Interpretable Pulmonary Disease Diagnosis with Graph Neural Network and Counterfactual Explanations

With recent fast development of artificial intelligence and related techniques, computer-aided diagnosis is increasingly emerging. During the process of automated diagnosis, the importance of explainable decision making has grown significantly. Traditional black-box models have often hindered the practical implementation of automated diagnosis algorithms. To address these challenges, we propose a novel framework that seamlessly integrates prototype learning with Graph Neural Networks (GNN) while also providing post-hoc explanations. On one hand, our model achieves intrinsic interpretability by reasoning based on the similarity calculations with prototypes for each disease. On the other hand, it employs counterfactual reasoning on graphs to pinpoint the most significant features for post-hoc explanations, supporting the diagnosis process. Building on these strengths, we integrate vision-language models to effectively harness multimodal patient information, thereby capturing a more detailed understanding of their medical condition. Our experiments on real Chinese EMRs of Pulmonary diseases demonstrate that our method not only delivers precise diagnoses but also accurately identifies medical findings substantiating the diagnoses.