CIAE: a consistency- and informativeness-aware explanation framework for improving type 2 diabetes prediction and mechanistic insights

The gut microbiome plays a significant role in the development of type 2 diabetes (T2D), yet its underlying mechanisms remain unclear. Explainable artificial intelligence offers a promising approach to elucidating machine learning models and enhancing our understanding of these mechanisms. However, identifying the most effective explainer for specific problems remains a significant challenge. To address this issue, we propose the consistency- and informativeness-aware explanation framework (CIAE), which integrates multiple explainers to improve the consistency and informativeness of explanations. Using the top 100 Kyoto Encyclopedia of Genes and Genomes (KEGG) Orthology (KO) features identified by CIAE, models achieved robust results on individual datasets (area under the curve (AUC): 0.870 and 0.808) and demonstrated strong cross-dataset generalization (AUC: 0.822 and 0.778). Notably, K00688 and K25026 were selected as key KOs in both datasets, and they are involved in the regulation of the starch and sucrose metabolism pathway and blood glucose homeostasis, which are closely related to the occurrence of diabetes.