Introduction
Artificial intelligence in medicine, also known as medical AI, is the application of AI techniques and algorithms in the field of medicine and healthcare. It involves the use of computer algorithms and machine learning models to analyze complex medical data, make predictions, and assist in decision-making processes. There are currently two main categories of medical AI as defined by the FDA: decision support software and autonomous medical AI. While most approved medical AI applications currently fall under the first category, the second both shows promise and is an extremely important goal in the development of medical technology.
Decision support software
The first category, that of decision support software, refers to software which does not replace clinicians in diagnosing or referring patients; rather, it helps medical professionals in the functions that they already fulfill during the normal course of their work, allowing them to allocate more time to patient care. The vast majority of medical AI falls under this category. One such example would be triage and notification software, which helps doctors prioritize patients in need of urgent care. Another is the use of decision support for clinical management - software of this type reminds patients of actions they need to take to maintain their own care, or may alert doctors to patients who are not following management plans or need additional assistance.
There are many additional types of decision support software, including cost containment, which attempt to save money in hospital systems and the like; diagnostic support in both general and specific fields, which attempt to simplify decision making for medical professionals; and decision support for patients, which communicate directly with patients and may do things like help patients decide whether their symptoms are serious enough to visit the doctor.
As the demand for doctors continues to rise and specialist visits come with lengthy wait times, AI systems present a promising opportunity to alleviate the strain on health systems by enhancing the efficiency of healthcare workers. By adopting AI technologies, doctors can streamline their workflows, optimizing their productivity and ultimately allowing them to dedicate more attention to treating patients. Consequently, these systems have the potential to significantly improve the overall efficiency and effectiveness of healthcare delivery.
Autonomous AI
The second category, that of autonomous AI, refers to software which replaces clinicians in diagnosing certain conditions. This is an extremely narrow category: as of 2023, only three diagnostic applications have been approved in this framework, all of which diagnose diabetic retinopathy. Diabetic retinopathy, an extremely common complication of diabetes which can lead to blindness, often remains undiagnosed due to patients not receiving yearly ophthalmological screenings from specialists. The three applications for autonomous diagnosis, created by Digital Diagnostics, Eyenuk, and AEYE Health, scan fundus images of patients’ retinas in order to detect the presence of diabetic retinopathy, cutting down on the arduous yearly referral process.
These applications have shown technological advancements over the past few years: while those by Digital Diagnostics and Eyenuk required two images per eye, AEYE Health’s requires only one fundus image per eye, significantly reducing the screening time to under two minutes and streamlining the process, as well as showing improved accuracy. The simplicity of the process, compared to the traditional yearly specialist visit, also offers opportunities for more accessible patient screening. An additional and significant benefit of AI screenings is their practicality. These screenings can be conveniently performed within the patient's point-of-care. This has been demonstrated in the use cases of these devices, most commonly in general practitioners’ offices, endocrinology offices, and public health services targeting the uninsured or underinsured. In all of the above practitioners are able to screen patients on the spot and receive immediate results, working to close the care gap between patients who need screening and patients who actually receive it.
While the field of autonomous medical AI is currently very narrow, many companies are working on developing further solutions, and many more believe that this is the logical next challenge for medical AI to conquer. As it stands, however, autonomous diabetic retinopathy screening is the only function which fulfills this promise of medical AI - the ability to diagnose patients without the traditional specialist on hand. Parties seeking to expand this field will have their work cut out for them, too, as autonomous AI is under enormous scrutiny from the FDA and other regulatory bodies.
Considerations in AI adoption
While both decision support and autonomous AI devices may perform well in the sense that they provide accurate results, their actual adoption into medical settings come with many considerations. One consideration is that of cost. Another critical consideration is that of the change to the current or accepted workflow - a system that does not mesh well with the way in which practitioners are accustomed to working will most likely be little used, or possibly even ignored. Furthermore, software which is unintuitive may require extensive training and onboarding, which are expensive and time consuming. Unintuitive software may also lead to incorrect data entry, which impacts the program’s calculations and may lead to incorrect answers.
Examples of decision support devices with minimal impact on workflow are two models developed by Microsoft, one of which gathers and records information intelligently during a doctor’s normal intake process, allowing the doctor to speak to the patient freely; the other of which monitors calls to emergency services and detects if the caller is experiencing a cardiac arrest. In both examples providers are not required to drastically change the routines they are accustomed to, making adoption easy and relatively painless.
Decision support systems with large impact, even to the point of disruption, can cause alert fatigue in providers and are often ignored. One way of avoiding this is by developing systems in cooperation with doctors, in which developers ask providers about their work practices and then create systems to augment this instead of requiring providers to learn new habits or procedures. An additional study done with nurses showed that experiential knowledge (i.e., how long a nurse has been doing their job) played a large factor in the adoption of decision support software: the longer a nurse had been doing their job, the less likely they were to adopt the technology, instead opting to override the technology. This provides further evidence that impact on providers’ workflow is critical - the more established the habits are, the harder they are to break or change.
Much like the integration of support software, the successful integration of autonomous AI solutions into workflows relies on multiple factors, including the design of the solution itself, operation time, and ease of use. A straightforward and practical solution is more likely to be seamlessly integrated into existing workflows. AEYE Health's solution exemplifies this approach, enabling patients to undergo a quick two-minute exam while already waiting, and discuss the results immediately with their physician. Another possible aspect of successful integration is the possibility of seamless integration into the existing Electronic Medical Record (EMR) system. While each solution operates differently, achieving true practicality necessitates smooth software integration.
Summary
Both decision support and autonomous AI are being increasingly adopted by medical professionals and patients, each with their own advantages. While most medical AI falls under the decision support classification currently, autonomous AI has immense potential to change the way we conduct medical procedures precisely because it doesn’t require as much medical professional support. Although the only autonomous medical AI currently offered is screening for diabetic retinopathy, the field shows extreme promise. We can reasonably expect to see more applications of autonomous medical AI in the coming years, as well as wider adoption of existing technologies, changing the way we think about and practice medicine.
