AI-Assisted OCT Clinical Phenotypes of Diabetic Macular Edema: A Large Cohort Clustering Study
Edoardo Midena, Marco Lupidi, Lisa Toto, Giuseppe Covello, Daniele Veritti, Elisabetta Pilotto, Maria Vittoria Cicinelli, Rosangela Lattanzio, Michele Figus, Giulia Midena, Luca Danieli, Enrico Borrelli, Michele Reibaldi, Daniele Tognetto, Leandro Inferrera, Simone Donati, Settimio Rossi, Paolo Melillo, Paolo Lanzetta, Valentina Sarao, Giulia Gregori, Carlo Cagini, Chiara M. Eandi, Adriano Carnevali, Vincenzo Scorcia, Emilia Maggio, Grazia Pertile, Ciro Costagliola, Gilda Cennamo, Paolo Mora, Roberto dell’Omo, Marzia Affatato, Marzia Passamonti, Mariacristina Parravano, Nicola Vito Lassandro, Marco Nassisi, Francesco Viola, N Castellino, F. Cappellani, Giuseppe Giannaccare, Francesco Boscia, Maria Oliva Grassi, Donatella Musetti, Valentina Folegani, Alessandro Invernizzi, Luca Rossetti, Tommaso Bacci, Federico Ricci, Marco Lombardo, Mary Romano, Nicola Valsecchi, Michele Coppola, Fabiano Cavarzeran, Luisa Frizziero
Abstract
Purpose: To characterize, using clustering analysis, the OCT morphological and clinical phenotypes of diabetic macular edema (DME) in a very large population (>2000 DME eyes) using standardized and validated OCT-based biomarkers. Methods: A cross-sectional study was conducted on OCT scans collected from 2355 eyes of 1688 patients with DME and performed during real-world clinical practice. OCT scans were automatically analyzed by a software able to automatically quantify OCT key biomarkers: intraretinal fluid (IRF), subretinal fluid (SRF), hyperreflective retinal foci (I-HRF), and external limiting membrane (ELM) and ellipsoid zone (EZ) interruption. Clustering analysis was performed using the above-mentioned biomarkers, including the distribution of IRF across the three ETDRS rings. Results: The overall population was predominantly composed of type 2 diabetes patients (89%), with a mean diabetes duration of 15.6 ± 10.7 years and mean best corrected visual acuity (BCVA) of 63 ± 18 ETDRS letters. Multivariate clustering identified four morphological phenotypes with distinct patterns of fluid distribution associated with different I-HRF counts, SRF volume, and percentages of ELM/EZ integrity (p < 0.0001). Conclusions: This large OCT analysis identified distinct morphological subtypes of DME, confirming the clinical relevance of key imaging biomarkers. The distribution and severity of DME features differ among clusters, supporting the importance of OCT-based phenotyping in tailoring treatment strategies and understanding disease evolution.