Defining the Limits of Pre-Transplant Risk Prediction in AML: Evidence From Machine Learning and Regression Models.

Area of Science:

• Hematology
• Oncology
• Medical Informatics

Background:

• Allogeneic hematopoietic cell transplantation (allo-HCT) for acute myeloid leukemia (AML) presents significant morbidity and mortality risks.
• Machine learning (ML) is increasingly utilized for predicting medical outcomes.

Purpose of the Study:

• To assess the utility of ML in predicting overall survival (OS) post-allo-HCT for AML.
• To compare the performance of ML models against traditional Cox regression.

Main Methods:

• Developed and compared three models: Cox-TVC, Cox-EN, and Random Survival Forest (RSF) using 2,253 internal patient data and 14 pre-transplant variables.
• Evaluated model performance using C-index, net reclassification improvement (NRI), and decision curve analysis (DCA).
• Performed external validation on 252 patients with uniform measurable residual disease (MRD) assessment.

Main Results:

• ML models significantly improved risk stratification compared to HCT-CI and aELN (NRI: 31-44%, p <0.001).
• Key predictors included age ≥ 60 years, MRD positivity, and adverse aELN risk.
• Model discrimination was modest but higher in the external cohort (0.69-0.71) than the internal cohort (0.60-0.61).

Conclusions:

• ML approaches offer superior risk stratification for AML allo-HCT compared to HCT-CI and aELN, performing comparably to Cox models.
• Static pre-transplant models show modest individual outcome prediction.
• Advancements require MRD standardization, richer datasets, and dynamic peri- and post-transplant modeling.