Comprehensive Summary
This retrospective study evaluated whether routine complete blood count (CBC) parameters could be used with machine learning (ML) algorithms to predict polycythaemia vera (PV) before advanced testing. Records of 1,484 patients presenting with elevated haemoglobin were reviewed; 82 had PV confirmed by JAK2, erythropoietin (EPO), and bone marrow biopsy, and 1,402 were classified as non-PV. To address imbalance, the synthetic minority oversampling technique (SMOTE) was applied. Four ML models were tested: Random Forest, Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), and k-Nearest Neighbours (KNN). XGBoost demonstrated the best performance with AUROC 0.99, accuracy 0.94, precision 0.93, recall 0.95, and F1-score 0.94. PLT count contributed most to model predictions (42.4%), followed by HCT (26.7%), WBC (18.7%), and HGB (12%).
Outcomes and Implications
These findings suggest that ML models trained on simple CBC data can accurately distinguish PV from other causes of erythrocytosis, potentially reducing reliance on costly and invasive diagnostics such as JAK2 mutation testing, bone marrow biopsy, and EPO assays. At the bedside, this could allow clinicians to stratify which patients with erythrocytosis require confirmatory molecular testing, thereby saving resources and avoiding unnecessary procedures. Broader validation across diverse populations will be needed, but the study demonstrates that AI-enhanced CBC interpretation may provide a practical, low-cost screening tool for earlier recognition of PV.