The field of medicine has entered the era of artificial intelligence (AI; Table). Increases in computational power, scale and availability of data, and the development of novel machine learning methods have enabled the creation of tools that ultimately may transform medical care and improve health. Each documented patient encounter, laboratory or imaging test result, along with additional data captured from a patient’s mobile device, can be aggregated and analyzed using increasingly complex analytic approaches.
| Term | Definition/example |
|---|---|
| AI | A computational science technique for creating systems that can complete tasks that typically would require human intelligence (diagnosis, treatment decisions) |
| Machine learning | A subset of AI in which systems can learn automatically from data and recognize patterns without explicit programming (eg, a sepsis prediction algorithm that improves with additional data analyses) |
| Generative AI | AI that uses existing data (text, images, videos) to create new content |
| LLM | A subset of generative AI that focuses only on text outputs (such as discharge summaries, responses to clinical questions) |
The rapidly evolving capabilities of AI-related tools have laid the groundwork for transformational research. Indeed, AI has been thoroughly embraced by the medical research enterprise. Ten years ago, there were approximately 5000 AI-related publications in PubMed annually, increasing to 10 000 in 2018. This year, we are on track for more than 50 000 publications worldwide. Yet the potential for translating this information into clinically relevant insights and meaningful and equitable improvements in care delivery remains largely theoretical at present.
JAMA Internal Medicine is issuing a call for AI-related studies that address key priority areas. While we will consider papers across many domains of AI, we have a particular interest in studies providing novel evidence that will help to improve clinical outcomes, as well as studies addressing risk stratification and diagnosis and health system redesign. Finally, we are interested in work that addresses key regulatory, ethical, economic, and equity issues related to the evaluation and implementation of AI technologies in clinical care.
There are several notable illustrations of the capacity for AI to improve patient outcomes. These include a cognitive behavioral therapy intervention for chronic pain in which treatment plans were adjusted using AI-based feedback, an automated predictive model that decreased mortality and intensive care unit admission by identifying hospital inpatients at high risk of clinical deterioration, and an AI-enabled decision aid that significantly improved functional outcomes in patients with knee osteoarthritis.1–5 However, important evidence gaps remain. First, randomized clinical trials evaluating AI-based interventions are a rarity. Despite the rapid growth of AI-focused studies in PubMed, fewer than 1% are clinical trials. Further, many of the AI-related trials reported to date have small sample sizes and focus on diagnostic accuracy, rather than assessing the efficacy of an AI intervention in terms of patient-centered outcomes.6 Existing trials also have only limited reporting on patient demographics and operational efficiency, and most have not been replicated, raising concerns about the generalizability and applicability of results.
To help fill this knowledge gap, JAMA Internal Medicine is especially interested in high-quality randomized clinical trials that clearly define the AI-based intervention, fully describe the manner in which it was evaluated, and directly address patient outcomes. As described in our recent call for randomized trials, alternate randomization approaches such as cluster, stepped-wedge, and waitlist designs are also of interest, as are studies that use AI in creative ways to facilitate the conduct of trials (eg, by facilitating study participant selection and retention).7
AI-related approaches also hold substantial promise for improving both risk stratification and diagnosis—2 tasks fundamental to the practice of medicine. As a diagnostic tool, AI can recognize data patterns and novel variables that are not easily identifiable to human observers and can synthesize data to perform diagnostic reasoning tasks.8 JAMA Internal Medicine is interested in studies that develop and validate algorithms for diagnosis and risk prediction of conditions commonly encountered in the practice of internal medicine. For these studies, it is critical to validate models across diverse populations and settings using external datasets (ie, independent from those used for model development) performance.9 For studies of diagnostic testing, validation may mean evaluating whether diagnoses identified through AI are clinically meaningful and understanding whether early identification of disease states improves health. For all studies, the “black box” of the AI model should be made as transparent as possible, so that readers can understand what factors are being included and how they were ascertained. We also welcome studies that rigorously evaluate how predictive or diagnostic models affect patient outcomes, equity, and clinical practice once implemented.
AI-based tools have the potential to reshape clinical care delivery. These tools can improve clinical efficiency by decreasing administrative and documentation-related burdens, facilitating patient education and communication.10,11 In one study of 100 hospital encounters, large language model (LLM)–generated discharge summaries were found to be of comparable quality to those that were generated by physicians.12 AI tools can also help with quality improvement efforts, by automating the laborious process of reviewing medical records and incident reports and integrating tools to improve patient outcomes and enhance safety.13 On the other hand, despite enthusiasm for ambient notetaking, not all studies support their benefit, reinforcing the importance of empirical investigation. We are particularly interested in studies that assess the impact of AI on health care redesign in real-world clinical practice.
JAMA Internal Medicine is interested in submissions that address the fundamental ethical, legal, and social implications of AI. For example, the application of LLMs in the health care setting raises ethical concerns about patient privacy, accountability, transparency, and data provenance (eg, whose data are being used and under what restrictions). The trajectory of AI in medicine highlights how existing regulatory frameworks cannot keep up with the rapid pace of discovery and implementation. The US Food and Drug Administration has already approved over 1000 AI-enabled medical devices.14 Unlike traditional medical devices, many of these AI devices constantly refine and adapt themselves, based on new data. This ability to “learn” and evolve is one of the strengths of AI, but it also renders regulatory oversight far more difficult. A second challenge, particularly for LLMs, is that there is a lack of clarity about how the models reach their conclusions given their dynamic and iterative nature. Moving ahead, novel regulatory approaches will be needed to ensure efficacy and safety while balancing the interests and rights of patients. Importantly, it is critical to identify and prevent algorithmic bias, specifically to ensure that the adoption of AI into clinical care does not exacerbate long-standing disparities in health. This concept of fairness is crucial to ensure that AI tools provide accurate diagnosis and treatment recommendations across patient populations. Additionally, the economic impact of AI-related tools could have a profound impact on access and equity. On the one hand, AI could potentially increase access to care for underserved populations, by informing diagnostic and treatment decision points, particularly where data-driven algorithms are equivalent or superior to clinical judgement. Conversely, AI-related tools that confer additional costs could potentially exacerbate disparities.
The Editors of JAMA Internal Medicine are excited to invite the scientific community to contribute studies that will help shape the future of AI in medical and health care research, policy, and practice. We are most interested in studies that can directly inform efforts to improve patient outcomes and clinical care. We encourage authors to review recent author instructions regarding the presentation of AI-related studies in JAMA, and to incorporate specific guidance for reporting study methods and outcomes, such as the TRIPOD-AI for reporting clinical prediction rules that incorporate AI methods or CONSORT-AI for clinical trials of interventions.15–18 We encourage authors to consult the JAMA Internal Medicine Instructions for Authors for guidance about manuscript preparation and submission.19 There is an important new JAMA Network resource that will help our authors reach an even broader audience across the entire JAMA Network: the recently launched JAMA+ AI channel.20 This resource showcases studies from across the full family of JAMA Network journals, supplementing this content with new multimedia materials, author interviews, and educational materials.
JAMA Internal Medicine aspires to publish practice-changing research in the burgeoning field of AI, and we invite our authors and readers to join us on this journey. Designing, conducting, and evaluating AI research will require expertise in computer science, clinical trial design, medical informatics, health equity, quality and outcomes research, medical ethics, health economics, and health policy. We need all hands on deck to harness AI to advance health and health care.
