Joshua Bliss ¹ Spencer Krichevsky ² Joseph Michael Scandura ¹ Ghaith Abu-Zeinah ^1*

• ¹ Weill Cornell Medical Center, NewYork-Presbyterian, New York City, United States
• ² Stony Brook University, Stony Brook, New York, United States

The myeloproliferative neoplasms (MPNs) – polycythemia vera, essential thrombocytosis, and primary myelofibrosis – are chronic blood cancers that originate from hematopoietic stem cells carrying driver mutations which activate cytokine signaling pathways in hematopoiesis. MPNs are associated with high symptom burden and potentially fatal events including thrombosis and progression to more aggressive myeloid neoplasms. Despite shared driver mutations and cell of origin, MPNs have an extremely heterogenous clinical course. Their phenotypic heterogeneity, coupled with their natural history spanning several years to decades, makes personalized risk assessment difficult. Risk assessment is necessary to identify patients with MPNs most likely to benefit from clinical trials aimed at improving thrombosis-free, progression-free and/or overall survival. For MPN trials to be powered for survival endpoints with a feasibly attained sample size and study duration, risk models with higher sensitivity and positive predictive value are required. Traditional MPN risk models, generally linear models comprised of binary variables, fall short in making such trials feasible for patients with heterogenous phenotypes. Accurate and personalized risk modelling to expedite survival-focused interventional MPN trials is potentially feasible using machine learning (ML) because models are trained to identify complex predictive patterns in large datasets. With automated retrievability of large, longitudinal data from electronic health records, there is tremendous potential in using these data to develop ML models for accurate and personalized risk assessment.