Hyperspectral retinal imaging as a biomarker for Alzheimer’s disease

Abstract Background Currently, Alzheimer’s disease (AD) diagnosis relies on biomarkers that are either expensive, invasive or time‐consuming. The retina is easily accessible and may be used as a patient‐friendly and cost‐effective diagnostic tool. Optina Diagnostics’ Mydriatic Hyperspectral Retinal Camera (MHRC) may improve diagnostic abilities of the retina by using rich datasets and artificial intelligence. Here we aim to explore diagnostic possibilities of the MHRC by identifying image features associated with the cerebral amyloid status. Method Six cognitively healthy participants with a negative amyloid‐PET scan and twenty‐five participants with a positive amyloid‐PET scan (clinical AD n = 4, preclinical AD n = 21, MMSE ≥17) were recruited from the EMIF‐AD PreclinAD Twin60++ study and Amsterdam Dementia Cohort (Table 1). Retinal imaging was performed using the MHRC that acquires 92 retinal images in an ∼1 second exposure, in steps of 5 nm increments across a spectral range of 450‐905 nm (visible and near‐infrared) on a 31° field‐of‐view (Figure 1). Spatial‐spectral features (n = 2304) were extracted from two or three hyperspectral cubes per participant using different combinations of anatomical masks, spectral regions and texture measures. Morphological features (n = 935) related to the blood vessels (diameter, tortuosity, density and fractal dimension) were also extracted from different retinal zones. Features were assessed with a Tukey’s test for statistical significance to classify the cerebral amyloid status determined by amyloid‐PET scans. Features were considered significant if their p ‐value was below 0.05 simultaneously for both our present cohort and an independent cohort of 499 subjects. Result Explorative analysis identified thirty significant spatial‐spectral features ( p ‐value range 0.0033‐0.049) for the classification of the cerebral amyloid‐PET status. Examples of such features covering different spectral ranges and retinal anatomic regions are presented in Figure 2. In contrast, only three morphological features were identified ( p ‐value range 0.017‐0.049). Conclusion Phenotypic features extracted from hyperspectral retinal images hold promise to discriminate between amyloid‐PET positive and negative individuals, beyond morphological features available with conventional retinal imaging. More data are collected from several centers to build a classifier of features with the aim to discriminate amyloid‐PET positive from negative subjects. This would yield a patient‐friendly and non‐invasive retinal biomarker for AD.