Disrupting clinical education: Using artificial intelligence to create training material

Abstract

In recent years, we have seen the exponential growth of research into artificial intelligence (AI) in health care.1 Most of the AI systems in development are deductive, in that they analyse large data sets to find patterns that would otherwise be impractical to recognise and program by humans. There are new uses of AI that are less well known, however, one of which is the development of generative adversarial networks (GANs). These systems are capable of creating novel data, when trained on existing data sets.2 My own interest in GANs was sparked during my intercalated degree in Management Studies. During and following this degree, I have worked on a range of AI projects and have continually focused on uses in clinical education. A GAN model consists of two AI systems working in tandem: one system learns from an existing data set in order to generate new data, whereas the other system attempts to discriminate these data and determine whether they are from the original data set or have been newly generated. An example of such a system is a GAN model used to create realistic images of human faces when the website www.thisp​erson​doesn​otexi​st.com is refreshed. In this model, the systems were trained on images of human faces, but this could easily be extended to radiographs or images of skin lesions. These systems can also generate written prose and video content. Further down the line, the systems could be used to generate simulated learning environments with in-built variation. There are, therefore, multiple uses through which the technology can disrupt conventional clinical education. The term disruptive innovation was coined by Clayton Christensen in 1995.3 It is characterised by two principles, that the innovation does not strictly conform to the existing market's priorities and that it quickly improves to the point that the market's priorities evolve in its direction. This definition is entirely appropriate for the development of GANs. They are neither demanded by the existing market nor offered by incumbent services. In fact, it is likely that the overwhelming majority of the target audience is oblivious to their potential. They are also capable of pronounced generativity and rapid enhancements to improvement, as these machine-learning systems naturally self-improve by practising data manipulation. A simple example of how these systems may be used is in radiological training. Based on the simplified career progression illustrated in Figure 1, we can infer that the limitation to clinical career progression in radiology is exposure to images. Over the course of a radiologist's early career, they will be exposed to thousands of images, which refine their judgement and decision making. The most useful learning cases, and those that are error-prone, are cases that are complicated and difficult to interpret. In theory, GANs could be used to produce synthetic images for training without data protection and confidentiality concerns. The systems would be able to generate difficult cases and speed up the radiologist's exposure to challenging images. Some early research into the use of GANs in the field of radiology has proven the ability of such networks to generate convincing X-ray images.4 Similar principles could be used for generating written patient presentations, practice exam questions and an array of other training material. Another related disruptive technology is the development of simulated learning environments. Simulations and e-learning provide an opportunity to streamline clinical education and reduce the staffing required to deliver it. These technologies are limited by the rate at which simulations can be developed and by their variety, as it takes great effort to create and refine the systems before use. GANs and simulated learning environments may work synergistically to overcome these limitations. Dynamic AI-based simulations that are able to respond, with novel data, to the strengths and weaknesses of the operator are much more useful than generic rigidly programmed simulations. The machine-learning aspect of these systems provides the added advantage that they are continually self-improving. At first glance, this may appear to be a distant prospect, but it is important to remember that when discussing AI systems we must always account for their extraordinary potential for exponential growth. Twenty-first century technological advancements have shifted clinical education away from knowledge acquisition towards experience acquisition, and further transformations in education are expected to follow. Deductive AI systems, such as those that are beginning to outperform human decision making in radiology and dermatology, threaten to automate much of an early clinician's training. Systems are actively being developed with the aim to outperform clinicians, offering lower running costs, longer working hours and higher efficiency. If these systems succeed, there will be a need to create alternative learning experiences for junior clinicians, for whom part of their experiential learning will have been automated. Innovations that are able to refine the experiential learning model will surely accelerate and enhance clinical education. I look forward to helping to bring these technologies to fruition as the field progresses.

Ähnliche Arbeiten