An Explainable AI Platform to Help Healthcare Professionals of Diabetes Understand Predictions Made by Machine-learning Approaches (Preprint)

Abstract

Machine learning, especially deep learning has been used for diagnosing and predicting diabetes. Machine learning-based approaches achieve good prediction accuracy and precision. However, machine learning approaches normally work as a black box and the rationale underneath is unknown to physicians and patients, which may cause confusion and distrust issues. This issue hinders machine learning’s application in diabetes and other healthcare practice.

This study aims to help healthcare professionals to understand how AI makes predictions and recommendations for diabetes. Specifically, we design, develop, and evaluate an explainable AI platform that can not only predict diabetes risk but also provide human-comprehensible explanations for complex, black-box machine learning models and prediction results.

An explainable AI framework, XAI4Diabetes, is designed with a multi-module explanation framework based on technologies of machine learning, knowledge graphs, and ontologies. XAI4Diabetes consists of four modules: (1) the knowledge base module, (2) the knowledge matching module, (3) the prediction module, and (4) the interpretation module. XAI4Diabetes applies AI techniques to predict diabetes risk and interpret the prediction process and results. A mobile application prototype system was developed. The application was evaluated by usability study and satisfaction surveys.

Results of the evaluation study demonstrate that medical professionals agreed that XAI4Diabetes helps them understand (1) how machine learning makes the diabetes prediction, (2) what datasets were used for the machine learning model, (3) data features used in the dataset, and (4) the prediction results in terms of feature importance. Most participating medical professionals acknowledge that XAI4Diabetes helps them better understand and trust predictions made by AI systems. The satisfaction survey shows that participants were satisfied with the tool in general.

In this research, we designed, developed, and evaluated a multi-model explainable perdition model, XAI4Diabetes, for diabetes care. XAI4Diabetes provides an easy-to-use interface to predict a patient’s risk of diabetes and explain the prediction process and the results. The experimental results show that the prototype mobile application system can help healthcare professionals’ understanding of the AI decision-making process, thus improving transparency and trust. This would potentially mitigate various kinds of bias and promote the application of AI in diabetes care.