Abstract
Azacitidine is an approved therapy for higher-risk myelodysplastic syndrome (MDS). However, only 30-40% patients respond to azacitidine, and the responses may take up to six cycles to become evident. Delayed responses and the myelosuppressive effects of azacitidine make it challenging to predict which patients will benefit. This is further compounded by a lack of uniform prognostic tools to identify patients at risk of early treatment failure. Hence, we performed a retrospective analysis of 273 consecutive azacytidine-treated patients. The median overall survival was 16.25 months with only 9% alive at 5 years. By using pre-treatment variables incorporated into a random forest machine learning model, we successfully identified those patients unlikely to benefit from azacytidine upfront (7.99 vs. 22.8 months, p < 0.0001). This model also identified those who required significantly more hospitalizations and transfusion support. Notably, it accurately predicted survival outcomes, outperforming the existing prognostic scoring system. By integrating somatic mutations, we further refined the model and identified three distinct risk groups with significant differences in survival (5.6 vs. 10.5 vs. 43.5 months, p < 0.0001). These real-world findings emphasize the urgent need for personalized prediction tools tailored to hypomethylating agents, reducing unnecessary complications and resource utilization in MDS treatment.