Recursive Partitioning to Differentiate Acquired From Inherited Bone Marrow Failure Syndromes

Distinguishing inherited bone marrow failure (IBMF) from acquired aplastic anemia (aAA) at diagnosis is a major clinical challenge and is essential for appropriate treatment and patients' counseling. Genetic testing to exclude IBMF is frequently subject to delays of several months and may not be available in all centers and could not be applied to all patients in a large number of countries. Our study aimed to develop a practical scoring system to identify patients who are unlikely to have IBMF. Patients in the training set (N = 150) were classified as having either aAA or IBMF based on genetic results and/or response to immunosuppressive therapy. We retrospectively recorded 33 clinical and laboratory characteristics at the time of diagnosis. The diagnostic algorithm was then constructed using recursive partitioning. This unbiased model handles missing values using conditional inference procedures. Model sensitivity addresses the ability to correctly predict aAA. Three of the variables of interest were selected by the model in the final algorithm: morphological abnormalities, PNH clone (≥ 0.1% on granulocytes), and acute onset of BMF (cytopenia < 1 year). In the training set, the algorithm achieved a sensitivity of 96.2% (IC95%: 91.4%-98.8%) and a specificity of 88.2% (IC95%: 63.6-98.5) in differentiating aAA from IBMF. We then applied this algorithm to a validation set of 465 patients (aAA n = 368 (79.1%); IBMF n = 97 (20.9%)) and obtained similar sensitivity of 95.7% (IC95: 93-97.5) and specificity of 91.8% (IC95%: 84.4-96.4) with a positive predictive value of 97.8% (IC95%: 95.7-99). This efficient and practical scoring system might help physicians to identify patients who do not need genetic screening, allowing prompt start of treatment. Positive predictive value may be improved in specific populations by adding telomere length, AFP and HbF values.