Exploration of Morphological Characteristics in Diabetic Macular Ischemia Utilizing a Large Language Model-Assisted...

To explore novel clinical terminologies in diabetic macular ischemia (DMI) using a large language model-assisted knowledge graph (KG) constructed from published literature and to validate the findings using clinical data.

A review incorporating KG construction and subsequent exploration of clinical terminologies, validated in an observational cohort study.

Sixty-six original and review articles on DMI published between July 2008 and March 2025 were identified through PubMed, MEDLINE, and Embase. Validation was performed using data from 156 eyes of 156 patients with vision-threatening DMI.

Using generative pre-trained transformer 4, article texts were processed into entity-relation triplets. Entities were annotated with 13 predefined clinical properties and assembled into a KG using Neo4j. Community detection via the Leiden algorithm grouped related entities into subgraphs. Interpretation of subgraphs led to 2 novel clinical concepts: Disorganization of MIddle retinal Layers (DMIL), defined as structural disruption between the inner nuclear layer and outer plexiform layer; and degenerative DMI, defined as DMI with retinal neurodegenerative findings.

Characteristics of KG-derived subgraphs and definition of novel clinical terminologies.

The final KG contained 2408 entities and 8133 relations. Simplified graphs composed of highly frequent entities revealed structured relationships among important terminologies; for example, disease concepts, imaging modalities, vascular parameters, and visual acuity (VA) in DMI. Community detection showed uneven distribution of entities between subgraphs. Interpretation of subgraphs divided by community detection led to the identification of DMIL and degenerative DMI. In the validation cohort, eyes with DMIL had significantly worse VA than those without (0.301 [0.064-0.523] vs 0.000 [-0.079 to 0.111]; P < 0.001). Degenerative DMI was also significantly associated with both greater capillary nonperfusion and VA reduction (P < 0.001 for both).

Large language model-assisted KG construction enables objective synthesis of clinical literature and facilitates the exploration of known and unknown clinical characteristics in DMI.

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.