The best route for artificial intelligence to support care delivery is through a stepwise approach — aiding physician decision-making first, before complete automation.
- Justin G. Norden, MD, MBA, MPhil and
- Nirav R. Shah, MD, MPH
March 7, 2022
Summary
Health care is often frustratingly slow to embrace technology; electronic medical records were a striking example. While artificial intelligence (AI) is clearly a part of the future of medical care, due to the understandably conservative nature of medicine, AI’s potential will be first realized elsewhere. One field in particular that has captured the attention of many is autonomous vehicles (AVs). Large and small tech companies are actively involved in AV development, which has attracted billions of dollars of investments and many of the sharpest minds from around the globe. While the effort has yet to reach the end goal of vehicles that are fully autonomous in all circumstances (vehicles known as level 5 systems, the highest on a 0 to 5 scale), today there are widely deployed driver augmentation AI systems that are saving lives (level 1 and level 2 systems). When most people think of AI, whether vehicles or health care, they think of fully replacing the driver or fully bypassing the doctor. While there are many good reasons to completely replace the driver for transportation, this thinking is counterproductive in health care. We must learn from the challenges and many steps in deploying “fully autonomous” AI in other fields. Moreover, we must recognize that in health care there are distinct advantages of augmentation over complete automation.
Today, the algorithms most widely deployed in health systems are rule-based systems or decision trees that are set to trigger alerts when certain conditions are met. For example, sepsis alert systems notify care teams based on specified values of labs and vital signs. These systems, based on clinical trials, use a simple rule-based decision tree on vital signs entered into the electronic medical record (EMR) to automatically flag patients who might be at risk. While these systems are not perfect, and have recently come under scrutiny, they do flag many of the highest-risk patients.1 We see these rule-based systems deployed widely across a few areas of our health care systems today: in clinical decision-support systems in the EMR, in triage from online symptom checkers, and in our medical billing systems. While we are in the early days of deploying artificial intelligence (AI) in health care, many rule-based systems are already deployed. (Brief explanations of AI categories and terms used in this article are provided in Table 1.)
Table 1.
Artificial Intelligence Categories and Terminology
| Evolution of AI Categories | ||
|---|---|---|
| Term | Explanation | Examples |
| Rule-based Systems | Translating human knowledge into predetermined actions | Sepsis alerts for flagging an abnormal set of vitals |
| Supervised Learning | A set of algorithms that builds predictive models from data with known outcomes | A radiologist groups 1,000 chest X-rays into those with lesions and those without, then a supervised learning algorithm is trained on this labeled data to determine a pattern to sort these chest X-rays automatically into these groups |
| Unsupervised Learning | Sets of algorithms that identify patterns in data without predetermined labels or known outcomes | Using large gene expression data sets, we have discovered 10 breast cancer subtypes that correspond to different patient outcomes and treatment responses |
| Reinforcement Learning | These are AI algorithms that train toward a certain goal using an iterative approach of trial and error, updating their models dynamically toward the more successful outcomes | Type of AI used to train AlphaGO through iterative self-play, ultimately outperforming the world championship GO player |
| AI Terminology | ||
|---|---|---|
| Term | Explanation | Examples |
| Synthetic Data | This is data created artificially by algorithms that is meant to mimic real-world data | 1. People train fall-detection algorithms using artificially generated falls; the synthetic fall data is made by moving stick figure models through human poses that portray how people may fall 2. Use cases today in health care are to create digital twins for control arms in clinical trials instead of recruiting patients who will receive placebo |
| Computer Vision and Deep Learning | Types of AI techniques that use multilayer artificial neural networks trained off large data sets (deep learning) to find information in images or videos | Hospitals today are using computer-vision algorithms to find strokes in head-computed tomography scans (CTs) |
| Natural Language Processing (NLP) | These algorithms work to extract or “understand” useful information from human speech or text | Physicians are using NLP tools to transcribe speech to write patient encounter notes |
Source: The authors
How do rule-based systems fit in the AI deployment time line across industries? They were deployed in the 1960s, early in the development of AI.2,3 These were the first types of AI systems, designed to automate tasks by taking “knowledge” from a human and programming that into a machine. Initially, there was a lot of excitement that we could create intelligent systems by simply programming all of the facts and rules that we know. As the work progressed, however, people soon realized we could only solve simple problems with these systems. For more complex problems, the number of rules needed grew unmanageably large. While the field of AI moved on to more advanced systems to handle these more complex problems, for the most part, AI in health care progressed little beyond rule-based systems. But now there’s reason for hope.
Supervised Learning in Health Care
In the 1990s, AI researchers moved from rule-based systems toward data-driven approaches. In these systems, rather than prescribing exactly how the system should act with human-derived insights, algorithms were trained from historical data to derive patterns and insights. This allowed computers to derive insights directly from the data and make connections that humans might not have previously known.
The first step in data-driven approaches were supervised learning–based methods. Such systems relied on historical data with input and output labels. With clear targets for what task we wanted the algorithm to complete, computers could work to create their own rules for how to split the groups apart. One prominent example of supervised learning happened in the field of computer vision using deep learning on labeled images. Deep learning is a type of AI technique that uses multilayer artificial neural networks to find patterns in large data sets. These deep-learning networks showed how computers could be excellent at image detection and approach human performance.4
While the field of AI moved on to more advanced systems to handle these more complex problems, for the most part, AI in health care progressed little beyond rule-based systems. But now there’s reason for hope.
In health care, computer vision and deep-learning AI systems are some of the first next-generation AI systems to be studied and deployed in academic settings. In February 2022, we did a PubMed search of papers published in 2010 with the keyword search “deep learning” and found 107 papers identified. For 2021, this search returns 14,685 citations. These papers are most prominent in the fields of radiology and pathology — where there are large, curated data sets with clean labels. Deep-learning computer vision algorithms are now being deployed in hospitals, used in triaging head-computed tomography (CT) scans (Viz.ai), and predicting the risk of deterioration for Covid-19 patients based on chest X-rays.5 The most success is achieved when AI tools are carefully integrated to augment existing workflows.
Notably, companies such as Viz.ai are augmenting the health system using AI rather than trying to replace doctors completely. For example, Viz.ai specifically focused on the problem of trying to reduce the time to initiate interventional care for a stroke by detecting large vessel occlusions on CT scans and alerting teams immediately. Research has been published showing that these integrated algorithms were associated with reduced time to treatment and ultimately better patient outcomes.6 With this level of evidence, Viz.ai created the first AI algorithm to be reimbursed by the Centers for Medicare & Medicaid using the New Technology Add-on Payment (NTAP) designation,7 and Viz.ai is currently deployed across hundreds of health systems. Viz.ai tackled a very clear problem using AI and built their solution into a workflow that could be used immediately. There is a graveyard full of health care computer-vision companies that promised much but failed to make any impact on our health care system. By focusing on augmentation rather than replacement, Viz.ai is having a large impact today.
Deep learning is also being applied to problems outside of computer vision in the AI community, with impressive results. When large amounts of data are available, these systems perform well. But deep learning systems are sometimes criticized for being a black box, i.e., they do not give insight into the nature of what goes into a prediction. While it may be dangerous to rely on a system like this to independently determine a course of treatment, we can still use these highly predictive systems to augment clinician judgment. For example, such approaches are used to predict early sepsis, leveraging data from thousands of sources.8,9 Flagging a deteriorating patient sooner can make a big difference and, despite false positives, these systems can help save lives. Finally, explainable AI is an area of active research that works to show why an algorithm is flagging something and to give humans insight into the data that led to the result.
Unsupervised Learning Systems in Health Care
The next major shift in AI was toward unsupervised learning systems. For example, by analyzing thousands of gene signatures in women with breast cancer, researchers found 10 distinct subtypes of the disease — all of which respond differently to treatment and have different survival expectations.10 This empiric approach allowed unsupervised learning systems to identify novel phenotypes — new knowledge that is now used in clinical settings.
Another success of unsupervised learning techniques is in producing synthetic data, which is created artificially by algorithms and is meant to mimic real-world data. These techniques have been popularized by deepfakes or models that can produce a high-quality image, video, or audio of anyone, doing just about anything. One health care application for synthetic data is to create a digital twin, which is a virtual representation that serves as the counterpart to a physical entity or system. In health care, people are actively exploring digital twins as control arms for clinical trials. For example, as we typically know how a group of Alzheimer’s patients will progress in a control group, maybe instead of large cohorts being left untreated, we can use digital twins to power these trials. Companies such as Unlearn.AI are attempting this now and working with the U.S. Food and Drug Administration to get new treatments on the market faster.11 In cardiology, a field with lots of data to work with, people are using digital twins for researching more personalized cardiac care based on a unique cardiopulmonary model.12
Advanced AI Techniques
A final large category of AI techniques is reinforcement learning, an AI that trains itself toward a certain goal and updates its own model dynamically, as it does better or worse. This type of system is behind the breakthrough of AlphaGO (developed by DeepMind, a part of Alphabet) that eventually became the champion of the board game Go, after playing against itself to improve its own performance.13 In health care, reinforcement learning is currently confined to academia, and it has been challenging to create a strong feedback loop with messy health data.14 While reinforcement-learning AI models may not be deployed in the real world for some time, iterative quality improvement efforts that aim to build a learning health system are first steps in this direction.15
One health care application for synthetic data is to create a digital twin, which is a virtual representation that serves as the counterpart to a physical entity or system. In health care, people are actively exploring digital twins as control arms for clinical trials.
Finally, natural language processing (NLP) algorithms work to extract or “understand” useful information from human speech or text. Notably, NLP is not a distinct type of AI like supervised learning, unsupervised learning, and reinforcement learning; rather, it is a specific focus incorporating all these techniques to solve a specific problem. It is worth calling out because of rapid advances made over the past few years that have resulted in wide deployment across technology systems. Improved models such as GPT3 (an advanced NLP system developed by OpenAI) show just how advanced automated text generation has become, and how AI writing can be indistinguishable from that of humans.16
Today, these systems are not widely deployed in health care, but many health tech start-ups are starting to sell promising chatbot solutions to handle administrative workflows. Companies such as Memora Health, Health Note, and Nuance are deploying NLP chatbot systems for care navigation, triage, check-in, automated note-writing, content, and more. Such NLP systems are focused on augmentation of and integration with the health care system, not bypassing it completely for full automation. The most notorious example of aiming for too much with NLP in health care is IBM’s Watson Health, where billions of dollars were invested, but outcomes did not meet expectations.17,18
Learning from AI in Autonomous Vehicles
Autonomous vehicles (AVs) use rule-based systems to understand and follow the laws of the road, supervised learning systems to train their computer vision systems, unsupervised methods to find features that are most important in a particular scene, and reinforcement learning techniques to follow a goal and find the right driving path through an intersection.
Yet self-driving car companies are largely split on how to eventually solve the problem of reaching fully autonomous driving, considered a level 5 system on a scale of levels from 0 (fully operated by a human) to 5 (fully operated by the vehicle) (Table 2).19
Table 2.
Levels of Automation for Vehicles
Scroll table to see more
| Level 0 | Level 1 | Level 2 | Level 3 | Level 4 | Level 5 | |
|---|---|---|---|---|---|---|
| Concept | No driving automation | Advanced driver-assistance system, assists human with some tasks | Advanced driver-assistance system, full human attention required | Automated driving system, for some circumstances | Automated driving system, for certain circumstances | Full driving automation, for all circumstances, humans as passengers |
| Explanation | The human driver does all the driving. | An advanced driver-assistance system (ADAS) on the vehicle can sometimes assist the human driver with either steering or braking/accelerating, but not both simultaneously. | An advanced driver-assistance system (ADAS) on the vehicle can control both steering and braking/accelerating simultaneously under some circumstances. The human driver must continue to pay full attention (monitor the driving environment) at all times and perform the rest of the driving task. | An automated driving system (ADS) on the vehicle can perform all aspects of the driving task under some circumstances. In those circumstances, the human driver must be ready to take back control at any time when the ADS requests the human driver to do so. In all other circumstances, the human driver performs the driving task. | An automated driving system (ADS) on the vehicle can perform all driving tasks and monitor the driving environment — essentially, do all the driving — in certain circumstances. The human need not pay attention in those circumstances. | An automated driving system (ADS) on the vehicle can do all the driving in all circumstances. The human occupants are just passengers and need never be involved in driving. |
Source: The authors, informed by Automated Vehicles for Safety. U.S. Department of Transportation, National Highway Traffic Safety Administration. Accessed February 7, 2022. https://www.nhtsa.gov/technology-innovation/automated-vehicles-safety.
The first camp (Tesla and traditional automakers) is focused on building simple AV systems that help with steering, braking, and acceleration, that first augment the human driver (level 2 systems) but leave the driver responsible. The second camp (Waymo, Cruise, Zoox), is focused on high automation and fully autonomous systems (level 4 and level 5, respectively) that remove any responsibility of the driver and take full control, in most or all circumstances, respectively.
Getting to 100% automation is the goal for AVs, because the advantages of taking the driver completely out of the equation are clear and compelling. Not so in health care. In fact, quite the contrary, as the doctor-patient relationship is critical to outcomes.
Notably, level 2 systems like Tesla’s autopilot have driven more than 3 billion miles on the road, with crash rates lower than human drivers. However, there have also been catastrophic failures due to system issues and humans overestimating the autonomous system’s performance. Meanwhile, the fully autonomous systems have yet to be broadly deployed (e.g., Waymo is operating in a small area around Phoenix, Arizona; Cruise is operating in San Francisco, California) and may still be many years away from widespread use.20 While there is a large reward in both cost and safety by completely removing the human driver, the final step is the hardest. Systems today perform at around human performance. However, getting to meaningful deployment means they will have to be 1,000 times better — and this is significantly more challenging.21,22
Getting to 100% automation is the goal for AVs, because the advantages of taking the driver completely out of the equation are clear and compelling. Not so in health care. In fact, quite the contrary, as the doctor-patient relationship is critical to outcomes. Moreover, augmentation systems are far easier to develop and deploy and can be ready to use in years, not decades. Even without full automation, we can catch errors, aid diagnosis, and save time. Finally, full automation brings a host of barriers in terms of acceptance by patients, physicians, and regulators. So, in health care, when it comes to AI, less is actually more.
- Justin G. Norden, MD, MBA, MPhilPartner, GSR Ventures, Palo Alto, California, USA
- Nirav R. Shah, MD, MPHSenior Scholar, Stanford University, Stanford, California, USA; Chief Medical Officer, Sharecare, Atlanta, Georgia, USA
Source: NEJM
ARYAN'S BLOG 