Random Survival Forest Versus Elastic-Net Regularized Cox Regression for Survival Prediction in Acute Myeloid...

Area of Science:

• Oncology
• Biostatistics
• Machine Learning

Background:

• Risk stratification for acute myeloid leukemia (AML) patients is complex, with conventional methods showing limited prognostic accuracy.
• Machine learning (ML) offers improved risk stratification compared to traditional approaches.
• Existing time-to-event (TTE) models often lack temporal constraints in their training data.

Purpose of the Study:

• To evaluate random survival forest (RSF) and CoxNet models for predicting AML patient survival.
• To assess model performance using data from specific time points within the AML17 trial.
• To investigate survival prediction within a defined censoring window.

Main Methods:

• Trained RSF and CoxNet models using data subsets from discrete time points of the AML17 trial.
• Constrained training data to specific trial stages to prevent data leakage.
• Employed k-fold cross-validation, permutation importance, and elastic net regularization for model selection and feature reduction.
• Utilized bootstrapping for robust re-evaluation of selected models.

Main Results:

• RSF and CoxNet models demonstrated high predictive accuracy across different trial stages.
• The best models achieved concordance indices ranging from 0.63 to 0.69.
• Model performance was highest when training data was constricted to the end of induction and stage 1.

Conclusions:

• Sequential training of ML models (RSF and CoxNet) using time-constrained data significantly enhances survival prediction accuracy in AML.
• These validated models show potential for integration into digital twin systems for simulating patient outcomes.
• The study validates a robust methodology for developing accurate TTE prediction models in oncology clinical trials.