Can polycythaemia vera disease be predicted from haematologic parameters? A machine learning-based study.

Area of Science:

• Hematology
• Computational Biology
• Medical Diagnostics

Background:

• Polycythaemia vera (PV) diagnosis traditionally relies on invasive and costly tests.
• Early and accurate diagnosis of PV is crucial for effective patient management and treatment.
• Complete blood count (CBC) parameters offer a potentially accessible dataset for diagnostic modeling.

Purpose of the Study:

• To evaluate the efficacy of various machine learning (ML) algorithms in diagnosing PV.
• To determine if CBC parameters alone can accurately predict PV before advanced testing.
• To explore ML's potential in streamlining the diagnostic pathway for PV.

Main Methods:

• Utilized data from 1484 patients, categorizing them into PV (n=82) and non-PV (n=1402) groups.
• Applied the Synthetic Minority Oversampling Technique (SMOTE) to address data imbalance.
• Trained and tested Random Forest, Support Vector Machine, Extreme Gradient Boosting (XGBoost), and K-Nearest Neighbors algorithms using CBC parameters (WBC, HCT, HGB, PLT).

Main Results:

• The XGBoost algorithm achieved the highest predictive performance (AUC=0.99, Accuracy=0.94, F1-Score=0.94).
• Platelet count (PLT) was the most significant predictor, contributing 42.4% to the model's accuracy.
• Significant differences (p<0.001) were observed between PV and non-PV groups across WBC, PLT, HGB, HCT, EPO, and JAK2 parameters.

Conclusions:

• Machine learning models, particularly XGBoost, can diagnose PV with high accuracy using readily available CBC parameters.
• This ML-driven approach shows potential to decrease dependency on expensive diagnostic methods like JAK2 mutation analysis, EPO levels, and bone marrow biopsy.
• Implementing ML in hematology diagnostics can lead to more efficient and cost-effective patient care pathways.