We first use real-world data to develop a statistical model that can predict visual acuity from specific input data from OCT measurements. We then modulate the input data to recreate clinical instances of varying levels of SRF associated with controlled amounts of co-pathologies and examine the outputs from the model.
Development of a Statistical Model
The first stage of this project utilised retrospective data described in detail in a previous publication, obtained by taking specific structural measurements from 1211 OCTs of patients with nAMD attending the macular treatment centre of Manchester Royal Eye Hospital, as well as corresponding patient visual acuity . Participants comprised all consecutive patients diagnosed with wet age-related macular degeneration (wAMD) who attended the centre for treatment, and included all stages of wAMD.
We used the data from measurements on the OCTs to develop a statistical model whereby the vision associated with any OCT scan could be estimated by calculations on these measures of key OCT structures. We only used OCT measures that could be independently varied without necessarily affecting other measures, to prevent potentially conflicting inputs. Specifically, we removed the automated macular thickness measure from the model, as this input information might conflict with various retinal thickness measures which we would be controlling for in the study.
The exact measurements of the OCT structure that we used as input values for our statistical model were measured by the vertical thickness at the centre of the fovea (confirmed by fundus image) of the following anatomical OCT components: neuro-retina, subretinal hyperreflective material (SHRM), SRF, subretinal pigment epithelial (sub-RPE) fibrotic hyperreflective material, sub-RPE fluid and choroidal thickness. Furthermore, in the absence of autofluorescence imaging, and in keeping with the other OCT measures, the level of atrophy at the centre of the fovea was represented as the vertical extent of choroidal shadow for input to the model. Finally, expert assessment of the patency of the external limiting membrane (ELM) at the fovea, profile of the retina at the fovea (concave or non-concave), number of intraretinal cysts in the central 1 mm and the age of the patient were also used as input measures to develop the model. The visual acuity (number of ETDRS letters) of the patient was used as the outcome characteristic for the model to predict.
The precise mathematical form of this statistical model is given in the supplementary information. Within the statistical model there is a parameter associated with each clinical input value described above—that parameter quantifies the impact of the clinical variable on visual acuity. Values for these parameters were determined using non-linear least squares fitting to the sample data, performed using the MATLAB (MATLAB R2020a, Optimization Toolbox R2020a, and Statistics and Machine Learning Toolbox R2020a, The MathWorks Inc., Natick, MA, USA). Confidence intervals (95%) and associated p values for the parameters were determined assuming asymptotic normality (large-sample approximation of the parameter standard errors) and are given in Table 1 .
Determining the Impact of SRF Using the Statistical Model
Having constructed the statistical model, we used it to determine the visual impact of subretinal fluid in clinical scenarios where there were co-pathologies of SHRM and atrophy but no co-existing intraretinal fluid that would necessarily require treatment. These are hypothetical scenarios that are able to be recreated using our model. A first calculation was for visual acuity when the level of subretinal fluid is set to zero, evaluated for a range of SHRM and atrophy values such that every combination from 0 to 500 µm thickness for each measure in 50 µm steps was determined. For these artificial OCT conditions, the number of intraretinal cysts was set to zero to represent the clinically required scenario of no intraretinal fluid. ELM was set to not intact and retinal profile to default concave, as these were the modal values in the data. The parameters of age and retinal thickness were set to their mean values of 80.3 years and 171.3 µm, respectively. Values for choroidal thickness, sub-RPE fluid and sub-RPE fibrosis were also set to mean values of 140.6 µm, 29.5 µm and 28.8 µm, respectively.
A second calculation then determined the level of vision in another hypothetical scenario when all the above parameters were set exactly as before except the subretinal fluid, which was set to a variety of different levels from 50 to 500 µm.
Finally, we were able to compare these two hypothetical scenarios to determine the difference in levels of vision between these two calculations. This difference represents the vision change between the two scenarios, i.e., when the varying levels of subretinal fluid are present or completely absent, for each specific combination of atrophy and subretinal hyperreflective material. This calculation produced a four-dimensional matrix containing information on the expected vision change in all combinations of a range of atrophy levels and subretinal hyperreflective material, when influenced by varying degrees of subretinal fluid.
The numerous dimensions of information were represented using a four-dimensional plot—this plot represented the visual impact of SRF in OCTs in the context of three additional dimensions of associated OCT characteristics: amount of SRF, atrophy and SHRM. In addition, a three-dimensional plot was created where the amount of SRF was kept at a constant arbitrary level of 400 µm to allow more graphical detail to be presented on the impact of changing levels of atrophy and SHRM.
Compliance with Ethics Guidelines
The project was institutionally approved as retrospective anonymised data analysis by the Research and Development office of the Manchester University NHS Foundation Trust (Approval Number: R03712) and adhered to the Declaration of Helsinki.