Photo from NewScientist
Ho Chi Minh city, 13 May 2023 8:00 am
Doctors possess extensive knowledge, experience, wisdom, and judgment by virtue of their medical training. However, with the abundance of information available on human health and disease, even the most capable human brain cannot recall or comprehend all of it. In the past, medical decisions were primarily based on the doctor's personal knowledge. Nevertheless, the rise of artificial intelligence is gradually changing this, enabling medical decisions to be informed by a broader range of knowledge and resources.
The development that led to the creation of ChatGPT and other large language models has resulted in artificial intelligence becoming one of the fastest-adopted technologies in history, potentially bringing significant changes to our way of life and work. One area where such changes will be most prominent is healthcare. As the technology powering these systems advances, AI is poised to become an integral part of the healthcare industry, working alongside doctors, nurses, pharmacies, and waiting rooms. In the near future, AI is likely to either replace or significantly redefine all of these components of the healthcare experience.
AI "decision support systems" are already aiding physicians in obtaining a vast amount of information at the point of care. These systems leverage computers' natural aptitude for storing, retrieving, and correlating vast amounts of data, and combine them with the human experts' capacity to reason intuitively and think creatively. As a result, physicians can access a wealth of information quickly and easily when making decisions.
During the 1980s and 1990s, the development of early "expert systems" faced resistance from many physicians who were concerned about computers taking over medical decision-making. They feared that the "doctor's touch" would be eliminated, and physicians' opinions would be disregarded in favor of the computer's analysis. However, this did not come to pass. Research has demonstrated that these systems have improved significantly, assisting doctors in identifying potential outcomes that they may have overlooked. Furthermore, these systems do not take away the decision-making authority from physicians' hands.
The time is rapidly approaching when "centaur doctors," blending the most exceptional human intelligence with AI assistance, will be empowered to make courageous medical decisions with considerably fewer unintended repercussions. This is crucial because medical errors are responsible for approximately 250,000 deaths each year in the United States alone. It is not an overstatement to assert that AI-enabled healthcare has already preserved innumerable lives.
Dr. Gidi Stein developed an AI program known as MedAware to assist physicians in avoiding prescribing incorrect medication. The impetus for this system was the tragic death of a 9-year-old boy who received a prescription for a blood thinner instead of asthma medication due to a doctor's accidental selection of the wrong box. Unfortunately, medication errors like this are alarmingly frequent, with approximately 70% of medication errors resulting in adverse effects stemming from prescription errors.
It's easy to understand why prescription errors are such a widespread issue. The FDA has approved tens of thousands of prescription drug products, many of which have incredibly similar names, such as Novolin and Novolog, vinblastine and vincristine, and hydroxyzine and hydralazine. Given that doctors are well-known for their poor handwriting, it's not hard to envision how this could have led to problems in the days when most prescriptions were written by hand. Even in the digital age, with the advent of electronic prescriptions, mistakes can still occur due to a simple typo or momentary lapse in memory.
If a doctor prescribes a medication that does not align with a patient's medical needs, MedAware alerts the physician. The system also notifies doctors if they attempt to prescribe a medication that could have a negative interaction with one of the patient's existing medications - a mistake that physicians rarely check for. In hospitals across the globe where MedAware is used, the doctor still makes the final decision, with the system providing an extra layer of review that is especially useful when doctors are overworked and fatigued.
Doctors are often limited in their ability to think creatively due to the fear of making mistakes, resulting in a restricted range of treatment options. However, with the assistance of AI and access to individual data clouds, doctors can go beyond traditional methods and explore a wider range of potential treatments that consider a patient's unique characteristics such as their genetics, biochemistry, lifestyle, and personal history. By evaluating tens of thousands of possible outcomes, an AI-assisted doctor can offer more informed and confident recommendations for treatment. This approach goes beyond relying on clinical trials and averages and instead tailors treatment options to the specific needs of each patient.
Clinical decision support systems can personalize test results by taking into account various factors such as age, sex, ethnicity, and disease subtypes. One example of an AI tool that assists with clinical decision-making is DXplain, developed at Massachusetts General Hospital's Laboratory of Computer Science. This tool uses clinical manifestations to provide a probable diagnosis. A randomized control trial demonstrated that using DXplain helped family medicine residents score significantly better on a 30-case diagnosis test, improving their accuracy rate from 74% to 84%.
Nowadays, clinical decision-support systems are utilized to assist in laboratory testing and analysis, notifying healthcare providers and flagging unusual lab results. They may even avoid the use of more hazardous or intrusive diagnostic procedures in favor of safer alternatives. For example, liver biopsies are believed to be the most accurate method for assessing the severity of damage caused by hepatitis B and C, but AI models can integrate various data sources, such as imaging, blood tests, and genetics, to achieve significantly higher accuracy rates without the need for a biopsy.
According to a study, AI can be used to analyze urinary bladder tumors with a high accuracy rate of 93%. This technology has many applications in healthcare, such as automated tumor grading and decision support systems for clinical management, patient safety, cost containment, and diagnostics support. For example, one system can alert physicians when a patient is eligible for a clinical trial, while another ensures the accuracy of documentation, such as confirming that a patient has received necessary vaccines after spleen surgery. These AI systems are particularly beneficial in areas where access to clinical experts is limited, providing specialized care that would otherwise be unavailable.
Although there are some doctors who still resist using AI systems as aids, the trend is increasingly moving in that direction. It's common for there to be pushback when the balance of diagnostic power is changing. However, healthcare providers who integrate these systems into their practices will benefit their patients and themselves. In their optimal form, AI systems provide access to thousands of experts working together quickly. As AI is usually inexpensive to maintain once it has been developed, there is great potential to improve patient care and make it more cost-effective.
The progress made in medical AI is largely attributed to the abundance of data available. Massive collections of electronic health records (EHRs) have been gathered on patient interactions at major health institutions. On average, each patient generates about 80 megabytes of information per year, including imaging data, basic test results, and patient outcomes. Furthermore, larger data sets will soon be added, including genomics, blood analyses, gut microbiome analyses, and information from wearable devices, which will all become part of each patient's data cloud.
Doctors can use deep knowledge systems to analyze and interpret patient data, which can help them develop advanced diagnostic and treatment approaches that are tailored to each individual. The University of Florida has developed a system called GatorTron, which utilizes large language models and has been trained using 90 billion words from electronic medical records. This system is able to extract clinical concepts and provide answers to medical questions.
To achieve a future where computer systems can make decisions and provide explanations to humans, new AI advancements will be necessary. Traditional expert systems have difficulty scaling up because they become overly complex as more rules are added, leading to complicated decision trees. In contrast, human thinking is not solely based on rules. We are adept at recognizing when rules don't apply to a specific situation or when the logic fails. The human brain is remarkable in its ability to handle unexpected situations.
A major innovation comparable to the one that accelerated data-centric AI systems is essential for knowledge-based AI systems. This would result in a scenario in which "deep learning" is accompanied by "deep reasoning," which would enable AI to comprehend implicit connections rather than just those that have been explicitly coded into its software. The difficulty of this challenge stems from the fact that, unlike deep learning, where the addition of extensive computing power and data-fueled significant progress, we require conceptual advancements to make deep reasoning possible.
In the near future, it is expected that AI systems will be indispensable for doctors to cope with the increasing amount of medical data and insights. In the initial stages of collaboration, AI will suggest potential decisions to physicians, explained in a way that they can comprehend, allowing them to analyze the underlying reasoning. This approach is similar to how a medical specialist might assist a general practitioner in making a final care recommendation to a patient. As a result, AI will not only be a useful tool for dispensing health but also a powerful teaching tool.
As AI systems become capable of "deep reasoning," they will be able to discover connections and concepts that are beyond human comprehension. In the future, as natural language processing and large language models like ChatGPT advance, humans and computers will be able to engage in sophisticated discussions and consider different possibilities together in real time. The doctor will still have the final say in the decision-making process, but they will be able to rely on AI to help integrate the vast amounts of data and knowledge necessary to make informed decisions. With the help of deep reasoning, AI will uncover connections and concepts that would otherwise be inaccessible to humans.
In the future, AI systems may start providing doctors with advanced insights beyond simple correlations. If these insights prove effective, doctors may incorporate them into their practice, even if they cannot fully explain why they work. While this may seem like a radical idea, it's important to remember that many pharmaceutical treatments work without a complete understanding of the underlying biochemistry. Additionally, human physiology is incredibly complex, and we are constantly discovering new aspects of it that we don't fully understand.
Throughout history, humans have used medicines whose mechanisms they did not fully understand. For example, salicylate-rich plants were used by ancient civilizations to alleviate pain and inflammation long before the active ingredient, aspirin, was discovered. Even today, despite the fact that aspirin has been in use for over a century and a half, its mechanisms are still being studied. Similarly, when AI begins to surface complex insights that we cannot fully comprehend, we may still use it because it works, just like we use aspirin without fully understanding how it works. Although it may require a leap of faith, innovation often involves exploring the unknown before it ultimately becomes familiar.