Interpretable Machine Learning for Course Prediction in Central Serous Chorioretinopathy Using Baseline Optical Coherence Tomography From a Multicenter Cohort
Related Experiment Video
Updated: Oct 11, 2026
Assessing Early Stage Open-Angle Glaucoma in Patients by Isolated-Check Visual Evoked Potential
Published on: May 25, 2020
Lorenzo Ferro Desideri1, Nina Eldridge2, Lieselotte E Berger3
1From the Department of Ophthalmology (L.F.D., N.E., L.E.B., M.S.Z., R.A.), Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland; Bern Photographic Reading Center (L.F.D., M.S.Z.), Inselspital, Bern University Hospital, University of Bern, Bern, Switzerland.
Purpose:
To predict disease clinical course in central serous chorioretinopathy (CSC) using an interpretable machine learning model applied to baseline artificial intelligence (AI)-derived optical coherence tomography (OCT) biomarkers acquired at first presentation.
Design:
Multicenter, retrospective cohort study.
Participants:
Overall, 329 treatment-naïve CSC eyes recruited across 8 international centers in Europe, the Middle East, Asia, and North Africa.
Methods:
Treatment-naïve CSC patients were retrospectively included and classified, based on follow-up, as acute or chronic CSC according to subretinal fluid (SRF) resolution. Baseline OCT biomarkers were extracted using an AI-based platform (Discovery OCT Biomarker Detector; RetinAI AG, Switzerland), which quantified retinal layer thicknesses, including the outer nuclear layer (ONL) and photoreceptor-retinal pigment epithelium (PR + RPE) complex, fluid volumes (SRF and RPE detachment [PED]), and biomarker probabilities, across central, parafoveal, and perifoveal ETDRS regions. These features used to develop an interpretable machine learning model predicting future CSC chronicity.
Main Outcome Measures:
Classification of eyes as acute or chronic CSC, determined by SRF resolution on follow-up; discriminative performance of the Random Forest classifier, assessed by area under the receiver operating characteristic curve (AUC); and biomarker importance, evaluated using SHapley Additive exPlanations (SHAP) values.
Results:
Of 329 eyes (291 patients), 92 (28.0%) were classified as acute and 237 (72.0%) as chronic CSC; mean follow-up was 4.2 ± 1.6 years. Chronic CSC patients were significantly older (51.7 ± 12.6 vs 43.2 ± 9.6 years, P < .001) with worse baseline visual acuity (0.25 ± 0.21 vs 0.16 ± 0.15 logMAR, P = .029). Among baseline OCT biomarkers, parafoveal ONL thickness showed the strongest association with disease chronicity (q < 0.0001; r = -0.414). The Random Forest classifier achieved an AUC of 0.852 (95% CI: 0.748-0.940). SHAP analysis identified parafoveal ONL thickness, perifoveal PR + RPE thickness, and central SRF volume as the strongest predictors of acute CSC, whereas greater RNFL thickness, PED volume, FLAT PED probability, and HF probability were the most influential predictors of chronic disease.
Conclusions:
Automatically detected OCT biomarkers can predict CSC chronicity at initial presentation with good discriminatory performance. Quantitative markers of outer retinal integrity, fluid configuration, and RPE dysfunction may improve prognostication, facilitate patient counseling, and help identify individuals at increased risk of persistent CSC.
