A Short Guide for Medical Professionals in the Era of Artificial Intelligence PDF

www.nature.com/npjdigitalmed PERSPECTIVE OPEN A short guide for medical professionals in the era of artificial intelligence 1,2 ✉ 1 Bertalan...

www.nature.com/npjdigitalmed PERSPECTIVE OPEN A short guide for medical professionals in the era of artificial intelligence 1,2 ✉ 1 Bertalan Meskó and Marton Görög Artificial intelligence (A.I.) is expected to significantly influence the practice of medicine and the delivery of healthcare in the near future. While there are only a handful of practical examples for its medical use with enough evidence, hype and attention around the topic are significant. There are so many papers, conference talks, misleading news headlines and study interpretations that a short and visual guide any medical professional can refer back to in their professional life might be useful. For this, it is critical that physicians understand the basics of the technology so they can see beyond the hype, evaluate A.I.-based studies and clinical validation; as well as acknowledge the limitations and opportunities of A.I. This paper aims to serve as a short, visual and digestible repository of information and details every physician might need to know in the age of A.I. We describe the simple definition of A.I., its levels, its methods, the differences between the methods with medical examples, the potential benefits, dangers, challenges of A.I., as well as attempt to provide a futuristic vision about using it in an everyday medical practice. npj Digital Medicine (2020)3:126 ; https://doi.org/10.1038/s41746-020-00333-z 1234567890():,; INTRODUCTION We describe the simple definition of A.I., its levels, its methods, Artificial intelligence (A.I.) is expected to significantly influence the the differences between the methods with medical examples, the practice of medicine and the delivery of healthcare in the near potential benefits, dangers, challenges of A.I., as well as we future. While there are only a handful of practical examples for its attempt at providing a futuristic vision about using it in an medical use with enough evidence, hype around the topic is everyday medical practice. unprecedented1. There is a growing list of publications on the subject in the form of academic articles, health policy reports, statements from professional societies, and popular media cover- DEFINITION AND LEVELS OF A.I. age (Fig. 1). A.I. is an interdisciplinary field spanning computer science, A.I. has been used extensively in industries such as transporta- psychology, linguistics, and philosophy, among others. According tion, entertainment or IT during the last decade. It has been used to its simplest definition, artificial intelligence (A.I.) is intelligence to control self-driving vehicles; to trade on the stock market; social demonstrated by machines. It is sometimes also described as media platforms, web browsers and search engines. It is likely that “machines that mimic cognitive functions that humans associate the reader of this paper used some forms of A.I. today for more with the human mind, such as learning and problem solving”4. than an hour by using services such as Google Maps, Waze, Nick Bostrom, philosopher at the University of Oxford, described Facebook, Instagram, LinkedIn or Google Search, among a myriad three major levels in the development of A.I. in his book of others. It has potentials in medicine, drug design and Superintelligence (Fig. 25). healthcare, yet, the proof and evidence are yet to be convincing enough for the general public and the medical community to Artificial Narrow Intelligence (ANI) adopt the technology2. ANI already has incredible pattern recognizing abilities in huge The technology is still in its infancy and more studies are data sets, which makes it perfect for solving text, voice, or image- published each year than the year before. There are so many based classification and clustering problems. It is an algorithm that papers, conference talks, misleading news headlines and study can excel at a precisely defined, single task. It can play chess like interpretations that a short and visual guide any medical nobody else ever, yet its IQ is zero. professional can refer back to in their professional life might be useful. Artificial General Intelligence (AGI) There is no doubt that A.I. will have a beneficial role in healthcare and can penetrate the boundaries of adoption only if one day could have a human being’s comprehensive and total cognitive capacity. This is human level A.I. It can reason, argue, medical professionals serve as knowledgeable and supportive memorize and solve issues like you do guides and leaders in the process3. For this, it is critical that physicians understand the basics of the technology so they can see beyond the hype, evaluate A.I.-based Artificial SuperIntelligence (ASI) studies and clinical validation; as well as acknowledge the theoretically could have humanity’s combined cognitive capacity limitations and opportunities A.I. has. This paper aims to serve or even more. Humanity, obviously, would not be able to grasp its as a short, visual and digestible repository of information and knowledge and understand its reasoning. Many organizations details every physician might need to know in the age of A.I. work hard to avoid ever reaching this stage. The Medical Futurist Institute, Budapest, Hungary. 2Semmelweis University, Budapest, Hungary. ✉email: [email protected] 1 Seoul National University Bundang Hospital  B. Meskó and M. Görög 2 1234567890():,; Fig. 1 Number of medical A.I. studies by year from 2010 to 2020; and by medical specialties. a The number of studies as found on Pubmed.com using the search term)“machine learning” OR “deep learning”) and choosing a year in advanced search. b The same search method was used followed by (AND specialty) without specifying a time frame. The number in the circles determine how many studies we found. HOW DOES A.I. WORK? least the physical/biological explanations behind it. In the case of A.I. works through a method called machine learning. As there are advanced A.I.-based algorithms, it seems we will not be able to challenges and tasks so complicated in healthcare that writing understand more than the theoretical basics. traditional algorithms for solving those was not enough anymore, a new method was needed. Machine learning gives computers the REAL-LIFE EXAMPLES FOR THE SUBTYPES OF MACHINE ability to learn without being explicitly programmed6. If you feed LEARNING the algorithm with enough data of good quality, machine learning allows them to create strategies for excelling at that Machine learning has many subtypes and combined methods, but particular task. we only feature the three major subtypes besides an advanced If I want to write a program that can spot cats on photos, I had method, deep learning (Fig. 3). better turn to machine learning. The reasons why quickly become clear if you try to devise rules that such a program should be Supervised learning based on. How can you spot a cat in a photo? If you devise Is used when we can precisely define the task we want the features you think easily depict cats such as having two ears, two algorithm to learn based on data that we already have. Let’s take eyes, four legs and such, you find yourself in a situation where you the following example. We have two sets of medical records of also have to define all these expressions. What is an ear for a patients, group A and B. In one set, we have family history, lab program that only “sees” pixels on a photo? markers and other details with the diagnosis. In the other set, we Therefore, the most efficient way is to feed a machine learning have the same kinds of data but without the diagnosis. We would algorithm with images with cats on them, preferably with cats that like to build a model that can learn to assign the right diagnosis to have been annotated manually by human beings to make sure the patients in group B based on the associations and labels the images do contain cats. The more such annotated images we feed algorithm learns about in group A. It is like learning with a teacher the algorithm with, the better it will become at recognizing cats because we know exactly what the algorithm should learn, and it on images. It will not understand what a cat is, but it will certainly is by far the most frequently used training mode. recognize what we think are cats in photos like we do with an increasing and ruthless efficiency. Unsupervised learning No matter what task we aim to solve, we feed the simpler Is like learning without a teacher. We have a group of patients machine learning algorithms with data and constantly iterate how with different sets of data, but we do not know their individual it digests it to become better at solving the task. With more diagnoses. We build a model, then to try to cluster patients based complex algorithms such as neural networks and deep learning, it on similar attributes such as the symptoms they presented with, is possible that the algorithm starts creating its own rules and their lab markers or age and gender. We might learn new strategies without human input. From there, not even its associations we have not looked at before. In another example, it developers might understand how it draws a conclusion or the can also be useful in clustering tissue samples based on similar strategy it uses to excel at a task. gene expression values; or in finding novel drug-drug interactions. When it comes to technologies we use to make medical In summary, we devise certain rules, let the algorithm learn by decisions, we like to understand the core of the machine or at itself and we do not modify the algorithm based on the outcome. npj Digital Medicine (2020) 126 Seoul National University Bundang Hospital  B. Meskó and M. Görög 3 Fig. 2 Levels of A.I. The three levels of A.I. as defined by Nick Bostrom in Superintelligence. The green dot indicates a theoretical threshold for what the ideal scenario would be. Reinforcement learning Also, machine learning is often better suited for large numbers of Allows the algorithm to learn how to complete the tasks with a input variables (think of images with thousands of pixels), while sequence of decisions by itself without being told how to do it. traditional analysis with statistical models was designed for data The teacher is only able to give feedback after a series of actions, with tens of input columns9. not for each item as it does with supervised learning. The model starts performing the task only knowing some basic rules, and after failing or succeeding in completing the task, the teacher weighs in to push it to use the winning strategy more. This way, MACHINE VS DEEP LEARNING the program can build its own experiences as it performs the task more and more. It is similar to how we train dogs. When the dog Machine learning is a broad set of methods, their majority were performs or tries to perform a task, we only give it a treat if it used even decades before the current A.I. revolution. Most performed well. breakthroughs are achieved nowadays with artificial neural The most famous example for this method is how AlphaZero networks, but there are several other models, each with its own can learn to become the best player in any 2-players game in advantages. hours by playing millions of games against itself. It starts playing a Deep learning is a subset of machine learning and while it has game knowing only its basic rules, and the developers let the similar functions, its capabilities are different. Deep learning uses a algorithm know when it won a match to prioritize that strategy layered structure of artificial neural networks that is inspired by while playing the next game7. the neural network of the human brain. The internal structure and In an example, authors used this method to determine clinical number of layers within a neural network is a field of active trial dosing, where the algorithm learnt the appropriate dosing research, but as a rule of thumb we can say that a deeper net with regimen to reduce mean tumor diameters in patients undergoing more layers can learn more complex tasks - at the same time chemo- and radiation therapy8. The main challenge with applying requiring more data and longer time to train. Deep models have reinforcement learning to healthcare is that we cannot play out a the capacity to process images, sound and other high dimension- large number of scenarios as the lives of patients are at stake. ality data with good results, while other machine learning models may perform better with data organized into a spreadsheet. MACHINE LEARNING VS TRADITIONAL STATISTICAL MODELS Looking at a medical example, let’s build a model that can cluster patients by diagnosis based on the data in their medical A.I. is a huge domain which includes machine learning, records. If a medical record contains the expression Type 1 evolutionary algorithms, but also methods that do not even process quantitative data. Therefore, it makes sense to compare Diabetes, a machine learning model will learn to put all such machine learning to traditional statistical methods. patients in the Type 1 Diabetes cluster. But a deep learning Traditional statistical models focus on discovering relationships algorithm could learn with time without human input that and confidence intervals between data points and outcomes. patients with medical records that only mention T1D should also Compared to this, machine learning methods aim to reach high be assigned to the same group. Programmers of other machine prediction accuracy with putting less emphasis on whether it is learning algorithms should add these alternatives themselves. possible to interpret the model. The former often analyses a given Another example might also demonstrate the power and dataset for its insights, while the latter is trained on a set of data, potentials of deep learning. We would like to build a model that evaluated on a second set, and then used on an unseen third. turns on the light if we shout the word dark. A deep learning Thus, prediction is key in machine learning to generate model, by time, would learn that saying “I can’t see” or “it’s dark otherwise unavailable (e.g., expensive or not yet known) data. here” should also turn on the light. Seoul National University Bundang Hospital npj Digital Medicine (2020) 126  B. Meskó and M. Görög 4 papers and news on A.I. There are some general questions we can ask ourselves while reading medical A.I. papers. (Table 1)10,11. Besides obvious features such as the quality of the journal the paper was published in (real breakthroughs tend to be able to reach the attention of high-level journals), the most important factor is the source of data. It matters where the data stems from, thus it is worth checking the ‘Methods’ section where authors describe how, where and what kind of data they received. No algorithm can be trained without a good amount of quality data. Moreover, the size of the dataset also matters: the more images, text or any other source materials the researchers have, the more precise the algorithms can become. Collaborations with clinicians and healthcare institutions are crucial in getting large amounts of quality data. If not, some research groups do tricks on their dataset to make it bigger (e.g., rotate images to double the database size). The reported performance (e.g., precision, speed) of the algorithm should be compared to previously existing solutions and human abilities. Even a state-of-the-art technical solution may perform so much worse than human professionals in a real-world clinical setting, that it may not be helpful at all. It is crucial for such technologies to be evaluated whether they can easily be implemented in clinical protocols, and whether the results are straightforward for the medical staff to interpret. It is also key to see whether the paper analyzed a real clinical problem. An algorithm might perform very well on a pre-selected dataset, however, it should also be tested on real clinical data. DeepMind claimed in its published study that its model was able to accurately predict that a patient will develop acute kidney failure “within a clinically actionable window” up to 48 h in advance12. While the algorithm did do that, it cannot be clinically validated without testing it prospectively in a clinical setting. Also, when you look at the latest pieces of news on smart algorithms, you should watch out for the following to be able to carefully assess the quality of the given A.I.-related article. The term “artificial intelligence” itself might be misleading as due to the overuse of the expression, its meaning started to get inflated. If an announcement or news article mentions A.I. without describing the exact method under it, it is a good sign to be skeptical and careful. A company or a research group should mention a subtype of machine or deep learning and be able to explain the method in detail with which they are aiming to create A.I. In a relevant example that took place in Thailand, a Google- Fig. 3 Visual guide to machine and deep learning subtypes. a In backed algorithm aimed at improving care. There, normally, supervised learning, the teacher (developer) knows what he wants nurses took photos of the eyes of patients during check-ups and to teach to the child (A.I.), defines the expected answer and the child shared those to be analyzed by a specialist elsewhere. This process learns to excel at the task. b In unsupervised learning, the teacher could take up to 10 weeks. The A.I. developed by Google Health does not influence how the child learns to play but observes the conclusions the child can draw from solving the task. c In identified signs of diabetic retinopathy from an eye scan with an reinforcement learning, the teacher knows what he wants to teach accuracy of more than 90% and provide a result in less than to the child but does not define step-by-step how the child should 10 min. When Google started putting this system into practice, learn it. Instead, the teacher only gives feedback after the task is they ran into real-life issues. Sometimes a bad internet connection completed and asks the child to find out his own strategy using stopped the whole system from working as all images had to be those outcomes the teacher rewarded. d In deep learning, it is uploaded to the cloud through a strong connection. If the quality possible to analyze vastly more complex data sets from images and videos to a sort of human reasoning. It is multi-layered and could of the scan did not meet a certain threshold, it simply did not give mimic how neural networks in the brain work. a result (the deep learning algorithm has to be fed with high quality images to get better). This way, nurses had to spend time editing some of the images the algorithm chose not to analyze. HOW TO EVALUATE NEWS AND STUDIES ABOUT A.I.? Developers had to travel to the site to help sort out these issues13. Hardly a day goes by without promising research papers and As a summary, the medical journal Radiology proposed a guide studies on how to apply machine and deep learning to medical for authors, reviewers, and readers for assessing radiology problems. However, as already just mentioning A.I. makes research on artificial intelligence10. Other medical associations companies’ prospects better on any market, overhyping and and journal editorial boards might adopt and customize it too. overmarketing what an algorithm can do is an everyday They also describe a checklist that every author should complete phenomenon. There are still ways that help evaluate research when publishing their A.I. research. npj Digital Medicine (2020) 126 Seoul National University Bundang Hospital  B. Meskó and M. Görög 5 Table 1. A summary of some general questions readers of an A.I. medical paper might want to ask themselves to evaluate the quality of the results of a research. Questions What aspects of existing clinical practice does this system reinforce? Are the sizes of the training, validation, and test sets justified? How can we be sure the training data matches what we expect to see in real life and does not contain bias? How can we be confident of the quality of the ‘labels’ the system is trained on? Was the A.I. algorithm trained using a standard of reference that is widely accepted in our field? Was the manner in which the A.I. algorithm makes decisions demonstrated? Were the results of the A.I. algorithm compared with experts in my field? Is the system applied to the same diagnostic context that it was trained in? Is the A.I. algorithm publicly available? EXAMPLES FOR HOW HEALTHCARE COULD BENEFIT FROM A.I. Precision medicine In short, tasks that are highly repetitive and involve the analysis of The company Deep Genomics aims at identifying patterns in quantifiable data might benefit the most from the use of A.I. We genetic data and medical records of patients trying to link feature a few examples that provide a picture about the whole mutations to medical conditions. Oncompass Medicine uses A.I.- range of opportunities. based algorithms to match genetic mutations found in patients’ tumor samples with ongoing clinical trials worldwide. This way, Improving in-person and online consultations patients can receive precisely targeted treatments specific to the kind of cancerous tissue they have. Babylon Health launched an application that offers A.I.-driven consultation. It uses the patient’s medical history and common medical knowledge. Patients report their symptoms through the Designing treatment plans app which checks those in a database of diseases using speech IBM Watson developed a software that provides evidence-based recognition. After that, it offers a course of action14. treatment options for oncologists. It was designed to analyze both structured and unstructured data in medical records that may contribute to decision-making about treatment pathways. The Health assistance and medication management software combines data from the patient’s medical record with The medical start-up Sense.ly developed Molly, a virtual nurse that clinical expertise, and research papers to suggest promising was designed to have a smiling face coupled with a pleasant treatment plans19. There are many similar examples in other voice. Its goal is to help assist patients with monitoring their specialties. Creating an optimized radiation therapy delivery plan health or disease management in-between doctor’s visits using usually takes days. A.I.-based technologies help speed this process, ML. It also provides customized follow-up care, with a focus on completing the process in a couple of minutes20,21. chronic diseases15. A similar approach was used by the AiCure app that uses a smartphone’s camera and A.I. to confirm that patients Drug creation are adhering to their prescriptions16. This could be useful for people with serious medical conditions or participants in clinical The way pharmaceutical companies develop new drugs through trials. clinical trials can take several years and can cost billions of dollars. Speeding this up while making it more cost-effective would have an enormous effect on healthcare. The company Atomwise uses A.I.-driven diagnostics supercomputers to root out treatments from a database of In 2020, the FDA approved a software programme from the molecular structures. They also launched a search for a previously company Caption Health that allows medical professionals to unknown combination of safe and existing medicines and in a few perform cardiac ultrasound imaging without specialized training. days found two drugs predicted by the company’s A.I. technology It uses A.I. to provide real-time guidance and also the ability to which may significantly reduce the infectivity of Ebola. Such save images of diagnostic quality. It acts as a “co-pilot” for those analysis typically would have taken months or years22. performing an ultrasound scan as it was designed to emulate the guidance that an expert sonographer would provide to optimize Triage tools the image. It gives instructions on how to manipulate the It is crucial to be able to predict how severe a patient’s medical transducer, and provides automated feedback on diagnostic condition is to support the early identification of those who are image quality17. vulnerable and at high-risk, especially in emergency medical services. In a study, authors developed and validated an A.I.-based Mining medical records algorithm using deep learning that accurately predicted the need Collecting, storing, normalizing, and tracking medical records is an for the critical care of patients and outperformed the conventional obvious step for A.I. As an example, Google Deepmind Health is triage tools and early warning scores23. In another study, authors cooperating with the Moorfields Eye Hospital NHS Foundation analyzed online triage tools from more than 150 000 patient Trust to improve eye treatment by analyzing retina scans. The interactions with a chatbot, and they found a decreased level of images are analyzed by DeepMind’s algorithms resulting in a the urgency of patients’ intended level of care in more than one- detailed diagnosis and a so-called “urgency score” in about 30 s. quarter of the cases24. Both studies indicate that A.I.-based The prototype system can detect diabetic retinopathy, glaucoma, technologies can facilitate triaging even before patients reach the and age-related macular degeneration18. point-of-care. Seoul National University Bundang Hospital npj Digital Medicine (2020) 126  B. Meskó and M. Görög 6 HOW CAN AN A.I.-BASED MEDICAL TECHNOLOGY BECOME intelligence rather than replaces it. It also refers to the value an PART OF AN EVERYDAY PRACTICE? A.I. can provide that comes from how we can combine the unique The success of A.I. and its place in medicine and healthcare highly capabilities of human experts with those of A.I. to provide better depend on whether it can penetrate the boundaries of evidence- care for patients. A similar term that is related to augmented based medicine, the lack of policies and reluctance from medical intelligence is “human-centered A.I.” which explores the need for professionals to use it. There is no reason to believe that its use the development of A.I.-based systems that learn from and will be able to become common practice without meeting the collaborate with humans in a deep and meaningful way. standards and requirements of previous technologies. However, as the demand for A.I. to be implemented into Quality and quantity of data everyday medicine is getting higher by patients, policy makers, A.I. feeds on data. The more and better quality data it gets access medical professionals and hospitals, its way from developers to to, the more it can excel at tasks. Advanced algorithms need practice will have to become faster25. A typical example of how it annotated data to make sure those can learn the task they were has worked so far is related to Kardia, formerly known as AliveCor. designed for. There are medical professionals who act as data They first made an FDA-approved smartphone case that worked annotators which is a time-consuming and monotonous task. as a single lead ECG in 2012. They launched two clinical trials to Some medical algorithms can only improve through large test the hardware and the app comparing it to a traditional 12- amounts of annotated data. Therefore the dedicated contribution lead device. Later, the evolution of its design resulted in a credit- of data annotators is of crucial importance for the benefit of card sized device and an even smaller version in 2019. The original implementing A.I. in the healthcare setting. Therefore, we can device could provide a one channel ECG by playing the user’s conclude that data annotators are the unsung heroes of the fingertips on the sensor for 30 s. The results were uploaded to the medical A.I. revolution27. cloud to make it accessible for physicians. In 2015, Alivecor received clearance by the FDA to use an algorithm for the analysis Privacy issues of the readings to determine issues related to heart rhythm without human help. Medical A.I. needs access to medical records, data from health By the end of 2017, they already used deep learning networks, sensors, medical algorithms, apps and whatever source of and the FDA cleared the company’s ECG reader called KardiaBand information it can learn from. The data can come from healthcare as a medical device accessory to the Apple Watch. A study institutions or from individuals. Even if institutions make data concluded that the device managed to distinguish between atrial anonymized, it was proven in many cases that individual profiles fibrillation and a normal heart rhythm with a sensitivity of 93, and can be traced back. a specificity of 94%, respectively. Its sensitivity increased to 99% when a medical professional reviewed the reading26. Legal issues and liability By 2020, products of Alivecor have been tested in over 40 What if a deep learning algorithm misses a diagnosis, the doctor clinical studies. Despite these accomplishments, the use of the accepts the judgment and the patient suffers from the device is still not common practice. And as other companies consequences? What if an autonomous surgical robot injures a producing A.I.-based medical technologies are lagging behind, it patient during a procedure? It is an ongoing debate about who might depict a long period of adoption. will be held liable in the future when robots and A.I., acting A.I. will only reach the status of an everyday medical technology autonomously, harm patients. Current consensus states that the if medical associations provide clear guidelines about implement- professional is open to liability if he or she used the tool in a ing it; if policy makers create policies that favor adoption; and if situation outside the scope of its regulatory approval, or misused the medical community does not look at A.I. as a threat, but rather, it, or applied it despite significant professional doubts of the as the stethoscope of the 21st century. validity of the evidence surrounding the tool, or with knowledge of the toolmaker obfuscating negative facts. In any other cases, liability falls back on the creators and the companies behind them. WHAT ARE THE MAJOR CHALLENGES AHEAD? Examples for ANI do exist today, but there are major issues the A.I. Trust developer and the medical community have to face and tackle We will need a lot of time to trust an autonomous car, to see how before A.I. can become mainstream in medicine. it reacts in situations we are familiar with or whether it makes similar decisions in an emergency. Consequently, it will take even Explainability more time not only for patients but for medical professionals too Medical professionals tend to make decisions using data that were to trust A.I. with medical diagnoses, supporting medical decision- obtained with technologies they either understand or understand making or designing new drugs. This should be taken into the basics enough to trust it. In the case of A.I., it might not be consideration when we decide to adopt the technology into the possible. However, millions of learned parameters (the connection healthcare setting. weights within the network) determine the output of a deep neural network, which makes it unintuitive to understand the Biased A.I. decision process. Even if we visualize the sensitivity of different A study concluded that commercial companies’ facial-recognition parts of a network and browse these thousands of noisy images, systems were more accurate on lighter-skinned individuals by we will still not see easy-to-grasp learned rules. Reasoning is not a 11–19%. Those produced especially inaccurate results when by-product of the algorithm. Thus, explainable A.I. would be identifying women of color. In another example, A.I. was crucial in providing insights into A.I.-based algorithms enough to implemented in the United States’ criminal justice system in gain trust in them. order to predict recidivism. They found that the algorithm predicted disproportionately high probability of black people Augmented intelligence committing future crimes, no matter how minor their initial This is a term often promoted by organizations such as the offense was. It is not only racial prejudice, but A.I. algorithms also American Medical Association. It focuses on A.I.’s assistive role in often discriminate against women, minorities, other cultures, or healthcare emphasizing that A.I.‘s design enhances human ideologies. For example, Amazon’s HR department had to stop npj Digital Medicine (2020) 126 Seoul National University Bundang Hospital  B. Meskó and M. Görög 7 using their A.I.-based machine learning tool which the company highly repetitive and data-based professions, or rather tasks under developed for sorting out the best job applicants, as it turned out those professions will probably be highly impacted by automation, that the smart algorithm favored men. As such algorithms learn the core of the medical profession is and still will be the human from the data they are fed with, A.I. programmers must know touch, empathy and compassionate care; attributes that are almost about the issue of bias in algorithms and actively fight against it impossible to mimic through a programming language. by tailoring them28. However, as a general rule of thumb, we might be able to assume that those medical professionals who use A.I. will replace Patient design those who do not do so. That is how such a profound role A.I. will have in the future of medicine. The debate early in the 21st When designing algorithms for medical purposes, patients should century should not be about whether A.I. takes away the human be involved on the highest level of decision making to make sure touch or the art of medicine, but what we should do to improve their needs are met, and issues and recommendations are built both with it. into the technology. An example about its importance is how a When an algorithm using reinforcement learning, thus not start-up developed an algorithm that could detect signs of being restricted by our human cognitive limitations, comes up Alzheimer’s disease in phone calls of patients in Canada. However, with a cure which we could never find with our knowledge about it showed different results with patients who had a French accent. biology, medicine and other life science, the real art of medicine By inviting patients at the early stages of development, such will be to find out and to understand how it did that35. issues could be avoided. With this basic knowledge about the definition, levels, methods, As there are positive ongoing efforts for solving each of challenges and potentials of A.I., we tried to give an overview these, it is still an open question whether those algorithms that about how we can make the medical profession more creative, become a common part of medical practices would be able to spending more time with patients than ever. address them all29. Received: 18 May 2020; Accepted: 31 August 2020; THE FUTURE ROLE OF A.I. IN MEDICINE AND HEALTHCARE Every A.I.-based technology that is considered for use in healthcare must be regulated, efficient and backed by evidence30. The US Food and Drug Administration (FDA) has been showing an REFERENCES example in achieving a regulatory environment that not only 1. Topol, E. J. High-performance medicine: the convergence of human and artificial welcomes such innovations but is also able to keep them safe for intelligence. Nat. Med. 25, 44–56 (2019). the public. The FDA launched a branch for digital health in 2019 2. Matheny, M. E., Whicher, D., Thadaney & Israni, S. Artificial intelligence in health and has attempted at designing new regulatory standards for A.I.- care: a report from the national academy of medicine. JAMA 323, 509–10 (2020). based technologies. 3. Faes, L. et al. A clinician’s guide to artificial intelligence: how to critically appraise The FDA realized that in the era of A.I., more algorithms will machine learning studies. Transl. Vis. Sci. Technol. 9, 7 (2020). become available than medical devices as there is a shift from 4. Russell, S. & Norvig, P. Artificial intelligence a modern approach third edition hardware to software. As the number of algorithms to be (Pearson, 2010). 5. Bostrom, N. Superintelligence: paths, dangers, strategies. http://ovidsp.ovid.com/ regulated grows exponentially, the current resources of regulatory ovidweb.cgi?T=JS&PAGE=reference&D=psyc11&NEWS=N&AN=2014-48585- bodies will not be sufficient for assessing each iteration and 000 (2014). update. There have been discussions around a new regulatory 6. Samuel, A. L. Some studies in machine learning using the game of checkers. IBM J framework for modifications to A.I./machine learning-based soft- Res Dev. 44, 206–226 (2000). ware as a medical device (SaMD)31. This could potentially lead to 7. Sadler, M. & Regan, N. Game changer: AlphaZero’s groundbreaking chess stra- regulations that make it possible for regulators to assess tegies and the promise of AI. (New in Chess, 2019). companies while the companies can roll out algorithms and 8. Moding, E. J., Kastan, M. B. & Kirsch, D. G. Strategies for optimizing the response of updates without the need to check them all. This is a viable way of cancer and normal tissues to radiation. Nat. Rev. Drug Discov. 12, 526–542 (2013). 9. Bzdok, D., Altman, N. & Krzywinski, M. Statistics versus machine learning. Nat. letting A.I.-based technologies become widespread while keeping Methods. 15, 234–235 (2018). them safe. Our research group has attempted at creating a 10. Bluemke, D. A., Moy, L., Bredella, M. A., Ertl-Wagner, B. B., Fowler, K. J. & Goh, V. J. constantly updated database of FDA-approved A.I.-based medical et al. Assessing radiology research on artificial intelligence: a brief guide for technologies. After cross-checking and validating every approval authors, reviewers, and readers-from the Radiology Editorial Board. Radiology we could find, we identified 64 A.I.-based and, at the same time, 294, 487–489 (2020). FDA approved medical technologies. Only 29 of those (45%) 11. Challen, R., Denny, J., Pitt, M., Gompels, L., Edwards, T. & Tsaneva-Atanasova, K. mentioned any A.I.-related terms or expressions in the announce- Artificial intelligence, bias and clinical safety. BMJ Quality Safety. 28, 231–237 ment released by the FDA. (2019). 12. Powles, J. & Hodson, H. Google DeepMind and healthcare in an age of algorithms. Such technologies can only be efficient if they are successfully Health Technol. 7, 351–367 (2017). implemented into the medical practice. The American Medical 13. Page, W. D. H. Google’s medical AI was super accurate in a lab. Real life was a Association (AMA) has addressed the importance of A.I., has different story. MIT Technological Review. https://www.technologyreview.com/ advocated for the use of the expression augmented intelligence, 2020/04/27/1000658/google-medical-ai-accurate-lab-real-life-clinic-covid- and has assumed thought leadership with its reports and diabetes-retina-disease/. guidelines for physicians. In their AI policy, they state that “as a 14. Lunden, I. Babylon Health is building an integrated, AI-based health app to serve leader in American medicine, our AMA has a unique opportunity a city of 300K in England. Techcrunch. https://techcrunch.com/2020/01/22/ to ensure that the evolution of AI in medicine benefits patients, babylon-health-is-building-an-integrated-ai-based-health-app-to-serve-a-city-of- 300k-in-england (2020). physicians and the health care community”32. 15. Ricci, M. Sensely and Mayo Clinic take virtual nurse one step further. Pharma- To make sure that A.I.-based technologies meet the standards phorum. https://pharmaphorum.com/news/sensely-mayo-clinic-develop-virtual- of evidence-based medicine, numerous editorial boards of doctor/ (2017). medical journals and prestigious medical associations such as 16. Belciug, S. & Gorunescu, F. Intelligent decision support systems–a journey to the WHO or CDC have released their recommendations for the smarter healthcare. (Springer International Publishing, 2020). medical community33,34. 17. Staines, R. FDA approves Caption Health’s AI-driven cardiac ultrasound soft- One of the potential obstacles in adoption can be a common ware. Pharmaphorum. https://pharmaphorum.com/news/fda-approves-caption- fear among medical professionals that A.I. will replace them. While healths-ai-driven-cardiac-imaging-software/ (2020). Seoul National University Bundang Hospital npj Digital Medicine (2020) 126  B. Meskó and M. Görög 8 18. Moorfields announces research partnership. http://www.moorfields.nhs.uk/news/ 34. World Health Organization. Big data and artificial intelligence. https://www.who. moorfields-announces-research-partnership. int/ethics/topics/big-data-artificial-intelligence/en/. 19. Zauderer, M. G. et al. Piloting IBM Watson Oncology within Memorial Sloan 35. Meskó, B., Radó, N. The real era of the art of medicine begins with artificial Kettering’s regional network. J. Clin. Oncol. 32, https://ascopubs.org/doi/abs/ intelligence. The Medical Futurist. https://medicalfuturist.com/artificial- 10.1200/jco.2014.32.15_suppl.e17653 (2014). intelligence-and-the-art-of-medicine/ (2019). 20. Fornell, D. Artificial intelligence greatly speeds radiation therapy treatment planning. itnonline.com. https://www.itnonline.com/article/artificial-intelligence- greatly-speeds-radiation-therapy-treatment-planning (2020). AUTHOR CONTRIBUTIONS 21. AI treatment plans used in patients. uhnresearch.ca. https://www.uhnresearch.ca/ B.M. designed the study and wrote the paper, M.G. contributed to the paper. news/ai-treatment-plans-used-patients. 22. Atomwise finds first evidence towards new Ebola treatments. http://www. atomwise.com/atomwise-finds-first-evidence-towards-new-ebola-treatments/. COMPETING INTERESTS 23. Kang, D. Y. et al. Artificial intelligence algorithm to predict the need for critical care in prehospital emergency medical services. Scand. J. Trauma Resusc. Emerg. The authors declare no competing interests. Med. 28, 17 (2020). 24. Winn, A. N., Somai, M., Fergestrom, N. & Crotty, B. H. Association of use of online symptom checkers with patients’ plans for seeking care. JAMA Netw. Open. 2, ADDITIONAL INFORMATION e1918561 (2019). Correspondence and requests for materials should be addressed to B.M. 25. Sendak, M. P. et al. A path for translation of machine learning products into healthcare delivery. EMJ Innov. https://doi.org/10.33590/emjinnov/19-00172 Reprints and permission information is available at http://www.nature.com/ (2020). reprints 26. Himmelreich, J. C. L. et al. Diagnostic accuracy of a smartphone-operated, single- lead electrocardiography device for detection of rhythm and conduction Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims abnormalities in primary care. Ann. Fam. Med. 17, 403–411 (2019). in published maps and institutional affiliations. 27. Meskó, B., Radó, N. Data annotators: the unsung heroes of artificial intelligence development. The Medical Futurist. https://medicalfuturist.com/data-annotation/ (2019). 28. Chen, I. Y., Szolovits, P. & Ghassemi, M. Can AI help reduce disparities in general Open Access This article is licensed under a Creative Commons medical and mental health care? AMA J Ethics. 21, E167–179 (2019). Attribution 4.0 International License, which permits use, sharing, 29. Panch, T., Mattie, H. & Celi, L. A. The “inconvenient truth” about AI in healthcare. adaptation, distribution and reproduction in any medium or format, as long as you give npj Digit Med. 2, 77 (2019). appropriate credit to the original author(s) and the source, provide a link to the Creative 30. Anderson, M. & Anderson, S. L. How should AI Be developed,validated and Commons license, and indicate if changes were made. The images or other third party implemented in patient care? AMA J Ethics. 21, E125–130 (2019). material in this article are included in the article’s Creative Commons license, unless 31. FDA. Proposed regulatory framework for modifications to artificial intelligence/ indicated otherwise in a credit line to the material. If material is not included in the machine learning (AI/ML)-based software as a medical device (SaMD) - Discussion article’s Creative Commons license and your intended use is not permitted by statutory Paper and Request for Feedback. https://www.regulations.gov/document? regulation or exceeds the permitted use, you will need to obtain permission directly D=FDA-2019-N-1185-0001 (2019). from the copyright holder. To view a copy of this license, visit http://creativecommons. 32. Association TAM. Artificial intelligence in medicine. https://www.ama-assn.org/ org/licenses/by/4.0/. amaone/augmented-intelligence-ai. 33. John Howard, M. Artificial intelligence: implications for the future of work CDC. https://blogs.cdc.gov/niosh-science-blog/2019/08/26/ai/ (2019). © The Author(s) 2020 npj Digital Medicine (2020) 126 Seoul National University Bundang Hospital