A machine learning and explainable artificial intelligence approach for predicting the efficacy of hematopoietic stem cell transplant in pediatric patients

Abstract

Cancer is a fatal disease that affects people of all ages, including children. It is one of the leading causes of death worldwide. According to World Health Organization, an estimated 400,000 children develop cancer yearly. Bone marrow transplantation (BMT) is a specialized treatment for patients suffering from certain types of cancer, such as myeloma, lymphoma, leukemia, and others. It usually includes extracting healthy cells from the donor’s bone marrow and replacing the existing ones in the patient’s body. However, the treatment can also cause complications such as graft-versus-host disease, organ damage, stem cell failure, new cancers, and infections. In this study, we use machine learning and explainable artificial intelligence (XAI) techniques to predict the survival rate of children undergoing Hematopoietic Stem Cell Transplants. Three feature selection techniques have been utilized for feature selection: Harris Hawks optimization, salp swarm optimization, and mutual information. The final custom stacked model delivered optimal results with accuracy, precision (89%), recall (88%), f1-score (88%), area under curve (AUC) (92%), and average precision (86%). In addition, XAI techniques such as Shapley additive values (SHAP), local interpretable model-agnostic explanations (LIME), ELI5, and QLattice have been used to make the models more precise, understandable, and interpretable. According to XAI, the most important features were relapse, donor age, recipient age, and platelet recovery time. The promising results point to the potential use of artificial intelligence in understanding the effectiveness of bone marrow transplants in children.

Keywords

• Machine learning
• Explainable Artificial Intelligence
• Bone marrow transplant
• Cancer
• Harris Hawk Optimization
• Salp Swarm Optimization