C009 | APPLICATION OF DATA-DRIVEN MACHINE LEARNING ALGORITHMS TO GUIDE DIFFERENTIAL DIAGNOSIS BETWEEN INHERITED AND ACQUIRED BONE MARROW FAILURE SYNDROMES

The significant overlaps among clinical features of inherited (iBMF) and acquired bone marrow failures (aBMF) are often the cause of diagnostic difficulties and delay in appropriate therapeutic approaches. Molecular tests for iBMF and telomere length analysis have limited availability and prolonged turnaround times, so that diagnosis mostly relies on clinical suspicion. The development of algorithms and score cards based on simple parameters may reveal useful tools to guide differential diagnosis and indication to genetic testing, particularly in cases with subtle presentation. Through the collection of granular data from a cohort of 597 patients (542 aBMF; 55 iBMF), variables as age at diagnosis, hemoglobin (Hb), platelets (PLT) and erythropoietin (EPO) levels, presence of PNH clone and a normal myeloid NGS at diagnosis were identified as of interest for differentiating inherited and acquired forms. Consequently, they were chosen to calculate a series of multivariate logistic models to predict the probability of having an inherited form. A 10-folds cross-validation method was used to validate the models. From a 1st dataset, including 139 patients (120 aBMF, 19 iBMF), factors showing a positive association with iBMF with 96% (95% CI: 81-100) accuracy and 97% AUC, were higher Hb (OR 2.35, 95% CI: 1.08-6.60, p=0.05), PLT (OR 1.62, 95% CI: 1.24-2.53, p<0.01), and EPO levels (OR 1.03, 95% CI: 1.01-1.07, p<0.01). Conversely older age (OR 0.63, 95% CI: 0.38-0.91, p=0.03) was negatively associated with iBMF diagnosis. By widening the size of the sample and through the imputation of missing data, we obtained other two datasets of 195 (169 aBMF, 26 iBMF) and 400 patients (359 aBMF, 41 iBMF) respectively, where also the presence of PNH clone (OR 0.18, 95% CI: 0.04-0.67, p=0.01) and a normal NGS at baseline (OR 0.19, 95% CI: 0.04-0.85, p<0.01) acquired significance, showing a negative association with iBMF. Finally, considering only iBMF patients with a known genetic lesion, we developed a score card to identify patients at high probability of iBMF [AUC 92%, Kolmogorov-Smirnov (KS) 0.76 (threshold for high quality >0.45) and Gini index 0.83 (threshold for high quality >0.55)] (Table 1). Our models may be practical tools in clinical practice but will require validation on an independent cohort to confirm its feasibility and prediction accuracy.