Interactions of Key Hematological Parameters with Red Cell Distribution Width (RDW) Are Associated with Incidence of Thromboembolic Events (TEs) in Polycythemia Vera (PV) Patients: A Machine Learning Study (PV-AIM)

Aims: We previously reported results of the first machine learning study to identify the novel biomarkers associated with crude incidence of TEs in PV patients (pts) treated with hydroxyurea (HU) (Verstovsek et al, Blood 2019). In the current study, we have expanded the database to perform an in-depth machine learning analysis of biomarkers predicting the annual standardized incidence rate (IR) of TEs in PV pts treated with HU. In addition, we examined the IR of TEs in HU-treated pts and pts treated first with HU, then switched to ruxolitinib (HU-RUX). Methods: The study is part of PV-AIM (polycythemia vera advanced integrated models for prevention of TEs). The US OPTUM database includes the electronic medical records from 105 million pts (2007-2020), including 82 960 PV pts with a median record length of 8.4 years. TEs were assessed (A) before the treatment (Tx) initiation in both groups; (B) while on HU (median: 29 months) and (C) while continuing HU or switching to RUX (median, 17 months). A total of 3852 HU-alone pts and 130 HU-RUX pts passed inclusion/exclusion criteria. To avoid selection bias, only HU-alone pts treated prior to RUX market launch were included (n=704). Cohorts were then matched by selecting the nearest cases from the HU-alone groups based on the propensity scores calculated from the total Tx period, gender, race, age at index, and geographical division (n=130 pts in both HU-alone and HU-RUX). TEs were identified from a restrictive list of ICD-CM diagnosis codes defined in the RESPONSE study (Kiladjian et al, Lancet Haematol 2020). Annualized IR of TEs were then calculated per 100 pts for each of the above Tx periods for each cohort. A random survival forest (RSF) model was then constructed for HU-alone pts, with at least 6 months of HU Tx and 18 months of follow-up, to predict the risk of TEs 6 to 18 months after the first HU Tx. Pts with at least 1 lab test and 1 observation obtained at 3 to 6 months post index were included in the model (n=1012). The features also included demographics, history of TEs, phlebotomy and anti-coagulant use. The performance was assessed on a 70:30 train:test split using Receiver Operating Characteristic-Area Under the Curve (ROC-AUC). RSF variable importance was then used to identify the variables with the largest impact on prediction of TEs. The feature interactions for the top 10 features were mapped and assessed in terms of risk of TEs via log-rank using a brute-force combinatorial approach. Risk was