JMIR Publications

ABSTRACT

Background:

Machine learning, particularly deep learning, has proven effective in diagnosing and predicting diabetes. However, these approaches often function as black boxes, leaving physicians and patients uncertain about the underlying rationale. This lack of transparency hinders the widespread adoption of machine learning in diabetes and other healthcare domains, causing confusion and eroding trust.

Objective:

This study aims to address this issue by developing and evaluating an explainable AI platform that enables healthcare professionals to comprehend the predictions and recommendations made by AI systems for diabetes. The platform, called XAI4Diabetes, not only predicts diabetes risk but also provides easily understandable explanations for complex machine learning models and their results.

Methods:

XAI4Diabetes incorporates a multi-module explanation framework, leveraging machine learning, knowledge graphs, and ontologies. The platform consists of four modules: (1) knowledge base, (2) knowledge matching, (3) prediction, and (4) interpretation. By applying AI techniques, XAI4Diabetes predicts diabetes risk and offers insights into the prediction process and results.

Results:

A mobile application prototype was developed and evaluated through usability studies and satisfaction surveys. The evaluation study demonstrates that XAI4Diabetes significantly enhances medical professionals' understanding of (1) the diabetes prediction process, (2) the datasets used for model training, (3) the data features utilized, and (4) the importance of different features in the prediction results. Most participants reported improved comprehension and trust in AI predictions after using XAI4Diabetes. The satisfaction survey indicated a high level of overall satisfaction with the tool.

Conclusions:

This research presents XAI4Diabetes, a multi-model explainable prediction platform for diabetes care. By facilitating predictions of diabetes risk and offering interpretable insights, XAI4Diabetes empowers healthcare professionals to comprehend the AI decision-making process, fostering transparency and trust. These advancements have the potential to mitigate biases and promote the wider adoption of AI in diabetes care.