MYSEC-Prognostic Model and previous thrombotic events predict the risk of thrombosis in post polycythemia vera and post essential thrombocythemia myelofibrosis: a study of the MYSEC Group

Background: Information on thrombosis risk factors in post polycythemia vera (PV) and post essential thrombocythemia (ET) myelofibrosis (also known as secondary MF, SMF) is scant, while survival is predicted by the MYSEC-PM (Myelofibrosis Secondary to PV and ET-Prognostic Model). Aims: The primary objective of this study is to assess predictors of thrombosis occurrence in a large cohort of SMF. Methods: The MYSEC project is an international effort that retrospectively collected 805 SMF. In order to select the most relevant predictors of thrombotic risk, we performed a machine learning approach called Random Forest (RF). RF uses a large number of individual binary decision regression trees based on bootstrap samples of the training data, in order to achieve a high degree of accuracy. Our model was trained in R (version 3.2.0) and composed by 1000 regression trees. Firstly, we built a RF model that considered phenotypic and genotypic features at SMF diagnosis. Model-1 included: gender, age, hemoglobin value, platelets and leukocytes count, peripheral blast cells percentage, bone marrow fibrosis (BMF) grade 2 vs 3, driver mutation status, prior thrombotic events, type of diagnosis (previous PV vs ET) and presence of constitutional symptoms. Then, in order to evaluate the role of the MYSEC-PM in predicting thrombotic risk, we created Model-2, where MYSEC-PM components have been aggregated. Relevance of different variables was evaluated in terms of Mean Minimal Depth (MMD), Accuracy Decrease (AD) and Times a Root (TR). MMD expresses the variable proximity to the tree “root”: the lower the MMD, the more powerful is the variable, since more trees are generated on its basis. AD measures the decrease in predictive accuracy secondary to variable permutation and its value is proportional to the importance of the variable. TR represents the total number of trees in which the variable is used for splitting the root node: the higher the TR, the stronger is the variable. Results: Within 675 SMF patients annotated for driver mutation status and BMF grading, 73 (10.8%) had at least one thrombotic complication during a median follow up of 3.1 (range, 0.6-27) years. Thrombosis incidence rate was 3 x 100 person-year of follow up (95% CI: 2.4-3.8). The involved vascular district was venous in 45 (6.7%) and arterial in 28 (4.1%) cases, with an incidence rate of 1.7 x 100 (95% CI: 1.3-2.3) and 1.1 x 100 (95% CI: 0.7-1.6) person-year of follow up, respectively. In RF Model-1, MMD was low for continuous variables like hemoglobin value, platelets and leukocytes count and age at SMF diagnosis, along with history of thrombotic complications before SMF. AD and TR were high only for the latter two parameters. Since some of the factors included in the MYSEC-PM had low MMD but not significant AD and TR when used as continuous variables in Model-1, we looked at their relevance when aggregated as in the prognostic score (Model-2). As shown in Figure 1a, the MYSEC-PM and previous thrombotic events had very low MMD in Model-2. Besides, the same variables were also characterized by the highest values of AD and TR (Figure 1b). Summary/Conclusion: In our dataset of 675 SMF patients, ~ 11% had at least one thrombotic event during follow up, with a median incidence rate of 3 x 100 person-year. In a random forest model, the MYSEC-PM and previous thrombosis predict with high accuracy thrombotic risk after SMF evolution, paving the way for new applications of MYSEC-PM in this disease.