Predicting secondary myeloid neoplasms in acquired aplastic anemia using machine learning models

Patients with acquired aplastic anemia (AA) treated with immunosuppressive therapy (IST) face up to a 20% long-term risk of developing secondary myeloid neoplasms (sMNs), including acute myeloid leukemia and myelodysplastic syndromes. Although hematopoietic stem cell transplantation (HSCT) is curative and prevents sMNs, older patients and those lacking suitable donors have historically received IST as first-line therapy. Recent improvements in HSCT outcomes have expanded transplant eligibility, highlighting the need for tools to better identify patients at high risk for sMN. Validated predictive models could help guide early HSCT consideration or tailor surveillance strategies. We developed 2 binary machine learning models to predict sMN development in patients with acquired AA at clinically relevant time points: diagnosis (model 1) and 6 months after IST response (model 2). We analyzed data from 275 adult patients with AA treated at University of Texas Southwestern, Cleveland Clinic, and the Hospital of the University of Pennsylvania between 1975 and 2023. Seventy-nine clinical variables were collected, including demographics, somatic mutations, and treatment response. Neural networks were trained with leave-1-out crossvalidation. Both models achieved strong performance (area under the curve, 0.82; sensitivity, 0.82, specificity, 0.73). Shared key predictors included <i>DNMT3A</i> mutation, <i>CUX1</i> mutation, total mutation count, and age. <i>TET2</i> mutation was specific to model 1; paroxysmal nocturnal hemoglobinuria clone presence was unique to model 2. High-risk classification was significantly associated with worse overall survival (<i>P</i> < .0001). These findings support the feasibility of machine learning-based sMN risk prediction in AA. With training on larger data sets and external validation, these models may support individualized decision-making around HSCT and post-IST surveillance.