Editorial: ML and AI Safety, Effectiveness and Explainability in Healthcare

The increasing performance of machine learning and artificial intelligence (ML/AI) models has led to them being encountered more frequently in daily life, including in clinical medicine (Bruckert et al.; Rosenfeld et al., 2021). While concerns about the opaque “black box” nature of ML/AI tools are not new, the need for practical solutions to the interpretability problem has become more pressing as ML/AI devices move from the laboratory, through regulatory processes that have yet to fully catch up to the state-of-the-art (Benrimoh et al., 2018a), and to the bedside. This special edition targets three key domains in which innovation and clearer best practices are required for the implementation of ML/AI approaches in healthcare: ensuring safety, demonstrating effectiveness, and providing explainability. Notably, the first two have long been staples in the evaluation of drugs and medical devices (i.e., in order to be approved for human use, products must prove that they are safe and effective—often compared to a reasonable comparator) (Spławiński and Kuźniar, 2004). The third requirement—that of explainability—appears to be unique to ML/AI, due to the challenge of explaining how models arrive at their increasingly accurate conclusions. Yet, upon closer examination, one might argue that the explainability criterion has been implied in the past: mechanisms of action of drugs and devices are generally described in their product documentation (Health Canada, 2014). However, this can be misleading. For instance, many drugs have known receptor binding profiles and putative mechanisms of actions, although the precise mechanisms by which they produce their effect remain unclear despite their widespread use in clinical practice. Prime examples of this are lithium (Shaldubina et al., 2001) and electroconvulsive therapy (Scott, 2011), both longstanding and highly effective treatments whosemechanisms of action remain controversial. Indeed, even the precise mechanism of general anesthesia is a subject of debate (Pleuvry, 2008). As such, wemust consider a compromise-that of sufficient explainability (Clarke and Kapelner). This involves answering the question: howmuchmust we know about a model in order to determine that it is safe to use in clinical practice? The articles in this special edition begin to explore possible answers to this as well as other key questions in the application of ML/AI to healthcare contexts. Bruckert et al. propose a Comprehensible Artificial Intelligence (cAI) framework, which they describe as a “cookbook” approach for integrating explainability into ML/AI systems intended to support medical decision-making. Notably, the authors do not limit explainability to an understanding of general rules a model might use to make predictions, but rather extend it to an example-level approach where human-interpretable semantic information is passed from the Edited and reviewed by: Thomas Hartung, Johns Hopkins University, United States