Primary Care Artificial Intelligence: A Branch Hiding in Plain Sight

THE ORIGINS AND CURRENT STATE OF PRIMARY CARE ARTIFICIAL INTELLIGENCE

In 1966, Ian McWhinney argued that general practice should be a discipline and predicted that its research would thrive in the twilight zones where branches of knowledge converge.¹ In a review in this issue of Annals, Kueper et al have uncovered one such zone at the interface between computer science and primary care by describing a collection of research that has been hiding in plain sight since 1986.² By connecting 2 disciplines, these 405 articles constitute an area of focus—primary care artificial intelligence—that may be new to primary care researchers but has already generated an impressive compilation.

Despite this body of work, primary care artificial intelligence has failed to transform primary care due to a lack of engagement from the primary care community. Similar to health information technology, primary care artificial intelligence should aim to improve care delivery and health outcomes³; using this benchmark, it has yet to make an impact. Even though its history spans 4 decades, primary care artificial intelligence remains in the “early stages of maturity” because few tools have been implemented.² Changing primary care is difficult when only 1 out of every 7 of these papers includes a primary care author.² Without input from primary care, these teams may fail to grasp the context of primary care data collection, its role within the health system, and the forces shaping its evolution.

THE 6 CHALLENGES

In a previous commentary, we outlined why the relationship between primary care and artificial intelligence scholars is synergistic⁴ and in this piece propose how the 2 fields can collaborate to address the challenges standing in the way of greater impact. To start, we asked 30 artificial intelligence scholars about the foundational challenges in artificial intelligence relevant to health care. The participants noted that, compared with other industries, health care poses 6 challenges, stemming from 2 unique features. The first feature is that health care is a safety-critical environment, where malfunctioning of an artificial intelligence system can lead to injury or death.⁵ This is in contrast to other industries, such as e-commerce, where the stakes are much lower. As a result, the data used needs to be timely and accurate, and the resulting tools need to be explainable and free from bias. The second feature is that health care data are proprietary and therefore restricted. This is in contrast to other domains such as meteorology, where important data are publicly available.

One threat to high performance within a safety-critical system is inefficient data entry (Challenge #1). In health care, the process is seldom automated and too often depends on physicians, who may lack time to enter data. Without timely data, artificial intelligence systems do not have the information they need to make decisions. Once entered, these data are not processed, raising doubts about their accuracy (Challenge #2). Clinical data are not only stored in heterogeneous formats but can also be inconsistently entered or missing entirely. In response, researchers omit or modify data according to arbitrary or inappropriate rules, which can lead artificial intelligence systems to learn the wrong lessons. Explaining how these models learn has become increasingly difficult with the proliferation of unsupervised and reinforcement learning (Challenge #3). For users to trust artificial intelligence systems, they need to understand why decisions are made. Policy makers have even suggested that artificial intelligence systems need to provide explanations for their decisions before being deployed in health care.⁶ Once implemented, another concern is that these tools magnify existing biases (Challenge #4). The systematic under- or over-prediction of probabilities for populations emerges for multiple reasons, including biased training data and outcomes influenced by earlier, biased decisions.⁶ For instance, one study found that an algorithm that identifies patients for a care management program was biased against African Americans.⁷ Despite having worse comorbidities, African-Americans were less likely to be referred because the algorithm predicted costs rather than illness. Because of unequal access, African Americans historically received fewer resources and were therefore less costly. These inequities were embedded within the training set used to predict future costs. Because data reside in siloes, training data may not be representative of the sample for which the models will be applied (Challenge #5). This leads to tools that perform worse when used at different institutions.⁸ Furthermore, the population on which the tool was trained may shift, causing its performance to suffer over time. Sharing data across health systems can enable generalizable models; however, this flow of information raises privacy concerns (Challenge #6). With the digitization of data, patients are increasingly unable to determine when, how, and to what extent information about them is communicated to others.⁹ Breaches and misuse erode trust in artificial intelligence systems and may make individuals reluctant to access care.

Inherent in these challenges are important research questions. For example, can voice-based assistants improve the efficiency of data entry? How can we distinguish between data entry errors and nonconforming results that have value? What processes enhance explainability? What practices prevent biases from being perpetuated? How can we develop models that are generalizable and adaptable? How can research teams operationalize privacy by design, where privacy is built into every step?⁶

THE 7TH CHALLENGE

To stimulate progress, artificial intelligence and primary care researchers need to move from mutual awareness to partnership so that these questions can be applied to primary care contexts.¹⁰ We agree with Kueper et al’s call to cultivate teams consisting of both artificial intelligence and primary care researchers and urge them to work in a transdisciplinary manner.²

By doing so, these teams will create new conceptual frameworks and research strategies that transcend their preexisting disciplinary boundaries.¹¹ In other words, we do not simply need the application of artificial intelligence to primary care, but rather, the development of new methods that are tailored to the breadth, complexity, and longitudinality of primary care, bringing us to a 7th and final challenge. McWhinney believed that generalists are set apart by our overriding interest in people, an interest that is vital to the creation of a bond between physician and patient.¹ With the proliferation of electronic health records, this connection is being interrupted and risks being severed completely as we add more and more layers of technology.¹² To avoid this fate, primary care artificial intelligence needs to narrow this divide by facilitating new opportunities for connection. Finding creative solutions to this challenge is necessary if we hope to restore the relationships that sustain us and our patients.