Machine learning identifies inflammation-related diagnostic biomarkers for primary myelofibrosis with clinical validation.

Primary myelofibrosis (PMF) is a heterogeneous bone marrow disorder, and substantial evidence indicates the involvement of inflammatory mediators in its progression. However, a diagnostic model based on inflammation-related genes has not yet been established. The aim of this study was to identify specific inflammation-related genes (IRGs) with potential value in myelofibrosis diagnosis and risk prediction. Transcriptomic data from the Gene Expression Omnibus (GEO) database were analysed to identify inflammation-related differentially expressed genes (DEGs). Machine learning approaches, including the least absolute shrinkage and selection operator (LASSO) and random forest, were used to select hub genes. A nomogram was constructed and validated externally using independent GEO datasets and local sequencing data. Immune cell infiltration and functional enrichment were also investigated. HBEGF, TIMP1 and PSEN1 show significant differences in expression between normal individuals and those with PMF. A nomogram based on three genes was established to assess the risk of PMF in healthy individuals. The ROC curve revealed that the three hub genes have outstanding diagnostic value for this disease (AUC = 0.994; 95% CI: 0.985-1.000); the results were subsequently validated in an external validation set (AUC = 0.807; 95% CI: 0.723-0.891), and a sequencing dataset from the First Affiliated Hospital of Zhejiang University (AUC = 0.982; 95% CI: 0.841-1). Enrichment analyses implicated cancer-related and immune pathways, and the model genes correlated significantly with immune cell infiltration and function. We developed and validated a robust three gene diagnostic model for PMF based on inflammation-related genes, offering a noninvasive molecular tool with potential clinical utility for auxiliary diagnosis.