Survival Prediction of Pediatric Leukemia under Model Uncertainty

Abstract

This dissertation addresses critical challenges in survival prediction for pediatric leukemia, particularly Acute Lymphoblastic Leukemia (ALL), by introducing novel predictive models that incorporate Bayesian principles and advanced machine learning and deep learning techniques. Recognizing the complexity and heterogeneity of leukemia, our research emphasizes the need for precise and individualized predictions that factor in the recurrence and survival probability, two pivotal aspects that significantly influence treatment outcomes in children and adolescents. In the first segment, we introduce a Bayesian survival model that diverges from traditional survival analysis by integrating full Bayesian inference, providing more accurate patient-specific survival predictions that account for model uncertainty. This allows for more confident decision-making in clinical settings. The second part of our work proposes a Transformer-based deep survival model that not only predicts the time to event but also employs Shapley Additive explanations (SHAP) for model interpretability, shedding light on how clinical variables influence predictions. Further, we propose a Bayesian Transformer-based survival model that combines the feature extraction capabilities of the Transformer encoder with a Bayesian Neural Network (BNN) layer. This model outputs recurrence probabilities under model uncertainty. The outputs from the Transformer encoder are sued for K-means clustering to evaluate model performance. Our work demonstrates strong potential in survival and recurrence prediction for children with leukemia, providing robust predictive survival models for clinicians to offer efficient and effective medical care to pediatric leukemia patients.