Requirements for AI education for medical professionals

The complexity and volume of data in healthcare and increased demands for personalised care means that artificial intelligence (AI) will be increasingly incorporated into medical workflows.

Currently, prominent applications of AI into in healthcare include diagnostic decision support tools, patient triage/stratification tools, targeted treatments and administrative supports.

The iToBoS project aims to develop an AI diagnostic platform for early detection of melanoma. iToBoS (Intelligent Total Body Scanner for Early Detection of Melanoma) will allow dermatologists to capture a full body image of the patient, considerably reducing the average time required to complete traditional dermascopy. The project will create and evaluate an AI cognitive assistant tool that provides risk assessments for every mole. In addition to individualising the diagnosis by incorporating all available patient data, it will offer techniques for visualising, deciphering, and interpreting AI models, overcoming the "black box" nature of current AI-enabled systems and giving dermatologists useful data for their clinical practice. In this blog, the opportunity of AI to automate certain aspects of medical diagnosis and some of the barriers for swift integration in healthcare practice, specifically medical education, are discussed 

One component of the iToBoS solution includes an AI-based lesion detection tool, which employs a statistical technique called Deep Learning, a branch of Machine Learning (ML), more specifically Convolutional Neural Network (CNNs). CNNs learn to take an input image (e.g., skin image) and assign a label (e.g., lesion class) to various components of the image (irregular borders on a mole vs. normal tissue). The CNN is first trained, by exposing it to a previously manually (by expert clinicians) labelled set of images. Through repeated exposure to this ‘training set’ the CNN gradually learns the characteristics and attributes of specific components within the images. Through this process, the model starts to understand how to classify new images that it is exposed to that it has not previously seen. Following the learning process, the model is validated through exposure to a ‘test set’. This is a set of manually labelled images that the model has not seen. The model should be able to classify this test set with a high degree of accuracy.  This is one of the most common forms of AI. In image based diagnostics, these AI enabled tools, can help detect early changes, sometimes those that can’t be seen by the human eye. Therefore offering arguably more effective and efficient diagnosis.

As AI becomes more prominent in healthcare, there is an urgent requirement to integrate computational sciences directly into medical education. The more sophisticated the field becomes, the greater the need that health care professionals are properly enabled through education to effectively apply this new knowledge to their domain. Specific shortcomings of technologies should also be taught, such as aspects like transparency, explanability and liability, so that practitioners fully understand both the benefits and the risks of emerging technologies as they are applied to the field. Recent surveys report that most dermatologists agree that AI will change their practice[1]. However, in another recent survey of medical students in Korea, only 14% stated they were familiar with ‘computing, electronics, and programming’, whilst those who reported some AI-based education, also stated the source was informal - either learning from newspapers or television.

One challenge of AI-based systems is understanding how to translate outputs of these tools into effective clinical decision making. The reliance on the users knowledge of how the system produced a certain result, the limitations of the system, and how the findings ‘fit’ within typical or traditional medical diagnosis and treatment are all massively important aspects of evolving AI-assisted medical care. A large part of medical training is memory based, consuming and retaining information and knowing how to apply this knowledge to patient care. Much less time is spent educating students or medical interns on the use of technologies or AI/ML. The scale of learning in these areas is usually determined by the ‘state-of-the-art’ status of the teaching hospital, or the research agendas of the leading consultants.

Leading experts in bioinformatics advocate that healthcare workers should receive training in digital health[2]. Given the volume and speed of technical development in this domain, it seems certain that AI will become ever more integrated into healthcare environments. Currently, the prevalence of AI or ‘health-tech’ based training is low, or non-existent. In another study of final year medical students in Ireland, it was revealed that 43.4% had not heard of the term ‘machine learning’. Irrespective, 78.6% agreed that AI/ML should form part of their training.

 As part of the iToBoS project, Trilateral Research is tasked with exploring the ethical and social impacts of integrating AI tools into dermatology. Ethical and social considerations of the iToBoS tools were  discussed with key stakeholders (dermatologists, technical developers and patient advocates)  during a virtual workshop in January 2022. Two of the themes of discussion focused on autonomy and transparency. When considering transparency of AI solutions, the ability to correctly interpret the ML outputs is a pre-requisite of a transparent clinician-patient engagement. Not just ‘why’ the tool delivered a certain result (xAI) but ‘how’ the result was generated. Understanding the ‘how’ and ‘why’, will ultimately support clinicians to make decisions on patient treatment, whilst supporting patients autonomy in their healthcare choices. Whilst clinicians don’t require the in-depth knowledge equivalent to a data scientist to approach AI applications, they will require additional skills and competencies, such as the ability to manage data, supervise AI tools, and use AI applications to make informed, and effective, decisions.

To achieve a change in the medical curriculum, many structural barriers will have to be overcome. A change can only be implemented when large amount of evidence is generated. We have not reached that stage of implementing changes for AI. A more appropriate solution may be to develop Continued Medical Education (CME) or Continuing Professional Development (CPD) programs, specifically tailored for AI. Trilateral Research are working with data scientists, clinicians, patients and patient advocacy groups to design both educational materials, and training guidelines, to compliment ongoing medical education. The overarching goal is to support the sustainable and ethical integration of AI-based medical devices into every day clinical practice – ensuring a more holistic and ultimately more effective healthcare profession for all.

[1] https://onlinelibrary.wiley.com/doi/abs/10.1111/jdv.16812

[2] https://dl.acm.org/doi/abs/10.1145/3410352.3410790