A Graph Attention Network-Based Multimodal Auxiliary Intelligent Grading Model for Uterine Prolapse Severity.

introduction and hypothesisUterine prolapse affects women's quality of life. Traditional diagnosis relies on subjective experience with limited accuracy (ACC). Fusing multimodal data (clinical features of uterine prolapse and pelvic floor magnetic resonance imaging [MRI]) is anticipated to improve grading ACC, but existing methods generally lack effective cross-modal fusion strategies. This study hypothesizes that the proposed Multimodal Aggregation Graph Attention Network (MAGANet) achieves objective and efficient uterine prolapse grading auxiliary diagnosis via multimodal data fusion.

methodsClinical features and pelvic floor MRI data of 564 patients (March 2019 to April 2023) were retrospectively collected. MAGANet consists of modules for clinical feature data extraction, fused MRI feature extraction, sequential MRI feature extraction, graph representation construction, graph-representation fusion, and model output. It extracts deep features from structured clinical data, fused MRI, and MRI sequences to build a multimodal shared feature space and graph structure, which are fused via a Multi-level Gated Graph Attention Network for grading results. Key metrics were ACC, precision, macro-averaged F1 Score (macro-F1), kappa, area under the curve (AUC), sensitivity, and specificity.

resultsOn the independent test set, MAGANet outperformed single-modal models and various fusion methods, achieving ACC 0.935, precision 0.868, macro-F1 0.908, kappa 0.803, AUC 0.980, sensitivity 0.962, and specificity 0.650.

conclusionsMAGANet effectively integrates multimodal data, yielding excellent grading performance. It provides an objective, efficient clinical diagnostic tool and new insights for multimodal intelligent diagnosis in this field.