Artificial intelligence in autoimmune gastritis: Clinical phenotyping and disease progression

Summaries 1 min 16

Font Size

Created On: 30/10/2026

Autoimmune gastritis (AIG) is a chronic inflammatory condition marked by damage to the gastric mucosa, with clinical and histological features that may vary across patients and over time.

See All

Key take away

A novel multimodal EHR phenotyping framework combining consensus clustering and fine-tuned transformer-based NLP identifies three distinct autoimmune gastritis patient subgroups and precisely models longitudinal mucosal damage progression.

Background

Autoimmune gastritis (AIG) is a chronic inflammatory condition marked by damage to the gastric mucosa, with clinical and histological features that may vary across patients and over time. Recognizing distinct disease subgroups and comprehending their progression require the integration of diverse clinical information, including structured health records and narrative endoscopic reports.

However, extracting meaningful phenotypic information from these heterogeneous data sources and accounting for irregular follow-up intervals remain crucial analytical challenges. This study sought to develop and evaluate a multimodal electronic health record (EHR)-based framework integrating structured clinical data, transformer-based natural language processing (NLP), and consensus clustering to identify AIG subgroups and characterize longitudinal disease progression.

Method

This retrospective, multicenter study included 1,598 volunteers with AIG, whose structured demographic and clinical variables were integrated with longitudinal endoscopic and histological data collected during routine care. Following data preprocessing, 607 volunteers were retained for analysis. A consensus clustering strategy combining K-medoids partitioning and hierarchical clustering was used to identify stable and clinically meaningful patient subgroups.

In parallel, a fine-tuned transformer-based NLP model processed free-text endoscopic reports to automatically extract structured phenotypic features. To accommodate irregular intervals between clinical assessments, the researchers developed time-normalized progression indices to determine both disease severity and temporal changes in disease evolution.

Result

Subgroup stratification: Consensus clustering successfully stratified the 607-patient cohort into three distinct, clinically relevant patient subgroups exhibiting unique pathophysiological profiles.
NLP performance: The fine-tuned transformer model portrayed superior extraction potency across unstructured endoscopic reports, achieving:
(a) Overall accuracy: 90.2%
(b) Balanced accuracy: 89.3%
Longitudinal disease evolution: Application of the time-normalized progression indices illustrated noteworthy variations in mucosal damage trajectories across the three clusters.

Conclusion

Integrating artificial intelligence, clinical NLP, and EHR data enabled scalable phenotyping and disease trajectory analysis, offering a promising approach for studying AIG and other chronic diseases.