The imperative of applying ethical perspectives to biomedical and health informatics

In this editorial, I highlight 5 papers with consideration of ethical perspectives. In the first, the authors of an AMIA position paper provide a set of guiding principles related to selection of venues for AMIA conferences and events using an ethical perspective that delineates a set of rights for AMIA members and associated institutional obligations for AMIA.¹ Subsequently, I apply a public health ethics framework comprising 7 principles² to the remaining 4 papers that address critical topics in biomedical and health informatics including bias in models, particularly artificial intelligence (AI) models,^3,4 and protection of individual privacy.^5,6 The public health ethics principles include well-established bioethical obligations related to non-maleficence (do no harm), beneficence (produce benefit), respect for autonomy (eg, informed consent, privacy), and justice (provide equal opportunity to benefit) as well as 3 additional principles. Population health maximization is the obligation to maximize the health of the population versus the individual. Efficiency is the obligation to use scarce resources efficiently. Proportionality involves the weighing and balancing individual freedom against wider social goods in a proportionate way.

The authors of an AMIA position paper delineate guiding principles and a member-informed process for selection of venues for AMIA conferences that communicate the institutional values of its membership.¹ The guiding principles are framed as commitments to 8 principles with required obligations for AMIA.

1. Right to benefit from science: Conduct meetings that permit and enable attendees to congregate, exchange, and debate scientific information, and to enjoy the right to benefit from science that is free from bias.

2. Right to safety and security: Ensure that attending the conference does not put the average attendee at risk of substantial harm compared to other potential venues.

3. Freedom to travel: Exclude venues that would restrict travel or participation based on factors such as religion, gender, sexual orientation, disability, race, ethnicity among others.

4. Freedom of speech: Ensure that attendees can share their scientific results and conclusions in a free and unfettered manner without censorship.

5. Right to nondiscrimination and civil discourse: Provide information about how the AMIA will assess allegations and the consequences for those who are found to violate the expectations of behavior.

6. Human rights: Exclude locations and venues with laws, policies, practices, or rules that undermine that commitment or allow violations of human rights (eg, discrimination, arbitrary arrest, or violation of privacy).

7. Access to professional development: Ensure adequate space for optimal learning, a variety of settings that allow for different size groups, and accommodation of an AMIA plenary session.

8. Transparency and veracity: Ensure a transparent process of decision-making that is free from conflict of interest and visible to AMIA members.

These principles will guide selection of venues for AMIA conferences and events. In addition, AMIA reserves the rights to request in writing the principles of potential venue vendors to assure that they align with AMIA principles.

Wang et al³ developed a bias evaluation checklist that enables model developers and health care providers to systematically appraise a model’s potential to introduce bias and illustrated application of the checklist to assess 4 30-day hospital readmission prediction models. Developed by a team of experts in machine learning, health services research, health disparities, and informatics, the checklist comprises 3 steps for the potential user of the algorithm: (1) definition of what the model predicts and how it should be used; (2) location of evidence supporting algorithm efficacy; and (3) answering 11 questions to identify 6 sources of potential bias: label bias, modeling bias, population bias, measurement bias, missing validation bias, and human use bias. The assessment of 4 models demonstrated the utility of the checklist identify variety of ways in which algorithms can perpetuate healthcare inequalities. While illustrated with 30-day hospital readmission prediction models, the checklist is designed for broader application. Through their evaluation checklist, the authors illustrate support for the obligations of beneficence, non-maleficence, justice, and efficiency.

Estri et al⁴ developed an objective framework for evaluating unrecognized bias in medical AI models predicting coronavirus disease 2019 (COVID-19) outcomes. They evaluated 4 AI models developed to predict risks of hospital admission, intensive care unit admission, mechanical ventilation, and death after a severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) infection using longitudinal electronic health records (2016 to 14 days before positive Polymerase Chain Reaction [PCR] test) of >56 000 Mass General Brigham patients. Considering an unbiased algorithm as one that reflects the same likelihood of the outcome regardless of an individual’s group membership, they used model-level metrics of discrimination, accuracy, and reliability, and a novel individual-level metric for error to assess biases related to factors such as race, ethnicity, gender, age, and time. The authors found inconsistent model-level bias in the prediction models but that from the individual-level perspective, most models performed with higher error rates for older patients. Objective approaches such as that proposed provide a foundation for reducing bias in AI models. In their approach, the authors attend to the obligations of non-maleficence, beneficence, and justice. In addition, by developing an objective approach for evaluating bias, they are promoting efficient use of scare resources.

Thomas et al⁵ on behalf of the National COVID Cohort Collaborative (N3C) evaluated the extent to which synthetic data from the N3C dataset could be used for geospatial and temporal epidemic analyses. They compared the original dataset (n = 1 854 968 SARS-CoV-2 tests) and its synthetic derivative on indicators of COVID-19 spread including aggregate and zip code-level epidemic curves, patient characteristics and outcomes, distribution of tests by zip code, and indicator counts stratified by month and zip code. In general, synthetic data closely matched original data for epidemic curves, patient characteristics, and outcomes supporting the utility of the synthetic data. However, utility was diminished for groups with small samples such as some patient subpopulations and zip codes with few tests due to purposeful data suppression to prevent attribute disclosure. The authors illustrate the obligation of efficiency by evaluating the utility of their synthetic data approach. This approach supports respect for autonomy by suppression of data that reflect small samples (non-maleficence) but consequently limits the opportunity for some populations to benefit.

Luo et al⁶ developed a lossless distributed algorithm for generalized linear mixed model analysis and applied it to the task of data sharing and comparisons across hospitals. The novel distributed penalized quasi-likelihood (dPQL) algorithm uses aggregated rather than individual patient data. They demonstrated its applicability by ranking 929 hospitals for COVID-19 mortality or referral to hospice. Moreover, they provided mathematical evidence that the dPQL algorithm is lossless, that is, it obtains identical results to pooling of individual patient data from all hospitals. This approach supports the obligations of respect for autonomy and efficiency.

Application of ethical perspectives is essential to advancing discovery and application in biomedical and health informatics in a way that promotes diversity, equity, and inclusion—key values of AMIA and of the Journal of the American Medical Informatics Association. In this third year of the COVID-19 pandemic, I remain convinced that the field of biomedical and health informatics is critical to maximizing population health. Moreover, the need to weigh and balance individual freedom against wider social goods (ie, proportionality) in our informatics innovations has gained prominence and should be carefully considered in design and implementation.

Conflict of interest statement

None declared.