Prediction of overall survival in acute myeloid leukemia using supervised machine learning

Risk stratification in acute myeloid leukemia (AML) is traditionally based on cytogenetic and genetic features alone and does not incorporate patient-related clinical variables. This study evaluated whether machine learning (ML) models integrating clinical, cytogenetic, and genetic features improve prediction of 1-year and overall survival (OS) compared with guideline-based risk stratification.

We conducted a study of 785 contemporary real-world patients with AML. Patients were stratified by treatment intensity (intensive and non-intensive). Several ML models for binary 1-year survival and time-to-event OS were trained using three feature sets: clinical alone; clinical plus cytogenetics; and combined clinical, cytogenetic, and genetic. Performance was evaluated using area under the receiver operating characteristic curve (ROC-AUC) and C-index and compared with current models by European LeukemiaNet (ELN) 2022 and 2024.

The binary classification performance was similar across treatment groups with ROC-AUC between 0.5 and 0.7. Predictive performance for OS varied by treatment group, where ML models for non-intensively treated patients achieved better performance (C-index >0.75), consistently outperforming ELN2024 (C-index ∼0.57). In intensively treated patients, performance was lower and did not outperform ELN2022 although best results were obtained using combined clinical and cytogenetic features. Adding genetic features did not consistently improve model performance across settings. Models using the full feature set indicate that clinical variables have the greatest impact.

ML-based models yielded better or equal prognostic accuracy compared with ELN 2022 and 2024, highlighting the value of integrating clinical features. Limited performance in intensively treated patients suggests a need for dynamic predictors, such as treatment response measures, to enhance risk stratification further.