Polycythemia vera (PV) is a chronic blood cancer driven by a mutation in the JAK2 gene, which causes the bone marrow to overproduce red blood cells. This excess thickens the blood, dramatically increasing the risk of dangerous blood clots (thromboembolic events or TEs) such as strokes, heart attacks, and deep vein thromboses.
Blood clots are the leading cause of death and serious complications in PV patients. Risk stratification currently relies primarily on age (over 60) and prior clot history, but these simple criteria leave many high-risk patients unidentified until a potentially fatal event occurs.
Hydroxyurea (HU) is the most commonly used first-line treatment for high-risk PV. It reduces the overproduction of blood cells and lowers clot risk. However, even with HU, some patients continue to experience dangerous thrombotic events, and up to 15% of patients develop resistance or intolerance to HU over time.
Ruxolitinib is a JAK1/JAK2 inhibitor approved for PV patients who fail HU. It has demonstrated superior control of blood counts and symptoms. The PV-AIM study aimed to use machine learning on real-world patient data to better identify which HU-treated patients are at highest risk of blood clots and may need treatment escalation.
The PV-AIM study analyzed the Optum de-identified Electronic Health Record (EHR) database, one of the largest real-world clinical databases in the United States. From nearly 83,000 patients with a PV diagnosis, researchers identified 3,852 patients treated with hydroxyurea alone and 130 who later switched to ruxolitinib.
Patient data included demographics, laboratory values (blood counts, hematocrit, red cell distribution width), clinical observations (blood pressure, BMI), anticoagulant medication use, and documented thrombotic events. This richly detailed dataset allowed the analysis of how multiple factors simultaneously affect clot risk.
A random survival forest (RSF) model - a machine learning technique well-suited to time-to-event medical data - was trained to predict whether a patient would experience a thrombotic event during the 6-18 months after starting HU. The model incorporated all available clinical and laboratory variables to identify the most predictive combinations.
The model was externally validated using an independent registry from Croatia including 100 PV patients from three hospitals, providing a critical check on whether findings from the American dataset would translate to a different healthcare system and patient population.
Before starting HU treatment, patients had a baseline clot rate of 8.7 events per 100 patients per year. During initial HU treatment, this fell to 5.6 per 100 - a meaningful but incomplete reduction, confirming that HU provides real but imperfect protection against blood clots.
Among patients who continued HU without switching, the clot rate rebounded sharply to 10.5 per 100 patients per year - actually higher than before treatment started. This concerning finding suggests that continuing inadequate HU therapy may paradoxically worsen outcomes for some patients.
By contrast, patients who switched from HU to ruxolitinib showed a different pattern. Their initial HU period rate fell from 10.8 to 8.4, and after switching to ruxolitinib, the rate remained stable at 8.3 per 100 patients per year - the clot risk was contained rather than rebounding.
These data highlight that timely identification of HU inadequacy and switching to ruxolitinib may prevent the dangerous clot rebound seen in patients who continue on failing HU therapy. The challenge is identifying which patients need to switch before a life-threatening event occurs.
The machine learning model identified ten clinical variables with the greatest impact on predicting blood clots. As expected, prior clot history was the strongest predictor - more than twice as influential as any other factor. After that, BMI, blood pressure, and anticoagulant use all contributed meaningfully.
Most notably, three routine laboratory values emerged as powerful and novel predictors: lymphocyte percentage (LYP), neutrophil percentage (NEP), and red cell distribution width (RDW). These are all measurements from a standard complete blood count - a simple and inexpensive test performed routinely in virtually every clinical encounter.
LYP and RDW showed the strongest synergistic interaction for predicting clots in patients without prior clot history. Patients with RDW below 14.05% and LYP below 19.3% faced significantly elevated risk. Similarly, NEP and RDW showed strong synergy: RDW below 14.3% combined with NEP above 72.05% defined a high-risk group.
The model achieved a receiver operating characteristic area under the curve (AUC) greater than 0.8, indicating strong predictive accuracy. Validation in the Croatian cohort confirmed that these findings were not limited to the US patient population, reinforcing their clinical relevance.
One of the study's most practical contributions was translating the machine learning findings into a simple clinical decision tree. Rather than requiring complex calculations, the decision tree allows a physician to assess clot risk by checking three laboratory values: RDW, NEP, and LYP from a routine blood count.
For patients with no prior clot history, the tree first checks RDW. A low RDW (below approximately 14%) combined with either high neutrophils or low lymphocytes places the patient in a high-risk category that warrants careful monitoring and possible treatment intensification.
This kind of tool addresses a practical gap: physicians treating PV need to make treatment decisions - including when to switch from HU to ruxolitinib - but until now lacked objective, evidence-based criteria for identifying patients who are failing HU before a blood clot occurs.
The decision tree was validated externally and is designed to be used alongside standard clinical judgment. It represents an early example of how machine learning discoveries can be translated into practical, actionable tools that fit naturally into routine clinical practice for blood cancer patients.
The central clinical implication of PV-AIM is that clot risk in PV is dynamic and detectable through routine laboratory measurements, but current practice rarely uses these values systematically to guide treatment decisions. Identifying high-risk patients before a clot occurs - rather than after - could prevent strokes, heart attacks, and deaths.
The finding that continuing inadequate HU is associated with a rebound in clot rates above baseline suggests that the window for switching therapy may be narrow. Patients who remain on HU beyond the point of inadequate control may face compounding risk that switching to ruxolitinib can stabilize but may not fully undo.
The use of real-world electronic health record data rather than idealized clinical trial populations strengthens the study's applicability. PV patients in the real world are older, have more comorbidities, and take more concurrent medications than clinical trial participants. The findings are therefore more likely to translate to everyday clinical practice.
More broadly, PV-AIM demonstrates the potential of applying machine learning to large real-world datasets to discover clinically actionable biomarkers in blood cancers. Similar approaches could be applied to other myeloproliferative neoplasms and to optimizing treatment selection across a wider range of hematological malignancies.