Exploration of Morphological Characteristics in Diabetic Macular Ischemia Utilizing a Large Language Model–Assisted Knowledge Graph

Abstract

Purpose

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.

Design

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

Participants

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.

Methods

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.

Main Outcome Measures

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

Results

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).

Conclusions

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

Financial Disclosure(s)

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

Diabetic retinopathy (DR) often leads to severe vision loss during working ages.¹ Hyperglycemia activates several molecular mechanisms and concomitantly induces nonperfusion areas.^2,3 VEGF is expressed in ischemic retinas and promotes angiogenesis and vascular permeability in proliferative DR and diabetic macular edema (DME), respectively.^4,5 In addition to such vision-threatening DR, capillary nonperfusion in the macula is known as diabetic macular ischemia (DMI) and often results in visual disturbance.6, 7, 8, 9 Despite its clinical relevance, further research is needed to establish its diagnostic criteria and therapeutic strategies.

Clinical research has accumulated a body of knowledge regarding the morphological lesions and functional impairment in DMI.10, 11, 12, 13 OCT angiography (OCTA) allows us to measure layer-by-layer perfusion or nonperfusion metrics in the fovea and parafovea.¹⁴ It promotes better understanding of DMI, and further comparative studies between structural OCT and OCTA suggest impairments in the neurovascular unit in diabetic retinal disease.^15,16 Such lesions may result in retinal dysfunction.¹⁷ Some cases with DMI suffer from a reduction of visual acuity (VA).18, 19, 20 Several publications documented reduced levels of electrical potential in electroretinogram and impaired retinal sensitivities in Humphrey field analyzer or microperimetry.^8,12 However, integrative analyses are needed to reveal the complicated pathogenesis of DMI with respect to multiple clinical aspects.

Recent advances in generative artificial intelligence are demonstrating that collaborative utilization of a large language model (LLM) and knowledge graph (KG) improves the quality of data structure and generated responses to queries.21, 22, 23 Large language model systems generate natural language based on pretrained models with past and present documents. Although they provide appropriate responses to simple queries, vector-based reasoning does not allow either generation of complex paragraphs or multihop reasoning.²⁴ A KG is composed of vertex (nodes) and edge (linkage) and represents the structured network of entities in graphical models. An LLM system splits sentences into entities, named entities, and relations, which allows us to construct a KG from documents written in natural language.²¹ It facilitates our understanding of complex and structured data sets.²⁵ Linkage-based clustering into subgraphs allows for comparative analyses of biased communications between entities.²⁶

Computer science has become more powerful with respect to scalability in handling a massive body of knowledge, whereas the human brain continues to outperform in creation, imagination, and counterfactual reasoning. This means that human–computer interaction may provide breakthroughs in ophthalmological research. In this study, we prepared a KG regarding DMI based on documents of original research and review articles and explored its clinical characteristics within a structured body of knowledge.

Methods

Search Strategy and Selection Criteria

To construct a DMI-related KG, we employed a stepwise process consisting of 4 stages: (1) selection of related articles, (2) entity–relation extraction, (3) expert review and curation by retinal specialists, and (4) KG generation (Fig S1, available at www.ophthalmologyscience.org).

Two retinal specialists (M.Y. and T.M.) first selected original and review articles that described the clinical aspects of DMI and were published between July 2008 and March 2025. The selection process was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses 2020 statement (Fig S2, available at www.ophthalmologyscience.org). Three electronic databases—MEDLINE, Embase, and PubMed—were searched using the keywords listed in Table S1 (available at www.ophthalmologyscience.org). The exclusion criteria were as follows: (1) absence of a digital object identifier; (2) lack of a digitized version of the main text; (3) original articles with fewer than 50 cases; (4) nonpeer-reviewed publications; and (5) commentaries. We further excluded articles based on titles and abstracts if they: (1) focused exclusively on molecular or cellular mechanisms; (2) included cases other than healthy or DR eyes; (3) contained no substantial discussion of macular ischemia; or (4) conflicted with general clinical findings of DMI. During full-text screening, we excluded articles that contained data outside of healthy or DR eyes, did not evaluate macular ischemia, or lacked clinical validity.

Construction of KG

We then proceeded to construct the KG based on the body of knowledge regarding DMI. To extract entities and relations from the selected articles, we first converted their main texts into plain text files. Each file was divided into segments of approximately 200 words, from which entity–relation triplets were extracted using generative pre-trained transformer 4 (GPT-4) via the OpenAI Application Programming Interface, specifically utilizing the openai.ChatCompletion.create function.²² The application programming interface was programmatically accessed through a custom Python script (version 3.11.5). The output was stored as comma-separated values files listing each triplet (subject, object, and relation). Redundant entities were then manually merged, and standardized formatting was applied to ensure consistency: relation verbs were converted to their base forms, and British spellings were unified to American English.²⁷ Each entity was manually assigned one of 13 properties based on clinical or imaging context: clinical relevance, treatment, VA, other functional tests, fluorescein angiography (FA), FA/OCTA, OCTA, OCT/OCTA, OCT, other imaging tests, pathogenesis, disease concept, and others. Specifically, the FA/OCTA and OCT/OCTA properties were used when an entity applied to both imaging modalities.²⁸ The finalized entity–relation–property data set was then formatted into a Cypher query script and imported into a Neo4j graph database (version 5.26.4), using the Awesome Procedures on Cypher plugin and the Graph Data Science Library (version 2.13.3). The KG was visualized using either property-specific color coding or community detection results generated by the Leiden algorithm.

Analysis of KG

We employed 2 major analyses: entity frequency and community detection. We counted 2 types of frequency. First, we counted how many times each entity appeared as either a subject or an object across all extracted relationships; the sum of these appearances was defined as entity frequency. Second, the frequency of each subject–predicate–object triplet was examined and referred to as relation frequency. In the graph visualization, node size and edge (arrow) thickness represented these frequencies.

The second analysis involved community detection based on connections between entities. The entire KG was divided into subgraphs by applying the Leiden community detection algorithm implemented in the Neo4j Graph Data Science Library.²⁹ Among them, we identified 8 subgraphs with more than 100 entities. These were numbered from #1 to #8 in descending order of entity count. Each community was characterized to explore the clinical features of DMI.

Proposal of Clinical Terminologies

To explore novel clinical findings or concepts, we employed 2 approaches: a module-based approach and a comparative approach. Entities with specific properties were enriched in certain communities, which allowed us to apply a module-based approach. In particular, we focused on a community in which VA—as well as OCT—and OCTA-related terminologies appeared frequently. Neurovascular lesions detected via OCT and OCTA were concentrated in communities enriched with VA-related entities. Although OCT findings for DME are often repurposed to describe neuronal lesions in DMI, layer-by-layer assessment revealed several OCT findings specific to DMI. Among them, Disorganization of MIddle retinal Layers (DMIL) was defined as a novel OCT finding. Disorganization of the retinal inner layers (DRIL) represents structural disruption involving the inner and middle retinal layers and is typically observed in eyes with DME.¹⁷ However, in DMI, the inner retinal architecture may remain intact, while only the middle layers—the inner nuclear layer (INL) and outer plexiform layer (OPL)—are disrupted. This distinct pattern was termed DMIL in this study.

In the comparative approach, we compared 2 communities to investigate the dissociation between VA and pathogenesis. One community was enriched in VA-related entities but had relatively few pathogenesis-related terms, while the other showed the reverse. This suggested a separation between VA and pathogenesis in prior publications. We therefore integrated them and proposed a novel disease subtype with VA reduction: degenerative DMI. We selected neurodegeneration from among several pathological concepts because clinical features such as DRIL, ellipsoid zone (EZ) disruption, and ganglion cell damage are all associated with VA reduction. Accordingly, we defined degenerative DMI as the coexistence of DMI and neurodegeneration, characterized by at least one of the following findings: the ice-pick sign, corresponding to a needle-like thinning of the ganglion cell layer/inner plexiform layer on OCT images; DMIL; EZ disruption; or choroidal hypertransmission.^30,31 In addition, we noted that some eyes exhibited both DMI and center-involving DME but did not meet the criteria for degenerative DMI; these were referred to as edematous DMI in this study.³²

Validation

We confirmed the clinical validity of these new terms based on a cohort of 156 eyes from 156 patients with vision-threatening DMI, as described in our previous publications.^18,19 The study was approved by the Kyoto University Graduate School and Faculty of Medicine Ethics Committee and adhered to the tenets of the Declaration of Helsinki. Written informed consent was obtained before enrollment in the study.

Swept-source OCTA and spectral-domain OCT images were analyzed to assess capillary nonperfusion and morphological features of the neuroretina. Central 3 × 3 mm swept-source OCTA images were acquired using the Plex Elite 9000 (Carl Zeiss Meditec, Inc.). The superficial layer was segmented using the device's default settings. To avoid segmentation errors in eyes with DME, deep slab images were generated from the inner border of the INL to 70 μm above the retinal pigment epithelium, as previously described.³³ Subsequently, nonperfusion squares were semiautomatically determined to quantify and map nonperfusion areas, following established methods.¹⁸ After applying Uniform Manifold Approximation and Projection analysis to cluster DR cases into 5 groups, the Moderate, Superficial, and Severe groups were categorized as vision-threatening DMI.

Spectral-domain OCT images were obtained using raster scan and 30-degree cross-hair modes on the Spectralis OCT (Heidelberg Engineering) to measure retinal thickness and assess qualitative structural findings. Eyes were diagnosed with center-involving DME if the central subfield thickness exceeded 320 μm in males or 305 μm in females.³² The presence of the ice-pick sign, DMIL, EZ disruption, and choroidal hypertransmission were evaluated on horizontal B-scan OCT images centered on the fovea (3 mm), as previously described.^30,31

Statistics

All values were presented as medians with interquartile ranges. A P value of < 0.05 was considered statistically significant. Cohen's kappa coefficient was calculated to assess interrater agreement for qualitative findings. The Mann–Whitney U test and Fisher exact test were used for continuous and categorical variables, respectively. All statistical analyses were performed using SPSS (version 24; IBM).

Results

LLM-Assisted KG

Of the 1291 publications initially retrieved, 66 studies (51 original and 15 review articles) were enrolled for KG construction. These texts were chunked based on GPT-4, resulting in 10 074 preliminary entities and 8133 relations. Redundant entities were standardized and merged into 2408 unique entities. We then constructed the KG with community assignments (Fig 3A) and property-based color coding (Fig 3B). Several communities exhibited complex intrasubgraph relationships, and some entities connected across multiple subgraphs. Notably, entities with specific properties were preferentially enriched in certain communities. The entity frequency followed Zipf's law to some extent (Fig 4A). Additionally, simplified KGs consisting of the top 10 or top 20 frequent entities were generated. The top 10 entities included disease concepts (DR, DME, DMI), imaging modalities (FA, OCT, OCTA), and VA (Fig 4B). Expanding to the top 20 revealed additional entities related to retinal function, morphological abnormalities, and pathogenesis (Fig S5, available at www.ophthalmologyscience.org). These graphical models revealed several strong connections between entities in the previous publications.

Figure 3.

The knowledge graph constructed from 66 publications on diabetic macular ischemia. Nodes and gray arrows represent entities and relations, respectively. Nodes are colored according to their assigned community (A) or property (B). FA = fluorescein angiography; OCTA = OCT angiography; VA = visual acuity.

Figure 4.

Structured relationships among frequently used entities in previous publications on DMI. The entity frequency distribution is shown in descending order (A). Graphical models of the top 10 entities are presented in (B). Node size indicates entity frequency, and arrow thickness represents the frequency of subject–predicate–object triplets. Arrows point from subject to object. DME = diabetic macular edema; DMI = diabetic macular ischemia; DR = diabetic retinopathy; FA = fluorescein angiography; FAZ = foveal avascular zone; OCTA = OCT angiography; VA = visual acuity; VD = vessel density.

Exploration of Research Subjects

Community detection divided the entire KG into 657 subgraphs and revealed biased distributions of entities with specific properties (Table 2). Visual acuity–related entities were enriched in communities #2 (Video 1) and #7 (Video 2). Based on the module-based approach, we examined OCT- and OCTA-related entities within these communities (Tables S3 and S4, available at www.ophthalmologyscience.org). They included lesions representing inner and outer retinal damage, as well as DRIL, which has been reported mainly in DME. Disorganization of the retinal inner layers was defined as disorganization from the retinal nerve fiber layer (RNFL) to the OPL. In contrast, some eyes with DMI had well-preserved inner retinal layers (RNFL, ganglion cell layer/inner plexiform layer) and a specific loss of the boundary between the INL and OPL, which was referred to as DMIL in this study (Fig 6). We preliminarily investigated and found that eyes with DMIL had poorer logarithm of the minimum angle of resolution (logMAR) than those without it (0.301 [0.064–0.523] vs 0.000 [–0.079 to 0.111]; P < 0.001).

Table 2.

Percentage of Entities with Each Property in 8 Major Communities

Property Item	Community Number (%)								Overall
	#1	#2	#3	#4	#5	#6	#7	#8
Clinical relevance	0.3	0.8	0.0	0.0	0.6	2.6	0.9	0.0	0.6
Treatment	1.9	1.2	0.9	0.0	1.9	7.1	4.3	1.8	3.6
VA	0.0	3.3	1.9	0.6	0.0	1.3	1.7	0.0	0.9
Other functional tests	0.9	5.7	4.7	1.3	3.2	7.1	0.9	1.8	3.5
FA	3.1	0.4	2.0	1.3	0.6	1.9	1.7	0.0	1.6
FA/OCTA	4.0	6.2	5.0	7.0	5.1	9.7	8.5	1.8	5.9
OCTA	17.0	11.6	11.5	4.5	13.5	9.1	11.1	2.7	11.2
OCT/OCTA	4.0	0.8	1.0	1.3	0.0	0.0	1.7	2.7	1.6
OCT	5.6	6.6	5.0	26.1	1.9	1.9	6.8	0.0	6.8
Other imaging tests	6.5	2.1	2.0	4.5	7.7	1.3	9.4	3.6	4.5
Pathogenesis	11.7	4.6	22.0	8.9	14.1	11.0	8.5	51.8	14.6
Disease concept	1.9	7.1	3.5	3.2	3.2	3.2	1.7	0.9	3.3
Others	42.6	44.4	36.5	41.4	48.1	43.5	42.7	33.0	41.9

FA = fluorescein angiography; OCTA = OCT angiography; VA = visual acuity.

Figure 6.

Disorganization of MIddle retinal Layers in representative cases of DMI with and without retinal edema. A–C, A representative case without DMIL. The boundary between the INL and OPL (arrowheads) is clearly defined. D–F, A representative case of DMIL with retinal edema. F, Cystoid spaces extending from the INL to the OPL result in DMIL (double-headed arrow). G–I, A representative case of DMIL without retinal edema. I, The OPL is poorly delineated, leading to DMIL (arrow and double-headed arrow); 3 × 3 mm OCT angiography images from the superficial (A, D, G) and deep (B, E, H) layers are shown. The green dashed lines indicate the positions of the corresponding horizontal OCT images (C, F, I). DMI = diabetic macular ischemia; DMIL = Disorganization of MIddle retinal Layers; INL = inner nuclear layer; OPL = outer plexiform layer.

We employed the comparative approach between communities #2 and #8 (Video 3) and found dissociation between VA and pathogenesis in previous publications. Pathology-related entities in community #8 were classified into molecular and cellular mechanisms, histological lesions, and pathological concepts, which included ischemia, vascular permeability, and neurodegeneration (Tables S5, available at www.ophthalmologyscience.org). Clinical findings encouraged us to combine VA reduction with pathological concepts rather than molecular mechanisms or histological lesions. In particular, retinal neurodegeneration might result from retinal ischemia and contribute to VA reduction. We therefore proposed degenerative DMI as a novel subtype, which represents a new connection between VA and pathogenesis in eyes with DMI. Additionally, edematous DMI was defined as DMI characterized by both vascular permeability and VA reduction.

Clinical Relevance of Degenerative DMI

To validate the clinical relevance of degenerative DMI, we analyzed 156 eyes from 156 patients, with their characteristics displayed in Table S6 (available at www.ophthalmologyscience.org). Four distinct findings were assessed on horizontal OCT images: the ice-pick sign (κ = 0.974), EZ disruption (κ = 0.981), choroidal hypertransmission (κ = 0.987), and DMIL (κ = 0.981). Univariate analyses revealed that eyes with degenerative DMI had poorer logMAR and greater amounts of nonperfusion squares than eyes without it (Table 7). In eyes with edematous DMI, logMAR was poorer, whereas nonperfusion square counts did not increase (Table S8, available at www.ophthalmologyscience.org).

Table 7.

Comparison of Clinical Parameters between Eyes with and without Degenerative DMI

Variables	Degenerative DMI	Nondegenerative DMI	P Value
Eyes/patients	61/61	95/95
Age (yrs)	60 (50–70)	64 (53–71)	0.214
Sex (male/female)	44/17	56/39	0.089
Hemoglobin A1c (%)	7.4 (6.8–8.4)	7.4 (6.7–8.6)	0.407
Duration of diabetes (yrs)	13 (9–20)	18 (10–23)	0.033
Systemic hypertension (present/absent)	42/19	56/39	0.238
Dyslipidemia (present/absent)	23/38	44/51	0.323
LogMAR VA	0.155 (0.046–0.398)	–0.079 (–0.079–0.046)	<0.001
Phakia/pseudophakia	31/30	61/34	0.133
International DR severity grade (eyes)
Mild NPDR	0	7	0.011
Moderate NPDR	21	46
Severe NPDR	3	7
PDR	37	35
DME (present/absent)	25/36	20/75	0.011
CST (μm)	285 (245–376)	282 (263–306)	0.162
Ice-pick sign (present/absent)	26/35	0/95	<0.001
DMIL (present/absent)	28/33	0/95	<0.001
EZ disruption (present/absent)	43/18	0/95	<0.001
Choroidal hypertransmission (present/absent)	13/48	0/95	<0.001
NPS counts in the superficial layer	420 (335–635)	295 (223–384)	<0.001
NPS counts in the deep layer	1119 (856–1415)	866 (724–1065)	<0.001
NPS counts in both layers	1562 (1251–1907)	1149 (982–1446)	<0.001
Prior PRP (present/absent)	43/18	35/60	<0.001
Prior STTA (present/absent)	8/53	4/91	0.062
Prior anti-VEGF injection (present/absent)	7/54	6/89	0.374
Prior vitrectomy (present/absent)	13/48	7/88	0.014

CST = central subfield thickness; DME = diabetic macular edema; DMI = diabetic macular ischemia; DMIL = disorganization of middle retinal layers; DR = diabetic retinopathy; EZ = ellipsoid zone; logMAR VA = logarithm of the minimum angle of resolution visual acuity; NPDR = nonproliferative diabetic retinopathy; NPS = nonperfusion square; PDR = proliferative diabetic retinopathy; PRP = panretinal photocoagulation; STTA = subTenon's injection of triamcinolone acetonide.

Discussion

In this study, we constructed a KG for DMI for the first time with support from an LLM. After selecting relevant original and review articles on DMI, GPT-4 was used to automatically divide their main texts into entities and relations. This simple approach to KG construction enabled us to objectively generate and visualize a structured and complex body of knowledge about DMI. Furthermore, community detection and subsequent expert interpretation led to the proposal of a novel OCT finding, DMIL, and a novel clinical concept, degenerative DMI. Validation analyses suggest that combined vascular obstruction and neurodegeneration may contribute to VA loss in degenerative DMI.

We introduced a novel method to review and summarize medical literature using LLM-assisted KG construction. Traditionally, review articles are compiled through expert-driven efforts following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines.³⁴ In contrast, our approach allows nonspecialists to produce objective, structured, and large-scale reviews. Notably, we did not assess risk of bias, individual study results, or certainty of evidence. In particular, tables and figures—which often contain critical data—were not included in our analysis. Future studies should develop methods for quality assurance and the construction of multimodal KGs.³⁵

We employed GPT-4 to objectively and automatically extract entity–relation pairs. However, these extractions were not always accurate, and redundant entities were not consistently merged, likely due to insufficient ophthalmology-specific knowledge within the model. Traditionally, KGs are created manually by selecting specific components. In contrast, our KG was generated by computer but proofread by retinal specialists. In the future, domain-specific LLMs trained on ophthalmology corpora may improve accuracy.^36,37 Notably, we did not verify whether each entity–relation–entity triplet was correct, to avoid assessment bias. This implies that not all relations in our KG are necessarily valid, and appropriate validation methods must be established.

Knowledge graphs enable visualization of complex, structured information and allow for multihop reasoning. In contrast, clinicians gain high-level insights through natural-language review articles. Knowledge graphs offer scalability and, with adequate indexing and citation methods, may enhance reasoning over complex mechanisms. However, natural-language reviews provide metacognitive insights across paragraphs and the entire text, and methods to extract such higher-level understanding from KGs remain to be developed. In addition, we did not obtain meaningful results regarding specific imaging modalities, such as ultra-widefield FA and adaptive optics imaging, because they were rarely discussed in the included articles. Future studies should aim to construct a KG focusing on these modalities to explore novel insights.

We utilized community detection and subsequent module-based or comparative approaches to identify novel clinical terminologies. We defined properties based on 4 broad clinical categories: disease concept, pathogenesis, fundus imaging, and visual function. This framework enabled us to propose 2 new VA-associated terms. Alternative property classification systems or existing ontologies, such as Systematized Nomenclature of Medicine-Clinical Terms or Human Phenotype Ontology, may offer other perspectives. For molecular or cellular mechanisms, ontologies such as Gene Ontology may be applicable.38, 39, 40

We defined DMIL for the first time in this study, in comparison to the well-known OCT finding of DRIL and disorganization of the outer retinal layers.^17,41 We also observed INL thinning in some DMI eyes, which we speculate reflects bipolar cell damage. In old retinal artery occlusion, the boundary between INL and OPL is typically visible, although both layers are thin.⁴² In contrast, in severe DME, cystoid spaces often span from INL to OPL.³⁰ It remains unclear whether DMIL may persist after resolution of such cystoid changes.

We also proposed degenerative DMI as a new disease concept, based on KG-derived community structure. Eyes with both macular ischemia and neurodegeneration showed reduced VA. Neurovascular crosstalk—either vessel-to-neuron or neuron-to-vessel degeneration—may underlie this pathogenesis.^15,16 Future studies should clarify the underlying mechanisms and develop new therapeutic strategies.^1,43 On the other hand, edematous DMI was associated with reduced VA without increased capillary nonperfusion. Clinicians should investigate whether anti-VEGF therapies improve VA in such cases.⁴⁴

Several limitations exist in this preliminary study. Some subjectivity in exclusion criteria during article selection may have introduced selection bias. Despite their importance, tables and figures were not incorporated into the KG. The general-purpose LLM used may not have been optimized for ophthalmologic content; domain-specific multimodal foundation models may enhance accuracy. The workflow included both automated and manual processes—automated text extraction by GPT-4 and KG construction in Neo4j, and manual proofreading and property assignment by retinal specialists—introducing potential variability. While entities were labeled with basic properties, richer ontological frameworks could support more nuanced, multilayered knowledge. We employed the Leiden algorithm alone for community detection, which may also have introduced bias. Clinical validation relied on previously collected data sets, which may include biases in patient selection, data acquisition, or outcome assessment. We employed univariate analyses but not multivariate ones, which limits our ability to draw definitive conclusions about statistical differences. Further studies are needed to assess generalizability.

In conclusion, we proposed a novel approach to synthesize and explore a scalable body of knowledge on DMI, leading to the definition of novel clinical terminologies, DMIL and degenerative DMI. This method of human–computer interaction may accelerate research progress in ophthalmology.

Manuscript no. XOPS-D-25-00497.