Answers are grounded in this paper and cite their sources.
Introduction Diabetes mellitus predisposes patients to bacterial, viral, and fungal infections owing to impaired immune responses and metabolic dysfunction, frequently leading to severe complications, prolonged hospitalization, and suboptimal therapeutic outcomes. Accurate risk stratification remains challenging due to the heterogeneous and high-dimensional nature of multi-omics data, which hinders effective integration across molecular layers. Methods To address this gap, we propose a graph attention network (GAT)-based multi-omics integration framework that jointly models genomic, transcriptomic, proteomic, metabolomic, and clinical variables for predicting infection severity and treatment outcomes in diabetic patients. A retrospective multicenter cohort of 2,800 diabetic patients was assembled, encompassing 18,500 genomic variants, 19,800 transcriptomic features, 4,250 proteomic features, 610 metabolites, 320 microbial taxa, and 28 clinical variables. Infection severity was categorized into four clinically defined levels: mild (n = 720), moderate (n = 940), severe (n = 760), and critical (n = 380). After preprocessing, 43,508 molecular features were reduced to 1,024 optimized biomarkers through a sequential pipeline of differential expression analysis, mutual information filtering, LASSO regression, and recursive feature elimination. A heterogeneous patient similarity graph comprising 2,800 nodes and approximately 96,400 weighted edges was constructed using cosine similarity. A three-layer graph attention network with eight attention heads per layer generated 256-dimensional node embeddings, which were subsequently utilized in a multi-task learning framework to jointly predict infection severity and therapeutic outcome. The proposed framework was evaluated via five-fold cross-validation and benchmarked against conventional machine learning models (SVM, Random Forest, XGBoost, MLP) and graph-based deep learning models (GCN, GraphSAGE). Results The GAT-based model achieved an accuracy of 98.34%, precision of 97.86%, recall of 97.63%, specificity of 98.11%, F1-score of 97.74%, Matthews correlation coefficient of 96.95%, AUC of 99.18%, and AUPRC of 98.82%, significantly outperforming all baseline models. Discussion These results demonstrate that the proposed GNN-driven multi-omics framework effectively integrates heterogeneous molecular and clinical information, offering a promising artificial intelligence-assisted precision medicine tool for personalized risk stratification, therapeutic decision-making, and improved clinical management of infections in diabetic patients.
No takes yet. Share an insight, caveat, or question.
Yuan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: