Artificial Intelligence in Neuro-ophthalmology

Despite the current hype in almost every area of the society and medicine, artificial intelligence (AI) is not very new. Machine learning (ML), one of the first AI methods, was described in the 50s, followed recently by an extraordinary technical development of computers and processing power leading to the current achievements of deep learning (DL). DL is the current state-of-art technique in ML, particularly well adapted for image analysis. DL techniques have allowed new algorithms to make predictions on various outcomes and diagnosis, based on « learning » (training) on large datasets, using either a supervised (labelled data) or un-supervised (unlabelled data) approach. As in other areas of DL, the input data is used for training purposes and needs to be selected according to the highest clinical diagnostic standards, because it represents the « ground truth » or the « reference standard ». All the results provided by a DL algorithm will be compared to this ground truth. After the training phase, the performance of the algorithm is tested, first via internal validation (cross-validation) and, more importantly, via external testing (in totally novel datasets). The datasets used for training, internal validation and external testing need to be distinct and should not intersect at any time. The performance of an algorithm is expressed as diagnostic accuracy, sensitivity, specificity and area under the receiving operating curve (AUC), compared to the reference standard. The current, extraordinary boom of AI requires a high need of rigorous, controlled, prospective evaluations to demonstrate the impact of AI systems in health outcomes. In response to this need, AI consensus groups have recently elaborated international guidelines regarding AI interventions, including instructions and skills required for use of AI systems, including considerations for the handling of input and output data, the human–AI interaction and analysis of error cases [1].