AI-based characterization of terminal erythroid differentiation as a prognostic biomarker in myelodysplastic syndromes

Abstract Flow cytometry has demonstrated significant alterations in terminal erythroid differentiation (TED) in patients with myelodysplastic syndromes (MDS), associating defective erythroid maturation with poorer prognosis. Precisely and objectively evaluating TED could represent a useful biomarker to enhance risk stratification, overcoming the limitations of subjectivity and low reproducibility inherent in conventional morphological assessment. The study aimed to establish normal TED ranges using artificial intelligence (AI) in healthy bone marrow samples, define thresholds for abnormalities in MDS, and analyze their prognostic significance. We analyzed 272 patients diagnosed with MDS between 2014 and 2024 at three Spanish hospitals, as well as 107 bone marrow samples without alterations. Samples were digitized using optical microscopy (100x) coupled to a smartphone and processed with an AI model that automatically classified (performance >95%) cells into five categories: non-erythroid, proerythroblast (PE), basophilic erythroblast (BA), polychromatic erythroblast (POLI), and orthochromatic erythroblast (ORTO). Based on these classifications, the total proportion of erythroid cells in the marrow and the ratios between consecutive maturation stages (BA/PE, POLI/BA, ORTO/POLI) were calculated. Finally, we evaluated the association between TED profiles and overall survival (OS), progression to acute myeloid leukemia (AML), transfusion dependence, and response to erythropoiesis-stimulating agents (ESA). In the control cohort without marrow involvement, 95% of samples had an erythroid proportion above 17.2%. This value was therefore used as a threshold to define TED absent (<17.2%) and TED present (≥17.2%) in patients with MDS. Among patients with TED present, two phenotypes were further distinguished based on the BA/PE maturation ratio: TED normal, when the BA/PE ratio fell within the 5th–95th percentile range observed in the control cohort; and TED abnormal, when the BA/PE ratio was outside this range, indicating an early maturation defect. As a result, patients were classified into three final groups: TED absent, TED normal, and TED abnormal. We found that patients with absent TED (n=39) were older (mean age 84.0 years), had higher baseline erythropoietin levels (328.3 mU/ml), lower hemoglobin (9.5 g/dL), and a greater number of transfusions in the 16 weeks prior to treatment (4.3 transfusions) compared to those with normal TED (n=213; mean age 81.0 years; erythropoietin 165.6 mU/ml; hemoglobin 10.4 g/dL; 2.4 transfusions). Patients classified as TED abnormal (n=20) showed intermediate characteristics (mean age 74.7 years; erythropoietin 121.7 mU/ml; hemoglobin 9.9 g/dL; 2.3 transfusions) but had notably higher bone marrow blast percentages (7.4%) and higher IPSS-R scores (5.0) compared to TED normal (blasts 3.3%; IPSS-R 3.0) and TED absent groups (blasts 5.6%; IPSS-R 4.2). Patients with absent TED had lower OS compared to those with normal TED (33 vs. 95 months; p=0.002), showing a prognosis similar to patients with abnormal TED. Moreover, absent TED was associated with faster progression to AML after adjusting for IPSS-R (HR=2.2; p=0.05), higher transfusion dependency (55.6% vs. 28.9%; p=0.001), and lower hematological response to ESA (20% vs. 66%; p=0.048). Automated TED evaluation using AI enables the identification of two high-risk phenotypes in MDS (erythroid aplasia and early maturation failure), associated with reduced survival, increased progression to AML, transfusion dependence, and poorer ESA response. These phenotypes serve as prognostic biomarkers independent of the IPSS-R and could be integrated as complementary tools to optimize clinical management and risk stratification.