A scoping review of artificial intelligence in medical education: BEME Guide No. 84

Figures & data

Figure 1. Prisma flow diagram.

Figure 2. Infographic summarising results.

Table 1. AI terminology.

Artificial Intelligence (AI)	AI is a concept, the most general of the terms, that spans any process that involves a machine acting ‘intelligent.’ Intelligence is most often defined as ‘human-like’ in its ability to make decisions, learn from mistakes, generate insights, or understand language (Coppin 2004).
Machine Learning (ML)	ML is a method and discipline. ML involves the specific mathematical and computational structures which produce computer programs/algorithms that can make decisions given input data. ML is most frequently the way in which we achieve (a semblance of) AI (Gopinath and Churiwala 2019).
Traditional Machine Learning	Traditional ML involves extensive human-engineered preprocessing of input data to extract salient ‘features’ which are then piped into a mathematical model/equation to produce decisions/results. Examples: logistic regression, support vector machines, decision trees, K-nearest neighbors, gradient boosted machines (Gopinath and Churiwala 2019).
Data Mining (DM)	DM often borrows basic statistics or algorithms from traditional ML (for example, logistic regression or k-nearest neighbors), but instead of trying to teach a machine to make a decision, it instead attempts to sort through large volumes of data to allow humans to make insights and better understand the data itself. Whereas ML is concerned with assessing the ability of machines to make prospective decisions on heretofore-unseen data, DM instead tries to glean insights from full, known sets of previously collected data (Islam et al. 2018).
Deep Learning (DL)	Deep ML involves relatively little human-engineered preprocessing of input data, which is often input in raw or near-raw form. In DL, the model itself creates its own internal representations of features, as represented by the connection strength between mathematical ‘neurons’ within the artificial neural network. Deep neural networks require significant time/ energy/ money/ data investments to initially train. However, these ‘pre-trained’ models can be ‘customized’ to new domains, such as MedEd, cheaply (Naylor 2018).
Natural Language Processing (NLP)	NLP is any type of computational or mathematical approach that deals with natural human (written or spoken) language. NLP is almost always paired with one of the approaches above, which is often why you see it written with a slash (e.g. NLP/ML). ChatGPT, for example, is NLP/ML (specifically deep learning), because it is a deep machine-learning artificial neural network which processes natural language (Iroju and Olaleke 2015).

Figure 3. Ethical considerations for AI in medical education.

Figure 4. FACETS framework to report future AI innovations in medical education.

Coppin B. 2004. Artificial intelligence illuminated. Sudbury (MA): Jones & Bartlett Publishers.

 Google Scholar

Gopinath RA, Churiwala S. 2019. Machine learning definition and basics. In: An introduction to machine learning. Cham (Switzerland): Springer Nature..

 Google Scholar

Islam M, Hasan M, Wang X, Germack H, Noor-E-Alam M. 2018. A systematic review on healthcare analytics: application and theoretical perspective of data mining. Healthcare. 6(2):54. doi: 10.3390/healthcare6020054.

 Web of Science ®Google Scholar

Naylor C. 2018. On the prospects for a (deep) learning health care system. JAMA. 320(11):1099–1100. doi: 10.1001/jama.2018.11103.

 PubMed Web of Science ®Google Scholar

Iroju O, Olaleke J. 2015. A systematic review of natural language processing in healthcare. IJITCS. 7(8):44–50. doi: 10.5815/ijitcs.2015.08.07.

 Google Scholar