Machine learning identifies the independent role of dysplasia in the prediction of response to chemotherapy in AML.

The independent prognostic impact of specific dysplastic features in acute myeloid leukemia (AML) remains controversial and may vary between genomic subtypes. We apply a machine learning framework to dissect the relative contribution of centrally reviewed dysplastic features and oncogenetics in 190 patients with de novo AML treated in ALFA clinical trials. One hundred and thirty-five (71%) patients achieved complete response after the first induction course (CR). Dysgranulopoiesis, dyserythropoiesis and dysmegakaryopoiesis were assessable in 84%, 83% and 63% patients, respectively. Multi-lineage dysplasia was present in 27% of assessable patients. Micromegakaryocytes (q = 0.01), hypolobulated megakaryocytes (q = 0.08) and hyposegmented granulocytes (q = 0.08) were associated with higher ELN-2017 risk. Using a supervised learning algorithm, the relative importance of morphological variables (34%) for the prediction of CR was higher than demographic (5%), clinical (2%), cytogenetic (25%), molecular (29%), and treatment (5%) variables. Though dysplasias had limited predictive impact on survival, a multivariate logistic regression identified the presence of hypolobulated megakaryocytes (p = 0.014) and micromegakaryocytes (p = 0.035) as predicting lower CR rates, independently of monosomy 7 (p = 0.013), TP53 (p = 0.004), and NPM1 mutations (p = 0.025). Assessment of these specific dysmegakarypoiesis traits, for which we identify a transcriptomic signature, may thus guide treatment allocation in AML.