EXPLAINABLE ARTIFICIAL INTELLIGENCE FOR INTERPRETING MACHINE LEARNING MODELS

Abstract

Nowadays, machine learning-based models are becoming increasingly popular in the health care environment for disease prediction and creating decision support systems. Although such algorithms usually demonstrate high predictive performance, they tend to be quite black box and, therefore, non-interpretable or at least partially uninterpretable. In case when the system is not interpretable enough, clinicians may not trust it, which ultimately leads to low adoption rates in healthcare environments. This dissertation introduces an approach to Explainable Artificial Intelligence and creates an explainable machine learning-based model in the context of healthcare. A Random Forest classifier is developed using Breast Cancer Wisconsin Diagnostic dataset to predict binary outcomes– benign vs malignant tumors. To increase transparency and thus improve inter pretability, two model-agnostic methods are employed to perform explainability– SHap ley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). Both methods are incorporated into the predictive model architecture to provide interpretability while retaining the complexity of the model. The proposed model achieved a relatively high level of predictive performance, having reached 96.4% accuracy and 96.3% F1-score. SHAP analysis helps to identify global fea tures affecting the prediction and LIME provides local explanations for individual predic tions.In the experiments, the outcomes indicate that combining ensemble learning with ex plainability methods increases predictive accuracy and improves interpretability of the model, at least based on the interpretability perspective.