Synergistic integration of clinical and multi-omics data for early MCI diagnosis using an attention-based graph fusion network.

BACKGROUND: Mild cognitive impairment (MCI), a precursor to Alzheimer's disease (AD), requires precise early diagnosis. Single-omics approaches often miss disease complexity, motivating integrative and interpretable solutions. NEW METHOD: We present the Attention-based Multimodal Graph Fusion Network (A-MGFN), which integrates clinical, genomic, epigenomic, and transcriptomic data via biologically curated features - Clinico-Genetic Risk Score (CGRS), Curated Epigenomic Signature (CES), and Differential Expression Signature (DES). Each modality is encoded by a modality-specific graph convolutional network to capture higher-order intra-modal interactions, and a downstream attention module adaptively weights modalities for fusion. RESULTS: On the ADNI cohort, A-MGFN achieved an AUC of 0.86 ± 0.03 and an F1-score of 0.88 ± 0.03. Ablation and attention-weight analyses confirmed multi-omics synergy, with CES providing the largest marginal performance gains. COMPARISON WITH EXISTING METHODS: A-MGFN outperformed traditional machine-learning baselines and Graph Convolutional Network (GCN) frameworks (MO-GCAN, AD-GCN), with 5-7 percentage-point gains in F1-score, attributable to attention-guided fusion rather than fixed or unified-graph schemes. CONCLUSIONS: A-MGFN offers a robust and interpretable multi-omics framework for early MCI detection and provides insights into modality contributions that may inform clinical translation. Its design is extensible to other neurodegenerative disorders (e.g., Parkinson's disease).