Dual-Modality Long-Context Learning for Clinical Diagnosis: A Hybrid Clinical-LLM and Graph Neural Network Framework with Optimized Feature Fusion

Abstract

The proliferation of electronic health records (EHRs) has generated extensive volumes of free-text clinical documentation, presenting considerable obstacles for automated diagnostic processes and clinical decision-making. This research introduces a bimodal, extended-context learning architecture that integrates large language models (LLMs) and graph neural networks (GNNs) to facilitate comprehensive and transparent clinical diagnosis classification. The implemented methodology comprises three consecutive stages: (1) dataset preparation and attribute refinement utilizing the Artificial Bee Colony (ABC) metaheuristic optimization algorithm; (2) extended-context transformer assessment (Bio_ClinicalBERT, Longformer) employing a two-branch design; and (3) integrative merging with a Gated Graph Neural Network (GGNN). The dataset, comprised of anonymized clinical documents (> 8,000 instances) categorized across Asthma, Hypertension, and Other diagnostic labels, was divided at a 70/15/15 proportion. Graph representations were generated from clinical entity co-occurrence relationships using Pointwise Mutual Information (PMI) scoring, and the integrative architecture underwent holistic optimization through ABC fitness evaluation. The findings validate the effectiveness of the multi-phase strategy: the ABC metaheuristic achieved a 40% attribute dimensionality decrease while preserving classification stability. The resultant integrative model exhibited a noteworthy enhancement compared to the optimal LLM benchmark, attaining a 4.0% Macro-F1 improvement (with a maximum Macro-F1 reaching 82.4%) and a 3.8% Area Under the Receiver Operating Characteristic Curve (AUROC) increase. Importantly, this performance gain was determined to be statistically significant (p < 0.01). This framework establishes a scalable, efficient, and interpretable approach for clinical text comprehension by effectively combining long-range textual analysis with relational graph reasoning.