AI-driven prognostics in pediatric bone marrow transplantation: a CAD approach with Bayesian and PSO optimization.

Abstract

Bone marrow transplantation (BMT) is a critical treatment for various hematological diseases in children, offering a potential cure and significantly improving patient outcomes. However, the complexity of matching donors and recipients and predicting post-transplant complications presents significant challenges. In this context, machine learning (ML) and artificial intelligence (AI) serve essential functions in enhancing the analytical processes associated with BMT. This study introduces a novel Computer-Aided Diagnosis (CAD) framework that analyzes critical factors such as genetic compatibility and human leukocyte antigen types for optimizing donor-recipient matches and increasing the success rates of allogeneic BMTs. The CAD framework employs Particle Swarm Optimization for efficient feature selection, seeking to determine the most significant features influencing classification accuracy. This is complemented by deploying diverse machine-learning models to guarantee strong and adaptable analytical capabilities. The Adaptive Tree of Parzen Estimators (TPE), a Bayesian optimization technique, is a key component of the proposed methodology. TPE is instrumental in navigating the complex hyperparameter space to optimize model performance, enhancing the overall effectiveness of the ML algorithms. Besides, the study investigates the impact of various scaling techniques on model performance, including L1 normalization and L2 normalization, ensuring that data preprocessing is optimized for the best possible outcomes. The Local Interpretable Model-Agnostic Explanations (LIME) framework is utilized to enhance model transparency and interpretability, bridging the gap between complex AI algorithms and clinical usability. This study uses a comprehensive dataset titled "Bone Marrow Transplant: Children, which is the analysis's foundation. The findings, validated by ANOVA and T-tests, reveal significant associations between several factors and survival status, highlighting the importance of Donorage, extcGvHD, PLTrecovery, and survival_time, among others. The optimal CAD framework employs a majority voting ensemble of seven finely-tuned machine learning algorithms, achieving remarkable performance metrics. The proposed CAD framework not only achieves high accuracy (98.07%), Balanced Accuracy (98.08%), precision (98.45%), recall (98.02%), specificity (98.14%), F1 score (98.23%), and Intersection over Union (96.53%) but also offers interpretable insights into the classification procedure, contributing significantly as a comprehensive tool for clinicians in the domain of childhood BMT.