Acute myeloid leukemia (AML) is a clinical emergency requiring treatment and results in high 30-day (D30) mortality. In this study, the prediction of D30 survival was studied using a machine learning (ML) method. The total cohort consisted of 1700 survivors and 130 non-survivors at D30. Eight clinical and 42 laboratory variables were collected at the time of diagnosis by pathology. Among them, six variables were selected by a feature selection method: induction chemotherapy (CTx), hemorrhage, infection, C-reactive protein, blood urea nitrogen, and lactate dehydrogenase. Clinical and laboratory data were entered into the training model for D30 survival prediction, followed by testing. Among the tested ML algorithms, the decision tree (DT) algorithm showed higher accuracy, the highest sensitivity, and specificity values (95% CI) of 90.6% (0.918-0.951), 70.4% (0.885-0.924), and 92.1% (0.885-0.924), respectively. DT classified patients into eight specific groups with distinct features. Group 1 with CTx showed a favorable outcome with a survival rate of 97.8% (1469/1502). Group 6, with hemorrhage and the lowest fibrinogen level at diagnosis, showed the worst survival rate of 45.5% (25/55) and 20.5 days. Prediction of D30 survival among AML patients by classification of patients with DT showed distinct features that might support clinical decision-making.
Keywords: acute myeloid leukemia; classification; decision tree; early death; fibrinogen; hemorrhage; infection; machine learning.