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

The aim of this research is to diagnose polycythaemia vera (PV) disease using different machine learning (ML) algorithms with complete blood count (CBC) parameters before further investigations such as Janus kinase 2 (JAK2), erythropoietin (EPO) and bone marrow biopsy (BMB).

The study included 1484 patients who presented to the adult haematology clinic with elevated haemoglobin. Participants were retrospectively screened for JAK2, EPO and BMB results, and patients were categorised as PV group (n=82) and non-PV (other) (n=1402). First, the synthetic minority oversampling technique (SMOTE) method was used to avoid data imbalance. Then, classification predictions were made using Random Forest, Support Vector Machine Technique, Extreme Gradient Boosting (XGBoost) and K-Nearest Neighbours algorithms according to the participants' CBC parameters of white cell count (WBC), haematocrit (HCT), haemoglobin (HGB) and platelet (PLT).

The XGBoost algorithm was found to be the most effective ML algorithm in predicting the model (area under the curve=0.99, accuracy=0.94, F1-Score=0.94). In addition, the most effective parameter in the prediction of the model was PLT with 42.4%. As a result of the t-test, there was a highly significant difference between the WBC, PLT, HGB, HCT, EPO, JAK2 and bone marrow density results of PV and other groups (p<0.001).

ML algorithms can diagnose PV with CBC parameters with high accuracy, thus emphasising the potential to reduce the dependence on costly diagnostic methods such as JAK2, EPO and BMB.