What Lies Beneath Diabetic Macular Edema: Latent Phenotypic Clustering and Differential Treatment Responses to Intravitreal Therapies Publisher



Cicinelli MV ; Leonardo B ; Maiucci G ; Martino G ; Ziafati M ; Bousyf S ; Frizziero L ; Lattanzio R ; Midena E ; Bandello F
Source: Ophthalmology Science Published:2026

Abstract

Purpose: To identify latent phenotypic subgroups of diabetic macular edema (DME) using artificial intelligence–based OCT metrics and evaluate whether treatment responses to anti-VEGF and dexamethasone (DEX) therapies differ across these phenotypic clusters. Methods: Retrospective study including 114 eyes (82 patients) with treatment-naive DME. Quantitative OCT metrics, including intraretinal fluid (IRF) and subretinal fluid volumes, IRF % distribution within central 0–1, 1–3, and 3–6 mm, hyperreflective foci counts, and ellipsoid zone (EZ) % disruption, were analyzed before and after treatment. Main Outcome Measures: Gaussian finite mixture modeling was used to identify distinct DME subgroups. Changes in visual acuity (VA) and OCT parameters following anti-VEGF or DEX therapy were analyzed using linear and generalized linear mixed-effects models, with false discovery rate correction applied to account for multiple comparisons. Results: Three phenotypic clusters of DME were identified, each demonstrating distinct structural and functional characteristics: cluster 1 (29%, 95% confidence interval [CI]: 20.0%–38.4%), characterized by localized central IRF (mean 0.34 mm3, 32% in the 0–1 mm zone), moderate structural damage (EZ disruption: 13%), and better VA (mean logarithm of the minimum angle of resolution [LogMAR] 0.29); cluster 2 (49%, 95% CI: 39.6%–57.9%), with diffuse IRF (60% in the 3–6 mm zone), the highest IRF volume (mean: 3.33 mm3), significant structural disruption (EZ disruption: 46%), and the poorest VA (mean LogMAR: 0.63); and cluster 3 (22%, 95% CI: 13.9%–31.2%), showing intermediate fluid levels and minimal structural damage (EZ disruption: 0.5%). Anti-VEGF therapy led to the greatest VA improvement in cluster 2 (–31.5%, standard deviation: 28.6). Pairwise contrasts showed no significant VA differences between DEX and anti-VEGF in cluster 1 (–26.6%, 95% CI: –64.7 to 11.6) or in cluster 3 (–12.4%, 95% CI: –58.2 to 33.4), although the direction of effect suggested a trend toward greater improvement with DEX. In contrast, cluster 2 showed a nonsignificant difference favoring anti-VEGF (+25.0%, 95% CI: –4.6 to 54.6). For central subfield thickness, DEX achieved a significantly greater reduction than anti-VEGF in cluster 3 (–20.9%, 95% CI: –37.0 to –4.9) and was also associated with a relative increase in peripheral IRF distribution in cluster 3 (+26.7%, 95% CI: 6.5 to 46.9), supporting phenotype-dependent treatment effects. Conclusions: Latent heterogeneity in DME presentations may influence treatment responses. Artificial intelligence–derived spectral-domain OCT metrics could support tailored therapeutic approaches to optimize patient outcomes. Financial Disclosure(s): Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article. © 2025 American Academy of Ophthalmology