Article Preview
TopIntroduction
Medicine, like many other fields, is experiencing the implementation of software tools described as artificial ‘intelligence’ (AI) or, more precisely, machine learning (ML). The purpose of this introductory paper is to review and explain basic concepts and caveats within this nascent topic to less familiar readers, not to present original research. Some definitions, basic concepts, and examples of AI in medicine will be reviewed with the goal of helping the reader understand how these tools may advance the process by which providers will improve standards of medical care. The topics discussed in this paper will be covered by the author in more depth in subsequent papers.
At its core, AI is the application of a type of mathematical tool that has been enabled by computers and databases. Computers make it possible to develop these tools. These tools are based on a body of mathematics combining Bayesian probability modeling, rules, and neural networks (NN). This transformational technology is a cause for much optimism and concern. The change enabled by any innovative technology is naturally followed by a gradual appreciation of the impact, but optimism is gradually tempered by real experiences, often as opinions are revised and refined in a dialectic manner. It takes time to understand the capabilities of these AI and ML tools before medical professionals can feel safe and comfortable working with them in critical tasks.
These AI or ML tools are not ‘intelligent’ or sentient, therefore, AI is an imprecise term. So, the author will refer to the implementation of AI and ML in a computational platform, as an AI Machine (AIM) tool. The transformations to be enabled by these AIMs have profound implications. What must follow, and is only beginning to develop, is a framework to guide thinking in terms of quality and standards for these new tools. Here again, there is a possibility for one AIM to supervise and train another AIM! There is some comfort in knowing that the same methods used by data scientists and knowledge engineers to develop these AIMs, may also be used to monitor and improve them.
Even the most powerful computers are still just machines: the only math they perform is the math they were programmed to execute by humans. And, because humans are not capable of anticipating all scenarios, machines will fail eventually, in circumstances that have not been anticipated and tested. Software, no matter how sophisticated, cannot anticipate every scenario—just as any expert will not be omniscient and consistent in all settings for all patients. In software these errors are called bugs. When applied to medicine, these errors would have potentially major consequences for all involved.
The methods used to develop these AIMs with expert input, intrinsically, use a process of continuous improvement while being monitored by other experts or end-users, or even by other machines. Using one AIM to monitor and ‘challenge’ another AIM to improve is referred to as a ‘competitive’ or ‘adversarial’ neural network. Therefore, one AIM can be used to monitor and train another. The monitoring of an AIM system for rational, ethical, and humane decision-making is a major concern in this nascent field. Methods like those used to develop these AIM systems can also be used to monitor and improve them. However, they must incorporate feedback from health professionals and patients who engage in the use of these tools.
There are some remarkably successful examples of machine learning, many of which have been comprehensively catalogued and reviewed in Eric Topol’s book, Deep Medicine, (e.g., detecting cancer patterns in mammograms, chest x-rays, and skin). While getting acquainted with the many AIM tools described, one cannot help but wonder how the medical industry will keep up with monitoring these tools for quality and safety. It is critical that medical professionals using these new tools continue to adhere to basic and fundamental principles, such as maintaining human oversight and ethical, humane care, with regards to decision-making.
One impressive development has been made in the diagnosis of breast cancer by Yala et al. (2021) who trained a computer to recognize patterns on mammograms to diagnose breast cancer. Trained on one major data set, the computational system predicted breast cancer by up to five years in advance with a concordance of 0.75 to 0.84, in images collected at seven hospitals across five countries. This application in the prediction of breast cancer, may be the largest single validation of an AI tool in image analysis.