Leveraging Ensemble Machine Learning Models for the Detection of Primary Myelofibrosis in Electronic Health Records

Area of Science:

• Hematology
• Medical Informatics
• Machine Learning

Background:

• Primary myelofibrosis (PMF) is a rare hematologic malignancy often diagnosed late due to non-specific symptoms.
• Screening for PMF using electronic health records (EHRs) is difficult due to data challenges like imbalance and incomplete labeling.

Purpose of the Study:

• To develop an EHR-based screening algorithm for identifying patients at risk of PMF.
• To evaluate the effectiveness of positive-unlabeled (PU) learning for PMF screening.

Main Methods:

• Utilized EHR data from 10 Polish hospitals.
• Evaluated various ensemble models, with LightGBM demonstrating superior performance.
• Applied positive-unlabeled (PU) learning techniques.

Main Results:

• LightGBM achieved an Area Under the Precision-Recall Curve (AP) of 20.83%, with sensitivity of 45.52% and precision of 14.72%.
• The model identified high-risk patients by analyzing interactions among Red Cell Distribution Width (RDW) and Platelet (PLT) parameters.
• PU methods improved sensitivity but decreased precision, highlighting a challenge for clinical implementation.

Conclusions:

• Ensemble learning, particularly LightGBM, shows potential for identifying PMF risk in clinical settings.
• While PU learning can enhance sensitivity, its impact on precision requires careful consideration for practical healthcare applications.