Keywords
artificial intelligence - reproductive medicine - infertility - medical ethics - research
ethics
Introduction
Artificial intelligence (AI) is developing rapidly, and AI applications are being
implemented in various sectors, including several medical fields [11]
[22]. In medicine, complex AI algorithms can be used to analyze large amounts of data
with the aim of improving
diagnoses, prognoses and preventive measures [33]. In recent years, image data analysis has become a promising area of application,
and it
has even been suggested that the results generated by AI are superior to the contributions
of experts [44]. Moreover, AI is expected to
increase the efficacy of workflow processes in hospitals and improve patient monitoring
[55].
In reproductive medicine, new areas of application of AI-based methods are being explored.
Persons or couples who desire to have children but cannot conceive are often in a
life crisis and
have a reduced quality of life [66]. An inability to conceive may be viewed as an obstacle that frustrates the generally
accepted need to
reproduce and care for a child; this may acquire a psycho-existential dimension, as
the inability to have children threatens people’s existential vision of their future
[77]. Over 40 years ago, the birth of Louise Brown marked the start of a new era that
offered a new hope to infertile persons and couples, as
the possibility of achieving a successful pregnancy via in vitro fertilization (IVF)
became a reality [88]. According to some estimates, up
to 186 million people worldwide are affected by infertility [99]. The continued development and improvement of reproductive technologies,
which include the cryopreservation of oocytes and embryos, IVF, preimplantation genetic
diagnostics (PGD), and the technological option to select preimplantation embryos
based on morphokinetic
criteria have greatly increased clinical pregnancy rates over the past 40 years [1010].
However, several challenges persist. On average, the current probability in Germany
of giving birth after a fresh embryo transfer following IVF or intracytoplasmic sperm
injection (ICSI) is
approximately 24%; the probability when using previously cryopreserved and thawed
pronuclear-stage or multicellular-stage embryos is 20%. In essence, despite successful
fertilization and cell
division only every fourth fresh embryo transfer and every fifth cryo-transfer will
result in the live birth of an infant [1111]. The age
of the oocytes of the woman desiring to have children and the quality of the embryos
are the most crucial factors determining whether an IVF treatment would result in
pregnancy [1111]
[1212]. Reliable methods that allow a more precise assessment of the quality of
oocytes, sperm, and embryos are lacking. While the option of using preimplantation
diagnostics prior transferring the embryo to the uterus exists, this approach is ethically
controversial. In
Germany, it is legal only within certain limits, furthermore, it is technically complex,
expensive, and poses the risk of injuring or destroying the embryo [1313]
[1414].
The driving force behind the development of AI-based technologies in reproductive
medicine is the intent to improve the treatment and prognosis of infertile patients
by taking large amounts
of data and combining them to obtain meaningful results [1414]
[1515]. It is in this
context that our ethical discussion is relevant. The discussion is framed in part
by the needs of persons with an unfulfilled desire to have children, and in part by
the use of specific AI
technologies to achieve medical goals. The principles of healing, helping, and alleviating,
which can be subsumed under the general principle of beneficience, provide the ethical
framework for
this discussion [1616]. Extracorporeal fertilization and selective reproduction (e.g., after PGD) are still
considered controversial [1717]. Research and the potential integration of AI into clinical care raise complex normative
questions with regard to content and procedural
issues. Thus, a proactive ethical debate is necessary, even if AI research and the
use of AI in reproductive medicine are still in very early experimental stages.
Given this context, we will start by providing a brief outline of the current status
of the research and development in this field, followed by a structured ethical analysis
of the potential
uses of AI-based methods in reproductive medicine. After that, we will discuss the
impact of the relevant technologies on the relationship between physicians and their
patients. In the
conclusion, we will consider possible future developments in AI-driven reproductive
medicine and briefly address possible social challenges associated with these trends.
Possible Applications of AI in Reproductive Medicine
Possible Applications of AI in Reproductive Medicine
In Europe, pregnancy rates and live birth rates for all treatment options vary widely
between countries. In 2015, the pregnancy rate following fresh cycles after IVF or
ICSI was between 19.6%
and 44%, and live birth rates ranged from 10.2% to 40%. The rate of live births after
the transfer of frozen-thawed embryos ranged between 12.8% and 37.5% in different
European countries [1818]. Overall success rates decline sharply as women become older [1919]. A
retrospective analysis by Stoop et al. of live births after fresh and cryopreserved
embryo transfers showed that the average rate per harvested ripe oocyte was 4.47%
for women between the ages
of 23 and 37 [2020]. Therefore, it can be assumed that assisted reproductive medicine could potentially
be improved in several areas. Now
that AI systems are coming into more general use, it is hoped that the automatic classification
of sperm, embryos, and oocytes will be possible, thereby increasing the success rates
of IVF
[1414].
Basics of AI
The definition of AI in the context of this text is as following: using complex algorithms
to imitate logical thinking and cognitive functions. Several heterogeneous techniques
are included
under the umbrella term AI. Machine learning (ML) is a particularly successful AI
application. ML identifies patterns between variables in large datasets. Previously
unknown correlations can
be identified with ML and used to generate new hypotheses and pioneer new areas of
research [2121]. Most ML approaches can be classified
as either supervised or unsupervised approaches. With supervised ML, labelled training
data are used to develop models in which the target result (e.g., a diagnosis) is
known. In contrast,
unsupervised ML, does not require labeled data. Instead, patterns or aggregations
which occur within the data are recognized [2222]. Deep
learning (DL) is a variant of ML. DL attempts to imitate the functions of the human
brain, using different levels of artificial neural networks to generate automatic
predictions based on
training datasets [2121].
Despite several potential pitfalls, a promising clinical approach involves making
decisions for infertility patients based on an analysis of a variety of medical data.
Reproductive medicine
specialists can use ML models to help identify suitable treatments for persons struggling
with infertility [2222]. We present concrete
approaches below.
Selection and prediction of sperm cells
Sperm morphology is a common cause of (partial) infertility in men. Computer-assisted
sperm analysis systems are already in use. Nevertheless, the analysis of sperm motility
remains
difficult due to the clumping of spermatozoa and other contributing factors [2323]. Moreover, analyses may vary between laboratories.
Finally, in approximately one third of men, no clear aetiology can be identified [2424], which means that the cause of their infertility
cannot be identified with standard methods [1414]. In the future, automatic methods based on image may help obtaining more objective
and
precise results [1414]. Goodson et al., for example, used retrospective data from 425 human spermiograms
to develop models which
identified chromosomal anomalies. Size, total testicular volume, follicle-stimulating
hormone, luteinizing hormone, total testosterone and ejaculate volume were used as
the input data; based
on these data, the predictive accuracy for chromosomal anomalies was over 95% [2525]. Using data-mining methods, a research team also
developed two specific artificial neural networks to predict human sperm concentrations
and motility according to environmental factors and the men’s lifestyles [2626].
Evaluation and selection of oocytes
Successful reproduction, whether spontaneous or assisted, depends significantly on
the quality of the oocytes. However, the mechanisms of embryo malformations that develop
from oocytes of
insufficient quality are not known [2727]. Yanez et al. attempted to predict the developmental potential of human oocytes
by measuring
the viscoelastic properties of human zygotes just a few hours after fertilization
without destroying the zygote. They were able to reliably predict viability and blastocyst
formation with an
accuracy of over 90%, a specificity of 95% and a sensitivity of 75%. The researchers
also investigated RNA-sequencing data using a support-vector-machine classifier and found that
non-viable embryos had significantly different transcriptomes, particularly with regard
to the expression of genes important for oocyte maturation [2727]. Cavalera et al. observed mouse oocytes during in vitro maturation and took pictures
for time-lapse analysis; the obtained data were analyzed using artificial neural
networks. They were able to identify oocytes capable or incapable of further development
with an accuracy of 91.03% [2828].
Evaluation of embryo quality
Saeedi et al. presented the first automatic method to segment two main components
of human blastocysts, the trophectoderm (i.e., the external cell layer of the blastocyst
between the
4th and 7th day after fertilization), and the inner cell mass. The two regions are strongly textured
and the quality of their textures is quite similar. On imaging,
they often look as though they are connected to each other. The automatic identification
of both regions facilitates the detailed evaluation of blastocysts. Saeedi et al.
reported an
accuracy of 86.6% for the identification of the trophectoderm and 91.3% for the inner
cell mass. The aim of their study was to achieve a better understanding of why the
transfer of certain
embryos results in higher pregnancy rates than the transfer of other embryos [2929]. In a study published in 2019, researchers from
Cornell University trained a deep-learning algorithm from Google to recognize embryos
with high, moderate, or low development potential, measured according to the probability
of them
successfully implanting after intrauterine transfer. To train the artificial neural
networks, they used more than 10000 time-lapse images of human embryos. They additionally
developed a
decision tree using the clinical data of 2182 embryos, which combined the quality
of the embryo with the age of the patient and defined scenarios which are associated
with the probability of
becoming pregnant. Their analysis showed that the probability of becoming pregnant
following the transfer of an individual embryo ranged between 13.8% (age ≥ 41 and
poor quality) and 66.3%
(age < 37 and good quality); this correlated with the results of the automatic evaluation
of blastocyst quality and the age of the patient [3030].
Predictive models for IVF
Already in 1997, Kaufmann et al. developed a model to predict the probability of successful
IVF. For this, they used an artificial neural network with a predictive value of 59%.
The
research group used four criteria: age of the intended mother, number of oocytes obtained,
number of transferred embryos, and whether the embryos had previously been cryopreserved
[3131]. In a similar study conducted in 2010, the datasets of 250 patients were collected
from IVF research centers, fertility clinics and
maternity hospitals. To train and test the algorithm, the research group used significantly
more criteria (age of the woman, duration of infertility, body mass index [BMI], previous
pregnancies, previous operations, endometriosis, fallopian tube problems, ovulation
factor, sperm concentration, sperm vitality, number of aspirated oocytes, number of
transferred embryos,
prior history of miscarriage, and psychological factors) and achieved an accuracy
of 73% for their results [3232].
In a recent retrospective study, the researchers aimed to determine whether a simple
prognostic algorithm could differentiate between couples who require treatment for
infertility and
couples who can initially be offered less invasive strategies, the couples were divided
into groups according to their medical need for IVF treatment and their prognosis
for achieving a
natural conception. A Kaplan-Meier curve was generated for each group to measure the
probability of conceiving naturally over time and the effect of infertility treatment.
The outcome was
that for couples with slight or unexplained male infertility, the chance of live birth
without treatment and poor prognosis increases significantly, in comparison to couples
with a good
prognosis. This prediction model provides the opportunity to individualize fertility
treatments, thus avoiding unnecessary IVF treatments without affecting the chances
of fertility. [3333].
Ethical Aspects
A discussion of the ethical aspects implicit in the use of AI in reproductive medicine
is presented below. We found that the ethical aspects can be divided into four main
areas, as shown in
the table ([Table 1Table 1]):
Table 1
Table 1
Ethical aspects of AI in reproductive medicine.
|
Potential opportunities
|
Potential risks
|
|
Research ethics
|
-
Developing improved treatments for infertility treatment options → increased probability
of reproduction
|
-
Difficulties to adequately inform patients/study participants’
-
Long-term monitoring not always possible
-
Unrealistic hopes of study participants
-
Moral status of human embryos
|
|
Impact on the chances and autonomy of patients
|
-
Higher baby-take-home rate
-
Lower therapy discontinuation rates as well as lower physical and psychological stresses
-
Social stigma of childlessness avoided
-
Support of reproductive autonomy
|
-
Lack of evidence on effectiveness
-
Insufficient information given to patients/study participants
|
|
Physician-patient relationship
|
-
Personalized information for patients
-
Better therapy options to treat infertility
-
Increases time resources in clinical practice
|
-
Informing patients is challenging because of the complexity of algorithms
-
Treatment results for the individual patient are unclear
-
Responsibilities, transparency and trust still not clear
|
|
Reproductive justice and chances of access
|
-
Lower financial burden
-
Creates a hierarchy of patients according to the probability of success
|
-
No widespread implementation
-
Follow-up costs due to complex technical operation (e.g., maintenance; liability)
-
Access to specialized “repro AI centres”
|
Research ethics
As shown by examples from reproductive medicine research, using AI in medicine opens
up new possibilities in clinical applications. However, up to now, even promising
developments have
rarely been brought to technical maturity, and even fewer have become part of standard
clinical practice. In ideal conditions, the methods of medical ML in particular could
be highly
effective, for example for analysing imaging data. But performance in clinical practice
is considerably poorer, which might also be the case for AI-supported imaging analysis
of
spermiograms. This is unsurprising, as the overwhelming majority of studies with AI
or ML in medical research are performed in a retrospective setting [3434]. Investigations into AI applications also have indicated that the study designs
have limitations, creating further difficulty making statements about their
effectiveness in the clinical setting [3535]. Recently, the dearth of randomized controlled studies (RCT) has led to some criticisms
about the validity of existing research and development in the field of medical AI
and ML [3636].
In recent decades, developments in reproductive medicine such as PGD or ICSI were
often transferred directly from laboratories to clinical applications without undergoing
a comprehensive
review of their effectiveness and safety [3737]. Studies in the field of reproductive medicine conducted in humans are particularly
sensitive from the standpoint of research ethics and also present very particular
challenges. While initially, the intended mother is the person who is directly affected
by the study (and,
in contrast to her male partner, also bears the majority of physical and psychological
burdens), research and expected innovations in the field of assisted reproduction
(e.g., AI
applications) can result in the birth of offspring. The offspring are inevitably affected
by the possible risks of experiments, despite not having had the option to consent
to participating
in the study. Similarly, potential long-term effects on offspring caused by an experimental
procedure performed during assisted reproduction may possibly only be fully understood
several
years later. This problem is compounded by the ethical questions that arise during
the testing of new procedures based on the use or non-use of human embryos.
In all future research into the use of AI in reproductive medicine, providing well-considered
information to the intended mother or intended parents will be crucial. It is also
important to
be aware of unfavourable situations that may arise when providing information to potential
mothers or parents: study participants may nurse false hopes or unrealistic expectations
when
modern, attention-grabbing innovations are being tested. Especially women or couples
with a longstanding painful experience of childlessness must be considered as a vulnerable
group. As
those participating in such studies may feel like this is their “last chance” to fulfil
their desire to have a child, special precautions are necessary to ensure that participation
is
voluntary.
Patients’ well-being and autonomy
One of the stated goals of using AI systems in predictive models for IVF is to improve
outcomes compared to results obtained using conventional reproductive medicine methods.
One measure of
such improvements would be a higher baby-take-home rate, i.e., a higher probability
that assisted fertilisation will result in a live birth. This could reduce the psychological
and the
physical suffering of patients. The normative principle of the primacy of the patient’s
welfare in assisted reproductive medicine includes providing the best-possible suitable
treatment,
based on objectively measurable medical parameters, and the patient’s subjective experience,
which involves taking into account of the patient’s preferences in the treatment setting
and the
patient’s satisfaction with the treatment. The use of AI technology could, in future,
offer benefits in both areas. An analysis of 122560 treatment cycles in Germany showed
that 45699
patients discontinued therapy after the birth of a child. The remaining 76861 (62.7%)
patients discontinued therapy before they were able to fulfil their desire to have
children [3838]. A variety of reasons were given for discontinuing therapy: the absence of transferable
embryos due to immature oocytes, the inability
to harvest oocytes, failed fertilization attempts, or arrested embryo development.
Such factors can discourage patients. Other reasons included no or insufficient response
to stimulation,
overstimulation syndrome or premature ovulation as well as incorrect administration
of hormone injections leading to an unsuccessful course of treatment. Failure to conceive
despite
undergoing many reproductive medical procedures over a long time is also believed
to be one of the reasons why patients discontinue therapy [3838]. Other studies have come to similar conclusions and, in addition to these physical
and psychological burdens, also have mentioned relationship difficulties and other
personal problems as potential reasons for discontinuing treatment [3939]. Furthermore, infertility is often tainted by social stigmas
such as shame and social exclusion [4040]
[4141]. More effective and faster
treatment, which is conceivable in future AI applications, could provide some couples
with a technological option to avoid these stresses, at least in partially, and thereby
contributing to
their well-being.
The use of AI in reproductive medical practice can also be analyzed in the context
of patients’ reproductive autonomy [4242]
[4343]. Reproductive autonomy is a normative concept and should be understood as the capacity
of individuals to decide freely, well-informed
and without interference from others about their own reproduction. From this perspective,
measures that support and enable patients to exercise their reproductive freedom such
as the use of
AI in reproductive medicine should ideally be available to everyone who want to have
children but cannot [4444]. Conversely, limiting
reproductive autonomy should only be permissible if the use of new technologies in
reproductive medicine would demonstrably result in an harm to the patient or their
potential offspring
[4545]. Reproductive autonomy can also be interpreted as a right of entitlement, because
being able to actuate one’s own reproduction
or non-reproduction can be viewed as a central aspect of a person’s identity. Thus,
it could be concluded that attempts must be made to support couples fulfilling their
desire to have
children [4444]. This includes, for example, ensuring that access to and use of future reproductive
technologies including AI
applications is granted to such couples. However, the risks posed by an unfavourable
information situation as outlined in the previous chapter must be avoided to allow
patients to fully
express their reproductive autonomy in this still experimental field.
Benefits and challenges for the physician–patient relationship
The potential clinical implementation of AI in reproductive medicine will inevitably
impact on a lynchpin of the medical practice: the relationship between physician and
patient. The use of
AI may require reproductive medicine specialists to rethink their professional role,
as they are the bridge between the algorithmic output and treatment-relevant decisions
[4646]. The physician must not only address the biological factors behind the patient’s
infertility but also patients’ particular
psychosocial and emotional stresses during treatment, which are well-known and have
been demonstrated in studies [4747]. In this context,
the future use of AI could be beneficial as it will be possible to predict the success
rate of individual patients better in terms of their likelihood of becoming pregnant.
There is existing
data that can be used to make probability statements, for example, based on the age
of the woman [4848]. However, an optimized and more
precise prognosis could provide physicians with the opportunity to ensure patients
are better informed and provide appropriate therapy recommendations. An option to
optimize the selection of
sperm, oocytes, and embryos would allow the treating physician to offer better and
more efficient treatment. While the contribution of potential AI applications to such
improvements may be
minimal, given the stresses of infertility treatment, even those could offer valuable
benefits. Moreover, it is hoped that the use of automated support systems in medicine
will allow
physicians to devote more time to the physician–patient relationship [4949].
The potential implementation of AI into reproductive medicine would also place demands
on medical staff. On the one hand, they are responsible for collecting and recording
personal data
including age, weight, and lifestyle information, etc., which will make it possible
to train algorithms that could help facilitate pregnancies in the future. However,
reproductive medicine
specialists must also explain the use of algorithmic decision support systems to their
patients and provide them with appropriate information: How can using AI technology
have a more
positive impact on the respective diagnosis and/or treatment? Is the physician convinced
of the benefit of the AI support or not, and why? [5050] On the other hand, the medical staff treating patients must ensure that the prognosis,
diagnosis or treatment recommendation supported by AI systems do not contravene
medical state-of-the-art and their professional judgment [5050].
If future AI systems are going to place a greater emphasis on the quality of the gametes
and determine their quality, the focus could move away from the individual patient.
Predictive
analysis models and the resulting treatment recommendations which are based on large
amounts of data may be able to improve the treatment results for specific patient
cohorts but they may
not be necessarily beneficial for an individual patient. Consequently, these circumstances
may come into conflict with the physician’s obligation to act in the best interests
of every
individual patient [5151]. Over the course of the various examinations, data collection procedures and analysis,
there is a considerable
risk that patients could become mere “data subjects” [33] and may not be perceived as persons [33]. Treating physicians must be aware of the potential dynamics involved in the datafication
of people and continue to pay attention to the individual patient.
Questions of liability are another problematic area associated with the use of AI.
With the increasing digitization of medicine and the use of ML algorithms, new players
are increasingly
entering the healthcare system [5252]. These include tech companies and programmers, who play an important part in developing,
training
and testing ML systems. If AI applications result in treatment errors and wrong diagnoses,
this will raise new questions regarding who is responsible [5353]. This problem is compounded by how ML applications can appear to be a type of black
box [5454]. At time
algorithms with a high validity can no longer be explained or the explanations would
entail considerable effort or expense. The opacity of AI applications can make medical
decision-making
more difficult for specialists, as it may be unclear when they can rely on automated
systems. The lack of transparency may also lessen patients’ trust in relevant AI applications
[5555]. In addition, human–machine interactions pose certain challenges. Physicians with
extensive experience in their field appear to have a
greater mistrust of AI systems, while less experienced physicians may place excessive
confidence in such systems [5656].
Reproductive justice, and access
In Europe, the regulations on the reimbursement of the costs of assisted reproduction
are very heterogeneous. Criteria regulating access to reproductive technology, such
as age (both of the
woman and the man intending to have children), whether the intending parent already
has children, or how many treatment cycles the couple or the woman has already had,
can differ between
countries. In some countries, even female BMI can be a criterion for receiving public
funds [5757]. There are also considerable
differences between countries with regard to the three main cost areas: medication,
personnel, and laboratories. In some countries it is also relevant whether the IVF
center is publicly
funded or a private facility [5757]. Hence infertile persons’ options for accessing assisted reproductive technology
may differ,
resulting in social inequalities [5858]. In Germany, for example, only 50% of the costs of a maximum of three treatment
cycles with IVF
or ICSI are covered for persons insured under the German public health insurance scheme,
which consists of approximately 90% of the total population. Moreover, there are age
limits that
further restrict claims for reimbursement: the woman must be between the age of 25
and 39 years and the man between 25 and 49. Given this background, it would appear
that access to assisted
reproduction differs significantly from access to other healthcare services and that
it is closely linked to cultural or moral norms and ideas of justice [5959].
The prospect that AI procedures could soon be ready for use in clinical practice raises
the hope that it will be possible to offer infertile patients more efficient therapy,
which could
lead to successful pregnancies and the reduction of the financial burden. However,
in addition to these possible positive aspects, it is also important to be aware of
other implications.
Reducing the costs for individual patients and for the collective payor will only
occur if the use of the new technology is efficient and if the AI-based technologies
do not involve
(disproportionately) higher costs generated by various factors, such as purchase,
operation, data processing and storage, maintenance and updating of the model, visualization,
the need for
skilled operators, rectification of mistakes, possible liability costs, etc. in clinical
practice [6060]. Furthermore, considering the
problem of fairness raises questions about access or barriers to access: while it
is debatable whether AI-based systems will soon become part of clinical practice,
it can be assumed that
they will not be widely available immediately. It is conceivable that only a few reproductive
medicine centres will initially include these in their list of services. Patients
who do not
have access to these centres may have to confront the likelihood their chances of
success being lower, and reproductive medicine facilities that are unable to offer
these services and do not
provide the possible new “gold standard” could suffer a comparative disadvantage in
terms of demand for their services. A concentration of AI-based reproductive medicine
services in
high-tech centres may cause the costs of procedures and measures that are considered
necessary or desirable to initially increase. While this dynamic response to the use
of AI may represent
an advantage for the reproductive autonomy of some individuals, from the perspectives
of reproductive justice and of reproductive medicine as a market economy, where healthcare
services and
costs are financed collectively, this concentration can represent a disadvantage for
certain groups of patients with insufficient financial resources or mobility [6161]. Hence, when conducted ethical assessments and weighing the consequences of technological
change, attention must also be paid to
potential barriers to access, issues of availability, and the financing of services.
A further aspect of justice concerns not only the question of costs but also the practical
implementation in hospital. Based on critical reflections on the increasing “quantification
of the
social fabric” and its associated effects [6262] it should be considered whether the seemly more precise measurement of the success
rates
of pregnancy could lead to a hierarchical categorization of persons requiring treatment
and thereby engendering additional inequalities. The literature on the (potential)
use of algorithms
has raised concerns that algorithms could reproduce existing inequalities through
the persons who design the algorithms or the data used to train the algorithms [6363]
[6464]. When weighing the (potential) use of AI in infertility treatments, it is
important to consider the question of which data are being used to train the algorithms,
ideally to prevent discrimination against certain groups of patients.
Discussion
As in various other medical fields, there are several obstacles and risks associated
with the potential use of supporting AI systems in reproductive medicine. It should
be noted that AI-based
methods in reproductive medicine are still in the early stages of development, making
it very difficult to weigh what the actual risks and opportunities could entail. Furthermore,
ML models
can often not be fully explained and may be perceived as a “black box” [6565]. This could result in a certain scepticism among clinicians
and patients with regard to the diagnoses and therapy recommendations. The new scenario
could result in patients feeling helpless when confronted with the use of non-transparent
tools and
automated decision-making processes that affect important aspects of their personal
lives [5050], creating additional uncertainty in a
group of patients who cannot have children and are already physically and psychologically
vulnerable. In contrast, for physicians, the opacity of AI applications may result
in over-reliance on
such systems or excessive distrust of them. Both scenarios could disadvantage patients.
On the one hand unrealistic expectations about the efficiency of new technologies
may give patients
unjustified hope regarding their desire to have children. On the other hand, unwarranted
scepticism about a highly effective AI system may result in the potential of such
innovations being
neglected.
According to recent studies, there are several limitations with regard to the quantity
and quality of data, which significantly influence the performance, applicability
and generalizability
of the trained model. The majority of studies in the field of reproductive medicine
have small sample sizes and are retrospective. Largescale randomised controlled studies
which could test the
validity of the algorithms and optimize their utilization are still lacking. Hence
more research into personalized diagnosis and treatment, medical expert systems, and
AI-supported
reproduction is necessary [1414].
Supervised ML and unsupervised ML raise the question of what physicians can know or
say about the use of AI to affect the outcomes of infertility treatments. Physicians
are obliged to handle
with AI-generated recommendations even though they occasionally may not entirely understand
the system and/or agree with the system’s recommendations regarding the diagnosis
and therapy [5050]. This could make it more difficult to provide transparent and patient-focused information
when using AI-supported data processing [6666].
Overall, it appears that it will be necessary to examine the moral hazards described
above in more detail from the standpoint of empirical ethics [6767]. In such a context, the focus would not be on making an “objective” assessment and
weighing the risks of different AI applications in reproductive medicine; rather
would be on the (subjectively conveyed) conditions that affect the desirability, acceptance
or rejection of AI in reproductive medicine. If these challenging areas are addressed
in the course
of qualitative interviews with physicians, persons with a desire to have children,
AI researchers, providers and other groups, additional insights into the expected
benefits and perceived
risks of AI in reproductive medicine could be gained from the perspective of the affected
persons before potential applications are integrated into standard care.
Conclusion: Future Direction of AI Applications in Reproductive Medicine and an Ethical
Assessment
Conclusion: Future Direction of AI Applications in Reproductive Medicine and an Ethical
Assessment
Improving the capability and capacity of AI technologies over time and integrating
them into the treatment process could benefit patients and physicians by making high-quality
reproductive
medicine more effective and precise and by providing support to the treating physicians
during decision-making. The future utilization of AI in reproductive medicine prior
the starting IVF
treatment could offer promising results and allow better prognoses to be made with
regard to treatment success. This would give persons with an unfulfilled desire to
have children the
opportunity to address their individual chances of having a child early on with the
aid of reproductive medicine specialists.
However, developments in the field of reproductive medicine must also be examined
to assess the extent to which they could promote undesirable medium- and long-term
effects and social
dynamics. At the global level, reproductive medicine is not merely a growing research
area but also a lucrative industry in which many different actors are vying for the
attention of potential
parents and clinics. In the context of popular AI applications in particular, innovations
promise new feasibilities, whether in prediction of treatment outcomes or prenatal
diagnostics. There
are already companies that offer algorithm-supported embryo selection even for polygenic
traits, for example, to exclude the risk of schizophrenia or cancer in offspring [6868]. While the scientific basis of such services remains insufficient, this will create
demand and generate further optimization
fantasies.
It also remains to be seen to what extent the existing processes and logics underpinning
the quantification of life phenomena could emerge or be reinforced by the use of AI-supported
decision-making systems in reproductive medical practice, for example, using AI systems
to reduce complex decisions to their metric values, resulting in a binary differentiation
between “good”
and “bad” embryos. According to Mau, descriptions represented in numerical units always
express “values attributions”, comparisons and “value orders” [6262] and hence are hardly unbiased. They mirror attitudes and social effects such as
acceptance and non-acceptance, and may thus steering behaviours [6969]. For instance, it has become apparent over several years that countries that offer
systematic prenatal screening for trisomy 21 have
higher abortion rates of such pregnancies [7070].
If decisions in fertility clinics will, in the future, increasingly be made with machine
support, it will be necessary to critically observe and examine whether and to what
extent
pre-programmed value categories are the defining standards that mold the opinions
of physicians as well as potential parents (cf. [6363]).
A machine calibrated to optimize outcomes may ultimately even become the pacemaker
of a new form of eugenics, even if no identifiable “eugenicists” are involved [7171]. This makes it all the more essential to ensure that the values underpinning the
algorithms are transparent and that they are discussed
publicly. This would help AI-supported medicine become a humane medicine.
