As predictive models proliferate, providers and decision makers require accessible information to guide their use. Preventing and combating bias must also be priorities in model development and in communication with providers and decision makers.
Objectives: As predictive analytics are increasingly used and developed by health care systems, recognition of the threat posed by bias has grown along with concerns about how providers can make informed decisions related to predictive models. To facilitate informed decision-making around the use of these models and limit the reification of bias, this study aimed to (1) identify user requirements for informed decision-making and utilization of predictive models and (2) anticipate and reflect equity concerns in the information provided about models.
Study Design: Qualitative analysis of user-centered design (n = 46) and expert interviews (n = 10).
Methods: We conducted a user-centered design study at an academic medical center with clinicians and stakeholders to identify informational elements required for decision-making related to predictive models with a product information label prototype. We also conducted equity-focused interviews with experts to extend the user design study and anticipate the ways in which models could interact with or reflect structural inequity.
Results: Four key informational elements were reported as necessary for informed decision-making and confidence in the use of predictive models: information on (1) model developers and users, (2) methodology, (3) peer review and model updates, and (4) population validation. In subsequent expert interviews, equity-related concerns included the purpose or application of a model and its relationship to structural inequity.
Conclusions: Health systems should provide key information about predictive models to clinicians and other users to facilitate informed decision-making about the use of these models. Implementation efforts should also expand to routinely incorporate equity considerations from inception through the model development process.
Am J Manag Care. 2022;28(1):18-24. https://doi.org/10.37765/ajmc.2022.88812
Communicating effectively about predictive models and their equity implications is critical for (1) informed decision-making and (2) preventing potential harms to patients.
Predictive models continue to gain attention and investment in the US health care system.1-3 By leveraging large quantities of data from electronic health records (EHRs), clinical research, and other sources, these models can be used to anticipate patient outcomes such as disease progression, predict health service utilization, or identify appropriate treatments.3-6 In addition to commercial vendors, health systems are building predictive analytics tools and using them for clinical decision support and administrative purposes.7-10 Although some promising applications of predictive models have been identified, such as sepsis prediction,11-14 the evidence base remains fragmented and marked by both methodological and implementation issues.5,15-17 Furthermore, regulation of predictive models is minimal, with only limited oversight of particular diagnostic models by the FDA.18 In the context of little regulatory oversight and a highly varied body of evidence, the risk of exacerbating existing inequalities or creating new types of inequity is high.
Because predictive models are built on historic data from the health care system, they reflect and can intensify long-standing racism and inequities in health care.19 Bias in these models has been documented, discussed, and conceptualized in a growing body of literature.20-23 In a well-known example, an algorithm used to allocate care management resources systematically discriminated against Black patients. The model used health expenditures as a proxy for health conditions. At any given risk score used to indicate predicted need, Black patients were sicker than their White counterparts who received the same score, resulting in comparatively healthier White patients receiving additional resources.24 There are many historical and contemporary racial inequities in health care expenditures, quality of care, and aggressive interventions that disadvantage Black patients.19,24 Reflecting these inequities, the algorithm underestimated the health needs of Black patients. Similar patterns have been identified in predictions related to kidney disease, whereby some models have contributed to or worsened racial inequity by building barriers to treatment for Black patients.25,26 These examples highlight cases in which harms have been identified post facto, in part because they did not account for inequity prior to model development. They also demonstrate how high the stakes of predictive models in health care are.
Models used to manage many health conditions are similarly reflective of social context and inequity. For example, hepatitis C is curable but very costly to treat. Although predictive models can be helpful in anticipating disease progression and targeting financial resources, there are serious equity concerns related to access to treatment.27 Hepatitis C disproportionately affects racial and ethnic minorities, and White patients are more likely to receive treatment.28 Given this context, it is important to vet models and account for existing inequity prior to their use.
Model auditing and analysis of potentially unequal outcomes have been recommended as potential safeguards against these harms.23 Although such steps could be highly useful, they generally take place and identify bias post implementation.24,29,30 To anticipate bias and inequity before they cause harm, social context and equity need to inform predictive models from inception through utilization.31 Traditional approaches to building and using predictive analytics are limited in this capacity. These efforts tend to fall into silos of expertise such that experts in equity or social sciences, who might anticipate the “unintended consequences” of predictive models, are not regularly included in their development.32
It is critically important that inequity is considered from inception and that clinicians are provided necessary information about predictive models to make informed decisions about their use.33 Much of the work on this topic to date has been focused on concepts and commentary. Although this literature is foundational to our understanding of predictive models, there is a dearth of related empirical analysis in health care. This exploratory study aimed to contribute to developing the literature by understanding how stakeholders assess the value of these models, identifying ways to facilitate informed decision-making about their use, and building equity into basic communication about the models from inception. To accomplish this, we applied user-centered design to develop a product information label that would help clinicians and administrators evaluate predictive models. We conducted follow-up expert interviews to investigate the equity implications inherent in the development and use of predictive analytics and to offer an alternative to the current approach to model development.
Phase 1: Eliciting User Needs Related to Predictive Model Use
In phase 1 of this project, our design team followed the double diamond method to identify the needs and priorities of predictive model users. This is a type of user-centered design method widely deployed for information technology (IT) development that is characterized by iterative development cycles that produce a solution or prototype that responds to user needs.34 In this study, the double diamond process was used to (1) define problems faced by clinicians and administrators in the use of predictive models, (2) develop a prototype product information label for a predictive model, and (3) elicit reactions from users.34
The first step of the double diamond method was a needs assessment.35 Interviews were conducted with 46 clinicians, administrators, model developers, IT staff, and project managers at a large academic medical institution with predictive analytics infrastructure. We interviewed clinicians from a variety of specialties, including emergency medicine, pediatrics, general medicine, oncology, and primary care. The semistructured interviews sought to understand participants’ needs, concerns, and preferences related to a hypothetical hepatitis C model prototype that predicted serious illness to allocate treatment resources. A prototype product information label was developed and updated iteratively through rapid prototyping as interviews were conducted.35 Participant needs and concerns informed the prototype design, which was then used to validate and expand on the findings from earlier interviews.34 Participant quotes were also used to construct analytic matrices for validation and comparison of interviewees’ priorities and concerns.
Equity and bias were raised in these interviews, although they were not central to the priorities of interviewed stakeholders. It remained unclear how communication about predictive models could better incorporate information related to equity. To better understand how equity could be incorporated and addressed in a product information label, we continued the double diamond method with a second phase of the research focused specifically on these topics.
Phase 2: Expert Interviews on Equity Implications
Building on the findings in phase 1 and recognizing the pressing concern of inequity in health care, the second phase of the study included 10 interviews with investigators and clinicians specializing in predictive models, clinical care, and health equity at the same institution who did not participate in phase 1 interviews. Using the hypothetical hepatitis C case developed in phase 1, these semistructured interviews focused on predictive models, high-cost treatment, and equity to better understand the ways that predictive models can either exacerbate or reflect inequity. All interviews were recorded and transcribed professionally. The rigorous and accelerated data reduction (RADaR) technique was used to produce concise data tables and identify themes.36 Using this approach, interview content is reduced through successive iterations of coding to identify and refine codes and produce core insights from interview content. One author (P.N.) conducted the RADaR process with regular review, input, discussion, and oversight from the senior author (J.P.). Text reduction was reviewed and any thematic coding issues were resolved during regular analysis meetings. Phase 2 interviews informed an updated product information label that communicated key equity concerns to potential users.
Trust in predictive models and willingness to use them depend on characteristics of and communication about the models.
Four key informational elements required for informed decision-making emerged from the phase 1 interviews with clinicians, administrators, and model developers (Table 1). These were (1) information about the developers and users of a given model, (2) the methodology with which it was developed, (3) the extent to which it has been peer reviewed and updated, and (4) the populations in which it has been validated. These types of information were necessary for evaluation and confidence in a given predictive model, especially among clinicians.
The informational elements required for informed engagement with a model, and potential use, are included in the product information label prototype (Figure 1).
When viewing the product information label prototype, interviewees searched for the developer and other users of a model. Specific trusted institutions, such as leading academic medical centers or professional societies, were critical for trust in a predictive model. Utilization of the model by well-known institutions was seen as evidence that it was vetted and reliable. Relatedly, peer review was an important mechanism of verification. Because administrators and clinicians expressed that they might not have the time to evaluate specific statistical methods used in a model, they required evidence that trusted institutions, experts, or organizations had approved it.
Although nondevelopers expressed that they did not have the time or specialized knowledge necessary to evaluate specific methodological decisions like the use of random forest models, key methodological information was required for model evaluation. The data used to create the model, specific evidence underlying its function, and its original purpose were of particular interest. Clinicians were especially concerned that a given model was validated on their patient population, or a very similar one, to ensure that the model was appropriate for their patients.
Figure 2 includes the equity implications of phase 1 key informational elements, as well as 2 additional informational elements that emerged from the interviews: (1) purpose and application and (2) relationship to structural inequity.
Implications of Phase 1 Informational Elements
Experts reported that ethics and equity were often treated as separate from model development processes, which prioritized pragmatism and speed over ethical concerns. This separation of model development from the consideration of social context meant that equity was largely absent from predictive model development and implementation. Experts emphasized the risks to marginalized patients, including exacerbating existing inequities and building barriers to care or treatment. Accounting for and communicating these risks in model development and implementation were critically important, although there was considerable variation in opinions about where and with whom this responsibility should lie.
The lack of standards and safeguards against bias in predictive analytics was also a key threat to equity. Without clear requirements for updating and modifying a model, for example, bias cannot be systematically identified. Model development, methodology, and peer review/updates were especially affected by this lack of standards (Table 2). The implications include harms to marginalized patients, inability to systematically identify bias, and inadequate population validation on minority populations. Validation and calibration are not currently standard or required, meaning that the efficacy of a given model for subpopulations may be unclear, perpetuating inequality by advantaging patients who are already advantaged.
Additional Informational Elements
Two additional key informational elements emerged from our interviews with experts. Questions about the purpose and application of a given model were central to evaluation among experts. A model applied to anticipate high cost and limit access to care for patients, for example, would have fundamentally different implications than a model facilitating access to necessary treatment (see eAppendix [available at ajmc.com] for example quotes). Clarity about the issue to be addressed by a model was fundamentally important for experts considering a model’s value.
A model’s interaction with or reflection of structural inequity was also critically important. For example, multiple experts separately described biased algorithms related to kidney disease and cardiovascular function that exacerbated existing racial disparities. Because Black patients face structural barriers to health care, for example, diagnosis of kidney disease is often delayed. Black patients historically have been sicker by the time they receive a diagnosis, and a “race correction” reflecting this dynamic was built into an algorithm for future treatment. Knowledge of these social and historical dynamics and information about how a model engages with or addresses them was of primary concern. For the experts interviewed, clear communication about this aspect of a given model was necessary both for consideration in model development and for providers to understand the risks to their patients.
In this study, we identified information necessary for informed decision-making related to predictive models and explored equity implications with experts. Our findings focus on 2 domains for health systems to consider in the development and use of predictive models: critical informational elements required for users and foundational equity concerns to address.
In phase 1, administrators and clinicians described some of the informational elements required to evaluate and potentially use predictive models. In addition to demonstrating an understanding of foundational methodological concerns relevant to clinical care such as population validation, they also noted the challenge of evaluating every detail of the specific statistical methods. For some, it was necessary to rely on information about the use of models by trusted institutions, experts, or organizations that adopted a model. This was especially true when combined with prioritized methodological information affecting patient care. Our interviews also revealed the consistent desire on the part of clinicians to know if and how a model is validated on their patient population, or a very similar one, to ensure the appropriateness of the model. This informational element is important for health equity because homogenous training data sets lead to less accurate predictions for minority populations and marginalized groups.15,37
Implications for Trust and Utilization
Our study results underscore ways in which informed decision-making and equity matter in the use of predictive models. The confidence and trust of clinicians are required for utilization of predictive models, but they must also be earned through validated, trustworthy models and communication of key informational elements, some of which we identify here.38 For some, leveraging trust in scientific processes such as peer review and trust in organizations that are using specific models will inform whether a user will adopt a model, suggesting the importance of those organizations themselves being trusted and trustworthy agents.
Our findings also highlight the importance of communicating evidence that models are trustworthy and responsive to equity concerns. Health systems would be well served by establishing transparent policies related to population validation and communicating this critical informational element to clinicians and other users. Communication about predictive models also needs to include information and evidence on their relationships to structural inequity. Clarifying the purpose of a model in practice will facilitate critical engagement with potential negative consequences for patients. Individuals and systems considering use of these models should engage in discussions from early stages to identify any potential limitations of the models and could even work with health systems and designers to ensure equity of the model design and its subsequent use. These critical reflections, at a system level, could be important for improving the validity of models.
Alternative Approach to Development and Use of Predictive Models in Health Care
These study findings revealed some of the negative consequences of excluding expertise in equity from model development. Based on this outcome, we propose that perspectives of equity experts are incorporated throughout the development and implementation of predictive models and in the evaluation of and communication about predictive models in order to anticipate negative consequences of a given predictive tool. In the traditional approach, models are often developed and used in silos of expertise. After implementation, analysis of negative consequences may be conducted and bias can be identified. However, this means that patients have already been harmed. To better anticipate these consequences, an alternative approach (Figure 3) elicits and integrates the perspectives of a more diverse group of experts who are equipped to identify potential risks for marginalized patients. Future studies should expand on existing literature to examine specific outcomes and metrics that could be used to assess the equity implications of predictive models in health care. Patient perspectives should also be sought to understand perceived risks and benefits associated with predictive models.
This is an exploratory study based in a single academic institution.39 Our findings may not generalize, especially to smaller, nonacademic institutions that do not have similar model development capacity or familiarity with predictive models. Because this varies significantly and could be associated with specific EHR characteristics, future work should analyze the perspectives of various health systems.40 Further validation, evaluation, and expansion in other settings will be important in future work. However, information labels, such as the one prototyped in this project, are increasingly in demand in consumer and health care domains, and our work provides an initial application that should be further explored.
As the health care system continuously works toward evidence-based medicine, the evidence on bias and inequity must be included and integrated into predictive analytics. When this expertise or evidence is treated as secondary and unrelated to model development, the negative consequences of predictive models will remain “unintended” but preventable. Designing policy around model transparency, communication of key informational elements, and analysis of equity implications are important steps that health systems can take to facilitate informed decision-making and to prevent harm to patients. If models are used caveat emptor, this leaves patients and clinicians vulnerable. More extensive evaluation and communication about model functions will be critical. Formal oversight mechanisms within and across organizations that incorporate expertise on equity in the key informational elements about predictive models will guide users in their assessments of model value and facilitate more broadly informed model utilization.
The design team for phase 1 of this study included Jade Crump, Katherine Jones, Sarang Modi, and Sam Bertin. They designed the product information label prototypes and elicited responses from interviewees. The authors also want to acknowledge Tevah Platt and Drs Marie Grace Trinidad, Daniel Thiel, and Sharon Kardia for their input and contributions to the research project.
Author Affiliations: Department of Health Management and Policy (PN) and Department of Learning Health Sciences (JP), University of Michigan, Ann Arbor, MI; Department of Kinesiology and Community Health, University of Illinois at Urbana Champaign (MR), Champaign, IL.
Source of Funding: Knowledge to Treatment Optimization Program through the Michigan Department of Health and Human Services.
Author Disclosures: The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.
Authorship Information: Concept and design (PN, JP); acquisition of data (PN, MR, JP); analysis and interpretation of data (PN, MR); drafting of the manuscript (PN); critical revision of the manuscript for important intellectual content (PN, MR, JP); obtaining funding (JP); administrative, technical, or logistic support (PN); and supervision (JP).
Address Correspondence to: Paige Nong, BA, Department of Health Management and Policy, University of Michigan, 1415 Washington Heights, Ann Arbor, MI 48109. Email: firstname.lastname@example.org.
1. Wang F, Preininger A. AI in health: state of the art, challenges, and future directions. Yearb Med Inform. 2019;28(1):16-26. doi:10.1055/s-0039-1677908
2. Naylor CD. On the prospects for a (deep) learning health care system. JAMA. 2018;320(11):1099-1100. doi:10.1001/jama.2018.11103
3. Ngiam KY, Khor IW. Big data and machine learning algorithms for health-care delivery. Lancet Oncol. 2019;20(5):e262-e273. doi:10.1016/S1470-2045(19)30149-4
4. Alanazi HO, Abdullah AH, Qureshi KN. A critical review for developing accurate and dynamic predictive models using machine learning methods in medicine and health care. J Med Syst. 2017;41(4):69. doi:10.1007/s10916-017-0715-6
5. Futoma J, Morris J, Lucas J. A comparison of models for predicting early hospital readmissions. J Biomed Inform. 2015;56:229-238. doi:10.1016/j.jbi.2015.05.016
6. Singh K, Valley TS, Tang S, et al. Validating a widely implemented deterioration index model among hospitalized COVID-19 patients. medRxiv. Preprint posted online April 29, 2020. doi:10.1101/2020.04.24.20079012
7. Bhardwaj R, Nambiar AR, Dutta D. A study of machine learning in healthcare. In: 2017 IEEE 41st Annual Computer Software and Applications Conference (COMPSAC). Institute of Electrical and Electronics Engineers; 2017:236-241. doi:10.1109/COMPSAC.2017.164
8. Tang C, Lorenzi N, Harle CA, Zhou X, Chen Y. Interactive systems for patient-centered care to enhance patient engagement. J Am Med Inform Assoc. 2016;23(1):2-4. doi:10.1093/jamia/ocv198
9. Rosella LC, Kornas K, Yao Z, et al. Predicting high health care resource utilization in a single-payer public health care system: development and validation of the high resource user population risk tool. Med Care. 2018;56(10):e61-e69. doi:10.1097/MLR.0000000000000837
10. Liu W, Stansbury C, Singh K, et al. Predicting 30-day hospital readmissions using artificial neural networks with medical code embedding. PLoS One. 2020;15(4):e0221606. doi:10.1371/journal.pone.0221606
11. Singh K, Betensky RA, Wright A, Curhan GC, Bates DW, Waikar SS. A concept-wide association study of clinical notes to discover new predictors of kidney failure. Clin J Am Soc Nephrol. 2016;11(12):2150-2158. doi:10.2215/CJN.02420316
12. Chen M, Tan X, Padman R. Social determinants of health in electronic health records and their impact on analysis and risk prediction: a systematic review. J Am Med Inform Assoc. 2020;27(11):1764-1773. doi:10.1093/jamia/ocaa143
13. Goldstein BA, Navar AM, Pencina MJ, Ioannidis JPA. Opportunities and challenges in developing risk prediction models with electronic health records data: a systematic review. J Am Med Inform Assoc. 2017;24(1):198-208. doi:10.1093/jamia/ocw042
14. Nemati S, Holder A, Razmi F, Stanley MD, Clifford GD, Buchman TG. An interpretable machine learning model for accurate prediction of sepsis in the ICU. Crit Care Med. 2018;46(4):547-553. doi:10.1097/CCM.0000000000002936
15. Barda N, Yona G, Rothblum GN, et al. Addressing bias in prediction models by improving subpopulation calibration. J Am Med Inform Assoc. 2021;28(3):549-558. doi:10.1093/jamia/ocaa283
16. Holmberg L, Vickers A. Evaluation of prediction models for decision-making: beyond calibration and discrimination. PLoS Med. 2013;10(7):e1001491. doi:10.1371/journal.pmed.1001491
17. Wynants L, Van Calster B, Collins GS, et al. Prediction models for diagnosis and prognosis of Covid-19: systematic review and critical appraisal. BMJ. 2020;369:m1328. doi:10.1136/bmj.m1328
18. Software as a medical device (SaMD). FDA. Updated December 4, 2018. Accessed March 28, 2021. https://www.fda.gov/medical-devices/digital-health-center-excellence/software-medical-device-samd
19. Benjamin R. Assessing risk, automating racism. Science. 2019;366(6464):421-422. doi:10.1126/science.aaz3873
20. McCradden MD, Joshi S, Mazwi M, Anderson JA. Ethical limitations of algorithmic fairness solutions in health care machine learning. Lancet Digit Health. 2020;2(5):e221-e223. doi:10.1016/S2589-7500(20)30065-0
21. McCradden MD, Joshi S, Anderson JA, Mazwi M, Goldenberg A, Zlotnik Shaul R. Patient safety and quality improvement: ethical principles for a regulatory approach to bias in healthcare machine learning. J Am Med Inform Assoc. 2020;27(12):2024-2027. doi:10.1093/jamia/ocaa085
22. Ferryman K. Addressing health disparities in the Food and Drug Administration’s artificial intelligence and machine learning regulatory framework. J Am Med Inform Assoc. 2020;27(12):2016-2019. doi:10.1093/jamia/ocaa133
23. Veinot TC, Mitchell H, Ancker JS. Good intentions are not enough: how informatics interventions can worsen inequality. J Am Med Inform Assoc. 2018;25(8):1080-1088. doi:10.1093/jamia/ocy052
24. Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447-453. doi:10.1126/science.aax2342
25. Braun L, Wentz A, Baker R, Richardson E, Tsai J. Racialized algorithms for kidney function: erasing social experience. Soc Sci Med. 2021;268:113548. doi:10.1016/j.socscimed.2020.113548
26. Vyas DA, Eisenstein LG, Jones DS. Hidden in plain sight — reconsidering the use of race correction in clinical algorithms. N Engl J Med. 2020;383(9):874-882. doi:10.1056/NEJMms2004740
27. Konerman MA, Beste LA, Van T, et al. Machine learning models to predict disease progression among veterans with hepatitis C virus. PLoS One. 2019;14(1):e0208141. doi:10.1371/journal.pone.0208141
28. Vutien P, Hoang J, Brooks L Jr, Nguyen NH, Nguyen MH. Racial disparities in treatment rates for chronic hepatitis C. Medicine (Baltimore). 2016;95(22):e3719. doi:10.1097/MD.0000000000003719
29. Harrison MI, Koppel R, Bar-Lev S. Unintended consequences of information technologies in health care—an interactive sociotechnical analysis. J Am Med Inform Assoc. 2007;14(5):542-549. doi:10.1197/jamia.M2384
30. Murray SG, Wachter RM, Cucina RJ. Discrimination by artificial intelligence in a commercial electronic health record—a case study. Health Affairs. January 31, 2020. Accessed November 21, 2021. https://www.healthaffairs.org/do/10.1377/hblog20200128.626576/full/
31. Rajkomar A, Hardt M, Howell MD, Corrado G, Chin MH. Ensuring fairness in machine learning to advance health equity. Ann Intern Med. 2018;169(12):866-872. doi:10.7326/M18-1990
32. d’Aquin M, Troullinou P, O’Connor NE, Cullen A, Faller G, Holden L. Towards an “ethics by design” methodology for AI research projects. In: Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society. Association for Computing Machinery; 2018:54-59. doi:10.1145/3278721.3278765
33. Sendak MP, Gao M, Brajer N, Balu S. Presenting machine learning model information to clinical end users with model facts labels. NPJ Digit Med. 2020;3:41. doi:10.1038/s41746-020-0253-3
34. What is the framework for innovation? Design Council’s evolved Double Diamond. Design Council. 2015. Accessed August 11, 2021. https://www.designcouncil.org.uk/news-opinion/what-framework-innovation-design-councils-evolved-double-diamond
35. Kinzie MB, Cohn WF, Julian MF, Knaus WA. A user-centered model for web site design: needs assessment, user interface design, and rapid prototyping. J Am Med Inform Assoc. 2002;9(4):320-330. doi:10.1197/jamia.m0822
36. Watkins DC. Rapid and rigorous qualitative data analysis: the “RADaR” technique for applied research. Int J Qual Methods. 2017;16:1-9. doi:10.1177/1609406917712131
37. Kim MP, Ghorbani A, Zou J. Multiaccuracy: black-box post-processing for fairness in classification. In: Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society. Association for Computing Machinery; 2019:247-254. doi:10.1145/3306618.3314287
38. Asan O, Bayrak AE, Choudhury A. Artificial intelligence and human trust in healthcare: focus on clinicians. J Med Internet Res. 2020;22(6):e15154. doi:10.2196/15154
39. Benda NC, Novak LL, Reale C, Ancker JS. Trust in AI: why we should be designing for APPROPRIATE reliance. J Am Med Inform Assoc. Published online November 2, 2021. doi:10.1093/jamia/ocab238
40. Apathy NC, Holmgren AJ, Adler-Milstein J. A decade post-HITECH: critical access hospitals have electronic health records but struggle to keep up with other advanced functions. J Am Med Inform Assoc. 2021;28(9):1947-1954. doi:10.1093/jamia/ocab102