AI holds promise in diagnosing GA
Trials needed to determine value of artificial intelligence in geographic atrophy
Artificial intelligence and machine learning-based approaches hold promise for the differential diagnosis of geographic atrophy (GA) and prediction of future progression rates, according to Maximilian Pfau MD.
“We have made a lot of progress in recent years in applying algorithms to the quantification of lesions and other disease-associated features such as drusen or hyper-reflective foci in eyes with GA. However, the real problem holding us back is the lack of publicly or commercially available software solutions in order for proper validation studies to be carried out by independent research groups,” he told delegates attending the European Society of Ophthalmology (SOE) meeting in Nice, France.
Dr Pfau noted that GA is the late manifestation of age-related macular degeneration (AMD) with the highest incidence in patients over 85 years of age. It is characterised by the atrophy of the choriocapillaris, retinal pigment epithelium and photoreceptors. No treatment is currently available despite more than 45 clinical trials of various therapeutic agents having been carried out in recent years.
There are a number of challenges in designing clinical trials for GA, said Dr Pfau, such as excluding patients with mimicking disease, sub-phenotyping within GA secondary to AMD and the need to define responsive structural and functional clinical endpoints.
In terms of differential diagnosis, the DeepMind algorithm at Moorfields Eye Hospital is already capable of multiclass classification for “gross disease categories” such as both forms of AMD (i.e. non-exudative versus exudative AMD), epiretinal membrane, vitreomacular traction and central serous retinopathy, among others.
“However, no algorithm has yet been published for recognition of GA-mimicking diseases such as late-onset Stargardt’s disease,” he said.
Another issue in GA clinical trials is setting inclusion criteria to select those patients with rapidly progressing lesions.
“Patients with non-progressing lesions will not provide much information concerning treatment efficacy. However, taking account of the precise fundus autofluorescence phenotypes and criteria such as minimum atrophy size, as well as lesion size, shape and location may ultimately allow us to build models that explain up to 40% of the variability in future progression rates,” he said.
In terms of clinical endpoints, Dr Pfau said that automated quantification using widely available multimodal imaging may prove useful as a functional outcome parameter in the near future.
In a recent study, Dr Pfau and co-workers investigated the impact of retinal microstructure on cone and rod function in 43 GA patients, and applied an algorithm to predict functional impairment based on multimodal imaging findings.
“We found that local retinal sensitivity can be estimated from retinal structure seen on commonly used clinical multimodal imaging, although again the results need to be validated on an independent data set,” he concluded.
Maximilian Pfau: Maximilian.Pfau@ukbonn.de