Abstract
PURPOSE: To identify clinical phenotypes in pachychoroid disease (PD), characterise long-term progression patterns, and evaluate the effect of photodynamic therapy (PDT) using machine learning and longitudinal probabilistic modelling. METHODS: This retrospective cohort study included 973 eyes from 663 patients (mean age 54.4 ± 12.8 years; 80% male) diagnosed with PD. Eyes were classified based on macular fluid status into no history, primary, recurrent, or resolved fluid categories. Multimodal imaging data were analysed at baseline and at the final follow-up visit ( > 5 years). Phenotypic clusters were identified using cross-validated K-means clustering on imaging and clinical features. Longitudinal changes in phenotype were evaluated with Markov modelling, including the effect of photodynamic therapy (PDT) on cluster transitions. Visual acuity (VA) changes were estimated with linear mixed models. RESULTS: Four distinct clusters were defined: Cluster 1 (233 eyes) consisted of younger patients exhibiting active fluid and mild structural alterations (LogMAR 0.16); Cluster 2 (317 eyes) represented mild or resolving PD with optimal VA (LogMAR 0.05); Cluster 3 (336 eyes) featured chronic fluid accumulation, significant structural damage, and moderate visual impairment (LogMAR 0.35); Cluster 4 (87 eyes) corresponded to severe bilateral disease and worst VA (LogMAR 0.66). At baseline, Clusters 1 to 4 were distributed as follows: 22.7%, 25.5%, 39.8%, and 12.0%. Over follow-up (mean 90.2 ± 24 months), the distribution shifted to 3.7%, 32.4%, 36.6%, and 27.3%, indicating progression toward more advanced phenotypes. Cluster 1 had frequent transitions, with 53% eyes progressing to Cluster 2 and 37% to Cluster 3. Cluster 4 showed minimal transition (96% stable). Interaction indicated greater visual deterioration in more severe baseline phenotypes (p = 0.01). PDT administration did not significantly alter disease progression (p > 0.5). CONCLUSION: PD exhibits distinct, dynamically evolving phenotypes with measurable probabilistic progression over time. As this represents an exploratory, data-driven analysis, the identified clusters should be interpreted as hypothesis-generating. Significant structural changes occurred despite PDT, underscoring the need for therapies capable of modifying the underlying disease course rather than its local manifestations.