Potential challenges in AI application in the clinical setting

Abstract

AbstractArtificial intelligence (AI) solutions have attracted much interests in ophthalmology. Whereas multiple papers have been published the translation to clinical ophthalmology is challenging. The present talk will summarize five key elements that hamper translation. 1. Applicability of AI solutions to external real world datasets. Convolutional neural networks (CNNs) may fail if they were trained on datasets that do not match the target population in terms of ethnicity, age, disease severity and/or diagnostic criteria. 2. Burden of false negatives and false positives. Although the Area under the Curve (AUC) for detection of disease is often high, the application of the CNN still results in many false positives due to the low prevalence. In addition, false security is feigned in false negatives. 3. Inadequate facilities. Whereas CNNs have often been proposed for screening, the screening facilities are frequently not in place. Establishing a screening facility is expensive and requires support from health care providers. 4. Inadequate image quality. Even the best CNN does not perform well if image quality is insufficient. Continuous training and quality control is an expensive task for achieving optimal results. 5. Human interaction studies. Little is known to which degree AI based solutions will be accepted by health care providers and by patients.