Enhancing interpretability in healthcare models using explainable artificial intelligence

ML models have shown remarkable performance in predictions in genomic cancer applications but experience the black-box problem. Where complicated processes of decision making remain non transparent to healthcare professionals and researchers. This absence of interpretability makes hurdles to clinical adoption, approval from regulators and discovering new scientific facts in healthcare AI. the latest research of explainable artificial intelligence (XAI) shows vital gaps. Poor comparative analysis over prediction tasks and cancer types, narrow combination of multiple interpretability methods and insufficient presentation of interpretability as a quality control mechanism for pointing out failures of model. This work evaluates and develops a combined interpretability pipeline for machine learning models used in prediction of genomic cancer. Random Forest feature selection, SHAP (SHapley Additive exPlanations) analysis and UMAP (Uniform Manifold Approximation and Projection) dimensionality reduction are used in combination to give transparency in models that works at multi-scale. The pipeline was used on two childhood cancer datasets. B-cell Acute Lymphoblastic Leukaemia (ALL) with n= 100 patients and 22,277 genes to make relapse prediction. Rhabdomyosarcoma (RMS) with n=101 patients and 54,675 genes for classification of subtype (ARMS/ ERMS). The methodology integrates three main techniques. (1) Random Forest feature selection, to find out the top 100 most important features, (2) SHAP (SHapley Additive exPlanations) analysis, in order to produce rankings of global feature importance and explanations of individual level prediction and (3) UMAP projection different scales of features (100, 20 and 10 genes) to capture the patterns of clusters and optimal dimensionality. Cross dataset examination found zero gene overlapping across ALL and RMS top 100 genes that showed genomic signatures which are disease specific. Cross method validation with merging of Random Forest importance, Shapley values and cluster separation in UMAP space improved confidence in detected biomarkers. This thesis shows that interpretability give two roles, uncovering failures of model and explaining successful outcomes. This combined technique gives a structural template for making transparent, actual working of ML models in healthcare in genomic medicine. With these results that around 20 genes gives maximum balance contributing practical guidance for designing clinical biomarker panels.