Non-invasive MDS diagnostic algorithm can also identify pre-MDS and distinguishes normal controls from MDS/pre-MDS patients without bone marrow examination

Abstract Background: Myelodysplastic syndromes (MDS) are often preceded by preMDS states, idiopathic cytopenia of undetermined significance (ICUS) and clonal cytopenia of undetermined significance (CCUS). We previously developed a gradient boosted model (GBM) to diagnose/exclude MDS. Using 10 variables a GBM score was calculated [Oster 2021]. A Score <0.68 or ≥0.82 probably excludes (low score) or diagnoses (high) MDS. This model was validated in two separate studies [Schulz 2024; Oster 2025]. Here we use the model to explore additional applications: to identify pre-MDS and differentiate it from MDS and controls. Methods: Patient data from 3 centers (Tel-Aviv, Munich, Pavia) were divided into 5 groups: MDS, CCUS, ICUS, controls (anemic, non-MDS after bone marrow examination), and a group of simulated healthy controls. The simulated cohort was generated assuming normally distributed variables based on standard clinical reference ranges. Patients were simulated from ages 30 to 90 years with the blood measurements drawn from their respective normal ranges. Using the 10 variables, age, gender, hemoglobin (Hb), mean corpuscular volume (MCV), white blood cell (WBC) count, absolute neutrophil count (ANC), monocytes, platelets (PLT), glucose, creatinine, GBM scores were calculated for all individuals. Receiver Operating Characteristic (ROC) curve analysis was applied to evaluate the discriminatory ability of the GBM score between pairs of diagnostic groups. Area under the curve (AUC) values with 95% confidence intervals (CI) were calculated to assess the performance of GBM scores in distinguishing between groups. Results: 510 subjects were studied (MDS 200, CCUS 47, ICUS 62, controls 101, simulated 100). Gender distribution was similar among the groups. The mean age for each group was 73.0, 69.3, 52.0, 69.5, 61.3 years, respectively. The median Hb levels were as follows: 9.4, 10.6, 12.8, 10.0, 14.1 g/dl, respectively (p<0.0001). The GBM score (median, IQR) decreased with milder disease/condition: 0.9 (0.8, 0.9), 0.7 (0.4, 0.8), 0.7 (0.5, 0.8), 0.5 (0.3, 0.7), 0.1 (0.1, 0.3), respectively (p<0.0001). The percentage of patients with high/low scores were 68%/16% (MDS), 23%/51% (CCUS), 15%/53% (ICUS), 13%/67% (controls), 0%/97% (simulated. p=0.0005). Finally, the AUC values (95% CI) to assess the model's group discrimination were as follows: MDS vs simulated 0.968 (0.950-0.985); MDS vs controls 0.838 (0.971-0.885); MDS/pre-MDS vs controls 0.844 (0.810-0.878); MDS vs preMDS 0.791 (0.741-0.858); CCUS vs controls 0.594 (0.493-0.695); ICUS vs simulated 0.855 (0.83-0.94). Practically, with a score above 0.5, there is an 80% chance (positive predictive value, PPV) of either MDS or preMDS, and below 0.5, a 74% chance (negative predictive value, NPV) of neither. Conclusions: GBM score can differentiate: 1) MDS vs controls, 2) MDS vs simulated healthy subjects, 3) MDS and Pre-MDS vs controls, 4) MDS vs pre-MDS. The model fails to differentiate CCUS or ICUS from controls, but it differentiates well ICUS from simulated. Future work will incorporate genetics and will further improve diagnostics.