Recursive Partitioning As a Tool for Differential Diagnosis between Acquired and Inherited Bone Marrow Failure Syndromes (BMF)

Background Distinguishing inherited BMF (IBMF) and acquired aplastic anemia (aAA) at diagnosis is a major clinical challenge and essential for determining appropriate initial treatment. However, no specific test can unequivocally affirm aAA, and the diagnosis of IBMF requires time consuming genetic analyses, may have limited availability, and thus can lead to detrimental delay of treatment initiation by immunosuppressive treatment (IST) or transplantation. The objective of our study was to develop a practical scoring system to identify patients who are unlikely to have IBMF and therefore may not systematically require genetic analysis before starting treatment. Methods The model was trained on a cohort from Robert-Debré and St-Louis Hospital and was validated on a large independent cohort (all these informations are gathered in a nationwide data base; RIME, after informed consent). Patients were classified as having either aAA or IBMF based on genetic results (chromosome breakage, pathogenic inherited variants by Next Generation Sequencing (dedicated IBMF NGS panel), Pr Soulier, St-Louis Hospital) and/or their response to IST. We retrospectively recorded 33 clinical and laboratory characteristics at the time of diagnosis and 15 were considered as candidate for a diagnostic algorithm. The diagnostic algorithm was then constructed using recursive partitioning (classification and regression tree). Model sensitivity represents the ability to correctly identify aAA among true aAA, whereas specificity represents the ability to correctly diagnostic IBMF among true IBMF. The positive predictive value was the probability of true aAA among those with a positive result for aAA according to the diagnostic algorithm. We decided to integrate in the model mean age, previous history of inflammatory or auto-immune disease, post-hepatitis BMF, BMF during pregnancy, familial history of cytopenia, cytogenetics abnormality, cytogenetics failure, mean corpuscular volume, mean serum alpha-fetoprotein level and variables which differ in univariate analysis. Acute onset of BMF was defined as history or worsening of cytopenia within 1 year. Results The training set included 150 patients. Median age was 35 years (range 2.2-91), and 61.3% were male: 133 with aAA and 17 with IBMF. Three of the 15 variables of interest were selected by the model in the final algorithm: clinical morphological abnormalities, PNH clone, and acute onset of BMF. This algorithm achieved in the training cohort a sensitivity of 96.2% (IC95%: 91.4-98.8%) and a specificity of 82.4% (IC95%: 56.6-96.2) in the ability to correctly differentiate aAA from IBMF. We then applied the algorithm to the validation cohort of 489 patients with at least 2 non-missing data among the 3 variables of interest (aAA n=383 (78.3%); IBMF n= 106 (21.7%)) and obtained similar sensitivity of 94.3% (IC95: 91.4-96.4) and specificity of 85.8% (IC95: 77.7-91.9) with a positive predictive value of 96.0% (IC95: 93.5-97.8). A sensitivity analysis on the subgroup of 129 patients of the validation cohort who underwent genetic testing at initial diagnosis (NGS IBMF and/or Fanconi chromosomal breakage analysis) led to similar results. Conclusions We developed an efficient and practical scoring system that incorporates 3 routine clinical and laboratory parameters (clinical morphological abnormalities, acute onset of cytopenia and the presence of a PNH clone) to identify patients who might not require genetical analysis before starting treatment at sensitivity level of 94.3%, and positive predictive value of 96.0%.