Evaluation of Integrated XAI Frameworks for Explaining Disease Prediction Models in Healthcare

Explainable AI was created to solve the black box problems of deep learning models. Various explainable AI algorithms can highlight the features used by a black box model in deciding. Integrated frameworks were developed with multiple machine learning capabilities including model explanation. This research aims to investigate which of the integrated frameworks is more implementable within the healthcare sector and to evaluate how well the selected frameworks can explain prediction done with a tabular, natural language processing and image dataset. We investigated OmniXAI and InterpretML frameworks and the selected explainers are LIME and SHAP. Prostate cancer, pneumonia chest x-ray and medical question and answer were used as the datasets for tabular, image and natural language processing predictions. The feature importance score was used to compare and evaluate the explanations of the algorithms. Findings show that OmniXAI supports more methods and data formats. It can implement and explain tabular, image and NLP predictions while InterpretML supports only tabular data. InterpretML had a better plot of the explanation for the tabular prediction. InterpretML is more user-friendly and easily implementable. The OmniXAI integrated framework can be used in laboratories where tests are carried out with results stored in text, image, or CSV format. As the results are fed into the models for prediction, the explainers of the integrated frameworks will give insights into the model predictions. InterpretML can be implemented in laboratories where tabular results are been generated.