A multimodal EHR-based phenotyping framework integrating consensus clustering and transformer-based clinical NLP: application to autoimmune gastritis

Objective To develop and evaluate a multimodal electronic health record (EHR)-based phenotyping pipeline integrating structured and unstructured clinical data to identify disease subgroups and characterize longitudinal trajectories in a real-world setting. Materials and methods We conducted a retrospective multicenter study including 1,598 patients with autoimmune gastritis. Structured demographic and clinical variables were combined with longitudinal endoscopic and histological data extracted from routine care. A consensus clustering strategy integrating partitioning (K-medoids) and hierarchical approaches was applied to identify robust patient subgroups. Free-text endoscopic reports were processed using a fine-tuned transformer-based natural language processing (NLP) model to automatically extract structured phenotypic features. To address irregular follow-up intervals, time-normalized progression indices were developed to capture both severity and temporal dynamics of disease evolution. Results After preprocessing, 607 patients were included in the analysis. The consensus clustering approach identified three clinically distinct subgroups. The NLP model demonstrated high performance in extracting endoscopic features (accuracy 90.2%, balanced accuracy 89.3%). Application of the proposed progression indices revealed significant differences in longitudinal patterns of mucosal damage across clusters (p < 0.01). Conclusion This study demonstrates the feasibility of integrating clustering techniques and transformer-based clinical NLP within a unified EHR phenotyping pipeline. The proposed approach supports scalable secondary use of structured and narrative clinical data for subgroup discovery and trajectory modeling in chronic diseases