTE (updated EHR-CAT model, https://doi.org/10.1200/JCO.22.01542) or CRB (clinical trial exclusions from AVERT and CASSINI) were defined using Epic Clarity. To facilitate ease-of-use and implementation, all data (including medications) were extracted and curated in a one-step SQL algorithm without additional statistical software processing. VTE outcomes were ascertained using a previously validated NLP algorithm (VTE-BERT). CRB outcomes were determined using ICD-10 codes associated with bleeding from inpatient billing final diagnosis. To improve coding accuracy, we performed a systematic review of peer-reviewed manuscripts on ICD-10 bleeding codes and identified 545 non-duplicated codes from 8 relevant studies. We then performed a two-physician independent review, followed by chart review to adjudicate ambiguous codes (i.e. posthemorrhagic anemia, unspecified hematuria, epistaxis). A final list of 279 ICD-10 codes were included to define non-surgical CRB. For patients with multiple LOT, we randomly selected one line as the index date and assessed the incidence of VTE or CRB at 6 months after therapy initiation. We used logistic regressions (odds ratio) and AUC/ROC to determine covariate importance and model discrimination.

Results: We identified 7,640 cancer patients with 14,151 distinct LOT over 6 years. After random sampling, 66% had 1^st, 23% 2^nd, and 11% 3^rd or more LOT at index date. At 6 months after index, VTE occurred in 10.1% and CRB in 5.7% of the cohort. Thirty percent (2300/7640) of patients would meet at least one clinical trial exclusion. Non-mutually exclusive reasons included 1) medication: anticoagulant (10%), non-aspirin antiplatelet (2%), CYP3A4/P-GP drug (3%); 2) cancer type: brain metastasis (6%), brain primary (2%), acute leukemia (3%); 3) laboratory: platelet <50 (2%), eGFR <30 (2%), elevated ALT or bilirubin (2%); 4) other: recent bleeding (6%), weight <40 kg (1%).

The 6-month CRB rate was 10.4% in the excluded group vs. 3.5% in the remaining. The multivariable model including all 11 trial predictors had an AUC of 0.66. All covariates had clinically meaningful association with CRB except primary brain cancer (OR of 0.69) and antiplatelet use (OR of 0.93). A refitted model including stratified cancer type, stage, brain metastasis, recent bleed, recent hospitalization, anticoagulant use, hemoglobin, platelet, eGFR, and albumin had an improved performance on internal validation (AUC=0.74).

After removing 30% of patients meeting the clinical trial exclusion criteria, the EHR-CAT risk score stratified the remaining 5,340 patients into 6 categories. The 6-month VTE rates were 1.3%, 3.2%, 5.0%, 5.6%, 7.8%, and 13.5% for scores 0, 1, 2, 3, 4, and 5, respectively (AUC=0.68). In contrast, the Khorana score had a lower discrimination (AUC=0.59). Neither the LOT count nor time from 1^st line significantly impacted risk prediction in either model.

Conclusion: We demonstrated the feasibility of an automated algorithm to extract and curate data directly from electronic health record to exclude patients at-risk for CRB and stratify those at-risk for VTE. The pre-specified clinical trial exclusion criteria performed reasonably well at identifying bleeding risk; however, the parameters could be further optimized for better prediction. The EHR-CAT model also retained its discrimination when applied beyond 1^st LOT and after exclusion for bleeding risk. With external site validation and adoption, this Epic-based algorithm may aid the development of individualized approach to ambulatory thromboprophylaxis in patients with cancer.