Interpretable machine learning model using peripheral blood for non-invasive detection of moderate-to-severe myelofibrosis in JAK2 V617F-positive MPNs: A multicentre pilot proof-of-concept study.

Progression to moderate-to-severe myelofibrosis (MF) in JAK2 V617F-positive myeloproliferative neoplasms (MPNs) is often clinically silent, and bone marrow biopsy is invasive and unsuitable for regular monitoring. In this multicentre pilot proof-of-concept study, we aimed to develop an interpretable, non-invasive model using peripheral blood parameters to detect significant fibrosis and to explore the inflammatory mechanisms involved. A total of 336 JAK2 V617F-positive MPN patients were enrolled for model training and 92 for external validation. Candidate features (blood counts, cytokines, clinical data) were screened using least absolute shrinkage and selection operator (LASSO) and Boruta algorithms. Ten machine learning methods were compared; model discrimination, calibration and clinical benefit were evaluated. Molecular validation focused on interleukin-1β (IL-1β) in Jak2V617F knock-in mice and co-culture assays. The support vector machine model yielded the best discrimination (area under the curve [AUC] = 0.916, 95% CI 0.83-0.99), good calibration and highest clinical utility. SHapley Additive exPlanations (SHAP) interpretation identified haemoglobin, IL-1β and age as the most influential features. Preclinical experiments confirmed IL-1β upregulation with fibrosis progression, reversed by ruxolitinib or pegylated interferon-α but not hydroxyurea (hydroxycarbamide). This multicentre pilot proof-of-concept study demonstrates that an interpretable blood-based tool shows preliminary accuracy for identifying moderate-to-severe MF and underscores IL-1β as a mechanistic biomarker. The data support potential utility as a complementary approach; bone marrow biopsy remains the reference standard.