Utilizing machine learning to predict thrombosis in polycythemia vera

Thrombosis remains one of the leading causes of morbidity and mortality in patients diagnosed with polycythemia vera (PV). As a result, there is a growing need for precise models to identify patients at the highest risk of thrombotic events. Artificial intelligence (AI) offers a potential solution in the form of sensitive, precise, and dynamic prediction models that can be used for risk prediction during clinical trials. The use of AI in the diagnosis of myeloproliferative neoplasms (MPN) has been previously reported by Royston in our recent MPN Hub Steering Committee meeting.

During the 15th International Congress on MPN, Abu-Zeinah gave a presentation on the use of AI and machine learning to predict thrombotic events in patients diagnosed with PV. Here, we summarize the key points.

The need to improve thrombotic prediction

• Patients who are both <60 years and >60 years old have double the cardiovascular mortality risk compared with the general population
• Over 3–5 years, the overall mortality risk for patients with PV is 10%
  • - A third of this risk is due to thrombotic complications
• For the feasibility of randomized clinical trials that investigate thrombotic risk, around 200 patients are required, together with a model that provides a predictive value of ≥20%
• A unifying diagnostic model is therefore urgently needed
  • Both the positive predictive value and sensitivity would need to be 75%

A unifying diagnostic model

• The proposed model should be dynamic and be able to predict short-term risk at any clinical visit
• The advantages of a machine learning model include:
  • Automated feature selection
    • Handle large data sets and identify new variables and patterns
  • Improved accuracy
  • Personalized and real-time assessment
    • Ability to learn and adapt to changing patient data
    • Individualized predictions

Results from a large PV cohort with longitudinal data

• A recent study of 470 patients accrued several variables:
  • 8,500 clinic visits;
  • 1.4 million data elements; and
  • >2,100 parameters.
• Results showed thrombosis was non-linear
  • There was an increased incidence of 4.4% in the first 2 years vs 1% in subsequent years.
  • This was similar for recurrent thrombotic events, whereby incidence in the first 2 years was 9.7% and incidence in subsequent years was 1.8%.

The machine learning approach

• The proposed prediction model identified several important variables, including:
  • time since diagnosis;
  • age;
  • lactate dehydrogenase levels at last check;
  • hematocrit levels; and
  • blood type.
• Important variables identified by the model that have historically been underappreciated in PV research included:
  • time since last thrombotic event;
  • body mass index; and
  • creatine levels.
• The model accrued a total of 292 variables and performed well when assessed. It was able to identify patients who were likely to have thrombosis within 1 year at every clinical instance.
• Sensitivity was calculated to be 85%, with 98% specificity and 97% positive predictive value.
• Unfortunately, a model with 300 variables is feasible. As a result over 300,000 models were trained with the top 3–9 variables selected.
• There was high consistency among the variables selected, which included:
  • age;
  • time since diagnosis;
  • blood type;
  • JAK2 percentage; and
  • body mass index.
• When analyzed, several of the trained models demonstrated the feasibility of a 1-year 1:1 randomized clinical trial investigating the reduction of thrombosis by 50% in 200 patients, with all models indicating an efficient accrual period of 2 years.

Conclusion

A precise prediction model is essential for patient selection in clinical trials, requiring high sensitivity and positive predictive value. Leveraging machine learning enables individualized predictions and the identification of unknown patterns in large datasets. The proposed model should undergo validation using prospective and external data sets.