Keywords
Artificial Intelligence - Big Data - Bioethics - Clinical Decision Support Systems
- Public Policy
Introduction
Ethics is a kind of lens we use to identify issues and a lever used to formulate and
motivate best practices. Applied ethics is a tool to employ widely shared if not universal
values to contemporary questions and challenges in science and the professions. Bioethics
is the branch of applied ethics that addresses issues in the health professions; it
is often linked to other kinds of applied ethics, including business ethics, computer
ethics, government ethics, and so on.
It has become commonplace to observe or argue that science usually or even always
outpaces, or advances, more swiftly than applied ethics. In the case ofhealth informatics,
this is a mistake. For some four decades, albeit with some exceptions, advances in
biomedical informatics have been matched step for step by scholars who have identified
and addressed the ethical, legal, and social issues (ELSI) raised by the expansion
of a new science. This acronym, borrowed from the Human Genome Project, has ably served
the informatics community as a label or guidepost for research and pedagogy.
Though bioethics has moved forward, the same cannot be said for the law which continues
to lag as a source of official governance and oversight in health informatics and
other domains. It might be that the relationship between ethics and public policy
represents the greatest challenge faced by health informatics and the society health
informaticians seek to serve.
What follows from these two observations is this: we have an extraordinary opportunity
at a crucial time to try to ensure that the insights and analyses provided by ethics
continue to mature and, as important, that they are taken up and incorporated by academic
and health care institutions, businesses, professional organizations, and governments.
In what follows, I expand on these points by filtering them through a number of contemporary
challenges. These include artificial intelligence and machine learning; Big Data,
data sharing and privacy; duties to use and manage new technology; and ethics and
public policy.
1 Artificial Intelligence, Machine Learning, and Ethics
1 Artificial Intelligence, Machine Learning, and Ethics
The ancient fantasy of an intelligent machine or a smart homunculus became a research
project in the 17th century when Gottfried Leibniz, the philosopher and logician who, with Newton, discovered
the infinitesimal calculus, suggested that human reason could be rendered in a universal
language such that argumentation could be reduced to calculation. He built a primitive
calculator[1], arguably the first machine to replicate an aspect of human thinking. Leibniz thought
that intelligent machines would formalize reason and end disagreements. They did not.
During the greatest disagreement in the history of civilization - World War II - code-breaking
machines represented nontrivial instantiations of Leibniz’ aspiration: “The first
digital, electronic and programmable computer was developed as an instrument of warcraft.
The Colossus was a room-sized collection of racks, pulleys, wires and some 2,400 bottle-sized
vacuum tubes built at Britain's Bletchley Park to decipher encrypted German messages....
It became operational in 1944 and was used to prepare the D-Day invasion of Normandy.
One could argue that it eventually saved more lives than most medical inventions”[2].
However, we want health informatics to save and improve lives, to reduce suffering,
to help to achieve the larger goals of the healthcare professions. If an intelligent
machine can help do that, why is there a problem or any controversy?
Indeed, there are several problems, and their identification has contributed to both
greater understanding and to a period of overheated handwringing.
Key and core ethical issues in the development and use of machine learning programs
have already been identified repeatedly and elsewhere. They can be framed as lessons
learned or as recommendations. Some are lessons for developers, some for users. Though
the context here is health and health care, the lessons may be useful in many other
domains.
1.1 Quality and Standards Are Ethical Issues
Good software conforms to certain standards for quality which can be assessed in terms
of trustworthiness and reproducibility. The mark or measure of good software will
include accuracy of documentation and transparency about the provenance or source
of any code components. This is no mere courtesy to code-writing colleagues - it is
an auditable track record of what code is intended to do and how it was modified along
the way. This facilitates understanding, corrections, and improvements. It follows
that if one is developing or modifying machine learning software, the automated learning
process itself must also be monitored and documented. Relatedly, careful software
version control is an essential part ofhigh-quali-ty programming. The values of transparency,
veracity, and accountability reinforce the connection among quality, standards, and
ethics[3]. This has long been true across the health professions.
1.2 Prevent and Eliminate Bias
A sure way to erode confidence in artificial intelligence is to identify ways in which
a deep learning algorithm embeds racial, ethnic, gender, or other biases which shape
or corrupt its results. The very nature of machine learning algorithms makes plain
that one might unintentionally develop a biased system or accept biased results. Though
intent matters in ethics - many actions are praiseworthy or blameworthy precisely
because of someone's intent - this problem makes equally clear that failure to attend
to and correct or mitigate a problem can also be blameworthy; that failure to do so
is a kind of negligence (no surgeon ever intends to cut off the wrong arm or operate
on the wrong patient). If datasets used for training machine learning algorithms include
or entail bias or foster biased interpretations, then careful scrutiny of such sets
is a difficult and labor-intensive approach to take. Another is careful screening
of output to identify and filter bias, and perhaps to identify ways to modify the
algorithm to suppress it. It is now even possible to incorporate anti-bias features
into the code itself[4]. This is an extraordinarily promising approach, and something like it might very
well be the best way to ensure that public health surveillance and prediction, disease
diagnosis and treatment, and health policy are not infected or corrupted by illicit
bias.
1.3 Use Machine Learning Software for Good and not Evil
Current relativist trends and fashions notwithstanding, some actions are good and
others are bad and there is little credible dispute about their moral status. Reducing
suffering, eliminating disparities, and improving health are good; depriving people
of rights, using people for political or economic purposes without permission given
voluntarily, and harming people for profit are bad. If we start there, great progress
can be made. It is the hard cases, or cases on which reasonable people disagree, that
pose the most significant challenges. Is there a chance that a clinical decision support
system will improve diagnosis, treatment, and prognosis and simultaneously erode confidence
in the clinician-patient relationship? May a nurse with a good computer system undertake
duties traditionally reserved for physicians? Will the use of intelligent machines
in health care cause the erosion of healthcare practitioner clinical skills? Addressing
these and other challenges successfully will require sustained research, education,
and debate. At the least, we will make great progress by acknowledging and grappling
with such challenges.
1.4 Insist on and Provide Robust Education and Evaluation
No technology will ever fulfill its promise or be used in an ethically optimized manner
unless we insist on and provide robust education and evaluation. Scientists and clinicians must not only be taught empirical methods and clinical
skills, but also the appropriate uses of these methods and skills. Given our concern
for appropriate uses of intelligent machines in healthcare, it follows we must also
identify appropriate users, and a core criterion for demonstrating such appropriateness
will be a user's fitness to use the tool. Making and acting on this assessment has
long been recognized as a key ethical challenge when computers are used in healthcare[5]. We are, moreover, in urgent need of a comprehensive ethical curriculum in and for
health informatics. We must also not lose sight of the importance of system (algorithm,
device) evaluation in the context of its actual use; this, too, is an obligation long
recognized[6]. Much more recently, system validation in real-world settings has been identified
as a duty for anyone who would use an intelligent system in healthcare[7]. This may be especially true and important given the rapid progress in precision
medicine.
2 Big Data, Data Sharing, and Privacy
2 Big Data, Data Sharing, and Privacy
Empirical sciences have evolved and grown by using data and information to support
inferences which, if correct, come to constitute knowledge. In biology and medicine,
we use that knowledge to prevent, predict, mitigate, and cure maladies that sicken,
disable, or kill us.
A favorite example from the history of medicine of “calculating” to improve health
care is that of bloodletting, used for millennia to treat and often unintentionally
kill people who, under the humoural theory of disease, had too much blood. For instance,
George Washington one day in 1799 awoke with a sore throat, asked to be bled, lost
some five pints of blood, and was dead four days later, probably because of the bloodletting.
The work of Pierre Charles Alexandre Louis (1787-1872) was likely unknown to him.
Louis, a French physician, analyzed cases from several hospitals using his “numerical
method” and determined that bloodletting was not a cure but, rather, usually harmful.
He wrote: “As to different methods of treatment, it is possible for us to assure ourselves
of the superiority of one or other.. .by enquiring if the greater number of individuals
have been cured by one means than another. Here it is necessary to count. And it is,
in great part at least, because hitherto this method has not at all, or rarely been
employed, that the science of therapeutics is so uncertain”[8].
Necessary to count? Numerical method? Is it - can it really be - that simple?
Yes, sometimes. Observational science and basic arithmetic were, at first, the best
tools we had to support efforts to learn how the world worked. Observe and count (or
measure), find the mean, note changes over time, and then make inferences about physical
reality. An early form of statistics, frequency analysis, originated in the Middle
Ages with the Arab philosopher Al-Kindi as a code-breaking tool more than a millennium
before Colossus. His method also required counting - of letters. With an adequate
“plaintext…long enough to ill one sheet or so,” one could determine the frequency
of letters and combinations of letters as a baseline and use it to determine the values
of letters in a coded document[9]. It was a very early anticipation or foreshadowing of the data sets used to train
artificial intelligence programs.
What became clear in the 17th and 18th centuries was the utility of using data about observations to make predictions. In
the 1660s, John Graunt of London's “Natural and Political Observations Made upon the
Bills of Mortality” constituted the first actuarial tables and spurred efforts to
predict and control the causes of death. It is also the basis for modern insurance,
including the business of using statistics to sell health “coverage” to people in
need of health care in countries with dysfunctional health care systems.
The preceding part of this section embeds several important ethical issues. Chief
among them are these:
-
Are data reliable and who is responsible for ensuring reliability?
-
How exactly does the calculation work?
-
How should it be determined who should use these calculators, and for what purposes?
These, of course, parallel the “lessons learned” we earlier identified for appropriate
use of machine learning or any other medical software.
The datasets created by Al-Kindi, Louis, and Graunt were actually tiny; we might even
call this “small data.” Yet their work suggests the first grains or nuclei of a vastly
larger project. Framed very broadly, one might argue that the work of these three
data collector-analyzers raised ethical issues that parallel those we find when collecting
and analyzing some of the more than 2,000 exabytes of medical data in the world today
(an exabyte, recall, is 1,000 petabytes, with one petabyte being 1,000 terabytes,
and one terabyte being 1,000 gigabytes, etc.). But their work did not cause any ethical
consternation at the time; it was just science advancing.
Though we ought to regard privacy and confidentiality as universal rights, they are
neither absolute nor have they ever been regarded as absolute. There are tradeoffs
to be made: public health surveillance gives us better public health, hospital patient
monitoring gives us better hospital care, the sharing of research results and other
data gives us better medical treatments. Put differently: if a physician learns something
about effective practice from Patient Epsilon's response to a treatment, does the
physician violate Epsilon's privacy or confidentiality six months later by using what
she learned to help Patient Omega[10]? Indeed, such a reductio ad absurdum argument makes clear that not only does the physician not violate confidentiality
but also that she would be irresponsible not to use the information from one patient
to help another.
With very large datasets, however, our challenge is magnified. Until recently, the
ability to reconnect a datum with a person, or “re-identify” the datum, was either
impossible or difficult. Now it is comparatively easy, and this entails that the tools
of applied ethics must in a world of Big Data rise to the occasion. One way to put
this, according to my colleague Richard Bookman in a personal communication, is that
“Ethical boundaries move with scale. As little data goes to Big Data, onetime practices
at small scale may be rendered offensive and unacceptable at broader scale.”
The same could be said for data sharing. Few would argue that science progresses when
investigators communicate openly and generously with each other. It also progresses
when scientists are motivated by curiosity and service instead of profit and priority
- at least historically. When the financial stakes were low, history shows discovery
science was fertile and productive. That slowed when fame attached to empirical success
and reputations were made by being the first to make a discovery. We then commodified
science so that not only would a good scientist learn something, he could also make
a lot of money. In 1955, Jonas Salk famously disdained patenting his polio vaccine
(“Could you patent the sun?”); it was the same year the Dartmouth Workshop, at which
Artificial Intelligence as a field was fledged, was proposed. In 2020, one would be
regarded as a fool for not leveraging a discovery or method for profit. Big Data means
big profit, and the ethical boundaries have correspondingly shifted.
It would perhaps not be so difficult if the accumulation of wealth from others’ information
could rely on their concurrence and support. Worse, the erosive use of Big Data for
profit in business and industry has polluted public perceptions of all uses of Big
Data. We learned earlier that applied ethics can contribute to the identification
of appropriate uses of information technology. Here we have an opportunity to protect
fundamental privacy rights by making a distinction between using social media, for
instance, to track people's behavior to target them for political purposes, and using
electronic health records to analyze their data to improve clinical outcomes. There
is indeed evidence that ordinary people are prepared to accept this distinction, at
least if the user is trustworthy[3].
We have several ways to meet the privacy challenges posed by Big Data for health care.
All of them should be seen as trust-enhancing.
-
The ffirst is security backed up by law. Make it difficult for unauthorized - inappropriate - parties to
access or use the data, and punish them severely if they do. Constant and evolving
research improves privacy-protecting software. The 2018 IMIA Yearbook of Medical Informatics, for instance, included a special section (“Between Access and Privacy: Challenges
in Sharing Health Data”) that featured a suite of articles describing research devoted
to improving digital privacy protection[11]. It is clear that the good guys write better software than the bad guys. Use it.
-
Next, insist on accountable oversight. From bioinformatics laboratories to hospitals, those who collect, store, and analyze
health information are stewards of a precious resource. They must, absolutely must,
be held to trust-enhancing standards. The “trusted broker” model of scrutiny and oversight
can ensure both innovative research and the protections that are demanded by the sources
of health data and information, i.e., usually, patients[3]. Institutions and governments need to require and help implement mechanisms to review,
scrutinize, and approve collections and analyses of health data. This builds trust
and makes clear that those who use health data are motivated by values different than
those who scrape online clickstreams to beguile shoppers, those who vacuum social
media postings to monetize the intimate behavior of strangers, and those who weaponize
the Web to undermine democracies. They wanted, rather, to heal the sick.
-
Finally, educate lay people to improve digital health literacy, and teach scientists
and clinicians to learn to appreciate that the opportunity to study others’ information is a privilege.
If the former are willing to share data for better health and the latter are motivated
by higher values than those that govern the marketplace, then we reinforce the foundation
of a collective, international edifice that both protects legitimate privacy interests
while simultaneously ensuring enjoyment of the right to benefit from science[12].
It might be that the greatest utility of applied ethics is in the balancing of rights,
in part by the identification of responsibilities.
3 Duties to Use New Technology
3 Duties to Use New Technology
Information technology has transformed daily life, business, entertainment, and health
care. The list could, of course, go on, and on. The proper criterion to use in determining
if any new tool should be used is, generally, whether its advantages outweigh its
disadvantages and, if so, whether any antecedent rights are being violated. We learned
more than three decades ago that if health information technology could improve health,
then, ceteris paribus, this entailed a duty to use it for that purpose[5].
Such an insight does not of itself begin to answer the many questions that follow
- that is the job of scholars and analysts who scrutinize appropriate uses, privacy
implications, and governance. It nevertheless signals a powerful moral commitment
to reject basic Luddism, the movement that began in the 19th century with British workers smashing mechanical looms because the machines imperiled
their jobs. The best approach is to control machines, not destroy them.
One of the most promising and systematic approaches to achieve that control at scale
is the Learning Health System (LHS) paradigm[13], an evidence-building, safety-promoting system “capable of continuous self-study
and improvement”[14]. Such a system is a data-driven creature; it learns from every patient encounter,
every lab result, every outcome tracked and measured; it learns from public health
prevention and intervention. It captures Louis’ fabulous and simple understanding
that counting can count for a lot[8]. Moreover, LHS embodies core values of evidence-based practice, most especially
that it is neglectful and blameworthy to lose information that could have been used
to improve the wellbeing of individuals and the health of populations.
The LHS also provides an excellent example of and justifies what has been called the
“secondary use” of health information. In many respects, the term “secondary use”
is a misnomer: it implies that health information collected in a clinic, for instance,
is and ought to be collected for a primary purpose, and that any other purpose enjoys
lesser warrant. Similarly, it would be peculiar to collect public health information
and then regard that information's use in a clinic as secondary and requiring additional
permission. Rather, it would be irresponsible to acquire information in a clinic,
research study or public health surveillance effort and not use it to support one
of the others[15]
[16]
[17]. Widespread views and focused regulations regarding how best to protect these three
kinds of data and information are largely historical artifacts. Clinic data is protected
by privacy laws that evolved to protect, well, clinical data. Research data use is
governed by rules aimed at protecting human subjects, including their privacy. Public
health findings are, and always have been, acquired with the presumed or tacit consent
of those who benefit from - and trust - competent public health authorities.
Developers of Learning Health Systems (or “the” system if there is one and it is global)
are wise and quick to identify the values that reinforce it: person-focused, privacy,
inclusiveness, transparency, accessibility, adaptability, governance, cooperative
and participatory leadership, scientific integrity and value[14]. Indeed, one would be forgiven for suggesting that those values should underlie
any health system.
How such values are embraced will continue to challenge our best inclinations. The
effort to make the most of health data and information, however, is a collective recognition
of the overarching duty to use tools that improve health. Perhaps the most difficult
challenge is shaped by the values of inclusiveness and accessibility; that is, inclusiveness
and accessibility for whom?
Any duty to use a tool is cheapened or diluted if the tool benefits only those in
some jurisdictions or some socio-economic classes. If we intend a health system to
serve all, and if we know that social determinants are key etiologic sources of illness,
then perhaps a Learning Health System can be used to reduce disparities. Universalizing
access to health care, which has so far eluded most countries, would require far more
than data analysis - but such analysis in conjunction with tailored social commitment
might at the least effect some improvement. To be sure, such a course would require
the inclusion of social determinants, including poverty, in electronic health records
or other repositories of patient information. This in turn could cause or magnify
stigma, and it might be the case that some people would disdain being thus identiied.[1]
A robust attention to ethics and informatics cannot provide a script or formula for
righting wrongs. It can however usefully identify issues to be addressed, scrutinize
approaches to confronting those issues, and help point the way to the best possible
solutions. There is, as well, an element of professionalism in this regard, and major
informatics organizations have developed codes of ethics to guide those working in
the field to follow such precepts as help to secure the integrity of the profession[18]
[19].
4 Ethics, Standards, and Public Policy
4 Ethics, Standards, and Public Policy
Once we figure out how to get something right, it can be irrational to insist it should
be ignored. The evolution of standards in health care has improved quality, increased
safety, and conserved resources. This is especially true for health informatics. What
follows should be obvious: if continuously reined and improved standards can achieve
so much, and if those achievements improve human health and welfare, then there is
an ethical imperative to develop and improve them.
Any laws in the discovery and communication process - there are a lot of standards,
some conflict, and others are not very good - are remediable. As was once suggested,
“What is wanted here, in informatics, is in an important if vague respect what is
wanted everywhere in the world of standards: we want to be as good as we can be, along
with our efforts and tools, and to improve. That part of 'standard’ that implies regularity
or consistency captures the intuitions that order is more successful than chaos, and
that clarity and order produce goods and services that are easier to explain and share.
We have something very important to accomplish in health information technology, and
good standards are necessary for the process”[3].
Standards are also an important component of public policy and one of the ways by
which applied ethics informs and shapes public governance.
From the foundations of software engineering to the design of electronic health records
to embedded privacy protections to evaluation and interoperability, ethical principles
and standards serve as both guiderails and signposts. Is there a problem with biased
algorithms? Adopt standards for better software and testing. Do members of communities
distrust those who collect and analyze personal information? Follow standards for
trust enhancement. Confused about whether to adopt a new technology? Then turn to
ethical standards for harm reduction and rights protection.
As elsewhere in life, we sometimes in health informatics are over-eager to make problems
seem intractable or lose sight of the fact that we have available many tried-and-true
conceptual devices and publicly accountable processes for solving these problems.
To say, as some are wont, for instance, “privacy is dead; get over it,” is a sad surrender
to an ancient challenge tricked out with very large data bases and powerful computing
tools. Such surrender is neither necessary nor apt.
If we care enough about these issues, and we should, then applied ethics emerges as
the silent partner or loyal opposition to speak truth to a power whose benefits we
deserve and whose harms we must protest. When appropriate, in civil society, the law
itself can be a partner in ensuring that standards capture best practices. Ethics
should always come ffirst, both to guide us and to reduce the likelihood of health-system
laws that permit or even encourage unethical conduct.
Conclusion
Ordinary language loves the locution “ethical dilemma.” But to be on the horns of
an ethical or moral dilemma is to be in a position such that no matter what one does,
one does something wrong. Most ethical problems or challenges are not dilemmas, and
even the dilemmas, if they be such, can, as in logic, be escaped. The project here
has been to argue and so to make clear that the tools of applied ethics are adequate
to the task of guiding developers, users, and institutions as they adopt and try to
make the most of health information technology. This has been framed as statements
of issues and problems (machine learning, privacy, and duty to use the new tools)
and as suggestions for addressing them. As elsewhere in the health professions, the
ethical issues raised by a new technology are sources of interesting and significant
challenges. These challenges are difficult, but not intractable. Finding their solutions
presents opportunities to use critical thinking in the service of shared values. Chief
among these values is that of health itself. Producing better and ethically optimized
tools for healers is a contribution requiring the collaboration of the informatics
and ethics communities. Such collaboration affords a rare and unparalleled opportunity.
