The Extreme Gradient Boosting algorithm predicted polycythemia vera with 94 percent accuracy using only routine complete blood count parameters, according to a recent study.
The researchers investigated whether polycythemia vera (PV) could be predicted using routine complete blood count (CBC) parameters and machine learning (ML) methods before conducting diagnostic tests such as Janus kinase 2 mutation analysis, erythropoietin (EPO) measurement, or bone marrow biopsy. The retrospective study, published in the Journal of Clinical Pathology, included 1,484 adults presenting with elevated hemoglobin to a hematology clinic between January 2010 and August 2021. Of these, 82 were diagnosed with PV and 1,402 were classified as non-PV. CBC parameters—hemoglobin, hematocrit, white blood cell count, and platelet count—were analyzed using four ML algorithms: Random Forest, Support Vector Machine, Extreme Gradient Boosting, and K-Nearest Neighbours.
Extreme Gradient Boosting achieved the highest predictive performance, with an area under the curve of 0.99, accuracy of 94 percent, and F1-score of 0.94. The analysis indicated platelet count as the most influential variable (42 percent), followed by hematocrit (27 percent), white blood cell count (19 percent), and hemoglobin (12 percent). Statistically significant differences (p<0.001) were observed between the PV and non-PV groups for all CBC parameters and EPO levels. Median EPO was 1.77 U/L in the PV group and 9.87 U/L in the non-PV group. Janus kinase 2 positivity was identified in 98 percent of PV cases, and 76 percent of PV patients had bone marrow biopsy results consistent with myeloproliferative neoplasm, compared with 9 percent in the non-PV group (p<0.001).
The study was conducted at a single center and used retrospective data. Clinical and biochemical variables beyond the selected CBC parameters were not included, and information on potential confounding factors such as smoking status and comorbidities was incomplete. The authors state that these limitations may affect generalizability and recommend further research with larger, multicenter datasets that incorporate additional clinical variables.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
