Artificial intelligence and neurodegeneration: state of the art and explainability

BACKGROUND: Alzheimer disease (AD) and Parkinson disease (PD) are an increasing healthcare concern and growing cause of disability in our century. The advent of the artificial intelligence (AI) era and the early changes in retina and optic nerve morphology detected in neurodegenerative diseases by recent literature opened the way to the perspective of a AI-based diagnosis for both conditions. MAIN BODY: In AD a generalized decrease in macular Ganglion Cells-Inner Plexiform Layer (GC-IPL) and retinal nerve fiber layer (RNFL) and peripapillary RNFL thickness has been documented. Moreover, clinical AD is characterized by a lower choroidal thickness and higher choroidal vascularity index. AI can be predicted with around 80% accuracy based on small fundus vessels characteristics. Similarly, AI based on Optical Coherence Tomography (OCT) images showed very good performance in detecting AD, especially based on total macular volume (TVM) and GCL thickness. A combination of OCT and color fundus photographs (CFP) imaging reached 85% accuracy in detecting MCI. In PD typical retinal anomalies include a generalized thinning of the peripapillary RNFL with a degeneration pattern similar to the one of mitochondrial optic neuropathies, foveal pit widening, generalized retinal and retinal pigmented epithelium thinning and outer plexiform layer thickening. AI CFP-based detection of PD showed only 65% accuracy due to a low specificity, while multimodal imaging-based AI was the best approach, reaching a 100% sensitivity and 85% specificity. CONCLUSION: In consideration of the overall very good performance, AD and PD are likely to benefit from the support of retinal-imaging based AI for early diagnosis. To optimize the performance, especially in terms of specificity, the use of multimodal imaging and possible integration with meta-data has been demonstrated to be the best technique, especially for PD diagnosis. However, training and testing on a larger sample size and longitudinal evaluation of the chosen models will be needed before application on general population. So far, explainability of the proposed models seems robust and coherent with the corresponding central nervous system damage.