The AI algorithms deployed today for clinical diagnostics are termed ‘narrow’ or ‘weak’ AI. These AI algorithms are trained to perform a single task: for example, to classify images of skin lesions into diagnostic categories or to provide a molecular diagnosis from a combination of genomic and phenotypic data. These algorithms do not display general intelligence and are not flexible enough to address other clinical diagnostic tasks. However, transfer learning approaches can be used to adapt a fully trained AI algorithm to accomplish closely related tasks. This is best exemplified by image-based diagnostic AI algorithms that benefit from advances in computer vision and neural networks trained for general image recognition tasks. Thus, the first step in the design of clinical diagnostic AI algorithms usually involves mapping the specific diagnostic task to a more general problem class. Here, we review these problem classes and briefly highlight the intersection of these techniques with genomics.
