Investigating AI Approaches for Survival Prediction in Chronic Lymphocytic Leukemia

Area of Science:

• Hematology
• Computational Biology
• Oncology

Background:

• Chronic Lymphocytic Leukemia (CLL) presents a variable clinical trajectory.
• Identifying reliable prognostic markers is crucial for patient management.
• MYC gene abnormalities are known to influence CLL outcomes.

Purpose of the Study:

• To evaluate machine learning (ML) models for predicting survival in CLL.
• To compare the predictive accuracy of Random Survival Forest (RSF), Decision Tree (DT), and Cox proportional hazards models.
• To assess model performance in both MYC-positive and general CLL patient cohorts.

Main Methods:

• Investigated three time-to-event outcomes: 10-year survival from diagnosis, 10-year survival from cytogenetic assessment, and time to first treatment.
• Employed RSF, DT, and Cox proportional hazards models for survival prediction.
• Evaluated model performance using C-index and Area Under the Curve (AUC) metrics.

Main Results:

• RSF and DT models demonstrated superior predictive accuracy compared to Cox models.
• Key predictive variables were identified using permutation importance.
• While RSF and DT offer higher accuracy, Cox models provide clearer interpretability through hazard ratios.

Conclusions:

• ML models, particularly RSF and DT, show significant promise for personalized survival predictions in CLL.
• These models are especially valuable for MYC-positive CLL cases.
• Further refinement of ML models could enhance their clinical utility and applicability.