Keywords
artificial intelligence - ethics - machine learning - patient care
In the 21st century, artificial intelligence (AI) has become synonymous with the bulk
of modern technology, where advancements in programming and computer science have
expedited change in numerous sectors of society. In the healthcare industry, AI will
have an impact on the tripartite missions of patient care, education, and research.
For example, it is being used to support clinician decision making,[1]
[2] engage patients as virtual conversational agents,[3] and predict surgical outcomes[4]
[5]; most recently, the March 2023 release of OpenAI's latest AI system Generative Pretrained
Transformer 4 (GPT4) has demonstrated abilities ranging from transcribing clinical
encounters into physician notes to correctly answering exam questions from the US
medical licensing examination.[6]
It is unsurprising that radiology, the specialty based on advanced digital imaging,
is at the forefront of AI innovation in medicine. Smarter machines programmed with
intelligent models have the potential to provide high-quality images with less artifactual
distortions, decrease the amount of radiation emitted during imaging, and assist in
image interpretation by providing automated measurements of tumor scans.[7] Similar to and separate from these exciting prospects, other fields have found their
own niche applications of the newest technology: Esteva et al[8] developed a model that successfully identified skin cancers from medical images
with an accuracy rate on par with multiple dermatologists; Bertsimas et al[4] developed a tool that, within 4 to 11 questions, accurately predicted the risk of
mortality and other postoperative complications following emergency surgery. Given
that the amount of data in electronic health records (EHRs) is proposed to grow by
nearly 50% every year,[9] increased development of such AI-powered models would make possible the ability
to harness incredible amounts of information to a degree unachievable by manpower
alone. Understandably, such seemingly limitless potential raises ethical questions
including, but not limited to, the infringement of AI on patient privacy, the accountability
of computers (versus physicians) for errors, and the development of biased algorithms
by using data that underrepresents certain demographics, like race or gender.[9]
[10]
[11]
In facial plastic surgery, the potential applications and ethical considerations of
AI are likewise widespread: from research to workflow and patient evaluation, the
high digital visibility of this field—via “before” and “after” pictures, for example—and
the demand for flawless aestheticism lend itself well for implementation of precise,
AI-supported care. Ironically, these same characteristics also produce unique ethical
concerns. The purpose of this article is to provide an overview of AI and discuss
current uses, future possibilities, and ethical implications of AI in facial plastic
surgery.
What Is AI?
AI is broadly used to describe computer programs or machines designed to complete
processes that mimic human intelligence, such as learning or problem solving.[1]
[2] These programs are composed of algorithms—sets of rules that produce a given output
from input data—that have the ability to recognize patterns in large data sets.[12]
[13] In one category of AI called machine learning (ML), developing an algorithm requires
initial “training” via some amount of input information, where the algorithm uses
one of three types of learning styles to identify specific features or patterns within
the data. Once trained, the algorithm is tested on a new dataset to ensure that it
is able to maintain its accuracy when applied to unseen data.[1]
[12] Depending on the learning style employed, the algorithm is better equipped to perform
a specific type of function, whether that be categorizing medical images into clinical
outcomes or finding subtle similarities between genetic sequences.[12] In short, ML can be described as programs that generate desired outputs (i.e., predict
an outcome) by using algorithms that find patterns in large datasets.
Subfields of ML include deep learning (DL), natural language processing (NLP), and
computer vision.[1]
[12] Briefly, DL is a type of ML that utilizes algorithms inspired by the human brain.
These algorithms are called neural networks, where multiple layers of processing allow
for self-directed learning and complex problem solving.[1]
[14]
[15] Facial recognition technology is one example of DL. NLP refers to language detection
and analysis—both speech and text—exemplified by medical chart interpretation and
chatbots, or virtual conversational agents.[3]
[12] Computer vision is defined as image and video analysis using various algorithms
for object detection, image classification, and segmentation of regions of interest
(such as human faces).
Current Applications
Workplace Efficiency
One of the most simple and direct applications of AI in transforming medical care
lies in optimizing nonclinical tasks such as documentation. The EHR was one of the
most influential changes to medical care[16]; however, due to significant regulatory changes, much of clinicians' workloads are
spent interacting with EHR.[17] To relieve some of this burden, AI may be used for documentation of patient-physician
communications in the form of a medical scribe. Using speech recognition and speech
to text conversion algorithms, transcribing communications via AI during a patient
visit can lead to enhanced workplace efficiency and less administrative time,[18]
[19] and companies like Augmedix and Deepscribe already offer such services.[20]
[21] As an interesting example, Patel and Lam[22]—recognizing the challenges of proper discharge summary documentation—demonstrated
the ability of ChatGPT (an AI chatbot that utilizes NLP technology) to write a discharge
summary for a theoretical postsurgical patient after it was given a basic prompt.
However, caution must be exercised when increasing adoption of AI in the setting of
documentation, as automation bias may ensue if clinicians develop an overreliance
on such technologies. For example, accepting documentation without confirmation of
its accuracy can lead to errors in the EHR.[23] Bearing this in mind, incorporation of AI into EHRs has the potential to help providers
optimize their workflow by allowing for more efficient and timely medical documentation.
Additionally, conversational agents utilizing NLP have been used to assist patients
for other purposes including appointment management, triage, or medical advice, and
patient perceptions of the effectiveness and usability of these technologies are generally
positive.[3] Baker et al[24] compared the diagnostic accuracy and the appropriateness of recommendations of an
AI-based triaging system with multiple physicians using clinical vignettes similar
to those used for simulation-based learning in medical education. Although there was
no statistical analysis, the system was able to diagnose conditions and provide recommendations
comparable to that of human doctors; in fact, the AI system's recommendations were
generally found to be safer than those provided by the physicians.
Preoperative Decision Making
Facial plastic surgery is unique in that the face is among the most distinct structures
of the human body. Pre- and postoperative pictures are already common practice and
large databases can be created to categorize facial anatomy and model outcomes.[25] While research using AI in aesthetic and facial reconstructive surgery is still
in the early phases, AI systems can take advantage of these databases to inform preoperative
decision making. For example, ML programs are able to make independent associations
using large datasets to identify patterns and trends, where the Google search engine
is one of many algorithms that use ML to predict a user's preferences. Borsting et
al[25] created a DL-based program called “RhinoNet” that, using data from thousands of
before and after images, was trained to predict whether the image depicted an individual
who had undergone rhinoplasty. The program successfully estimated rhinoplasty status
in 85% of the test images, where it performed better (although not statistically significant)
compared to expert opinions. Similarly, Phillips et al[26] demonstrated the ability of an AI-based detection algorithm to accurately identify
melanoma with performance similar to dermatologists.
Such algorithms could be used in facial plastic surgery in various ways; for example,
AI may analyze large sets of pictures and videos of patients to create prediction
models to categorize favorable versus unfavorable patient anatomy. Formeister et al[27] showed the ability of an ML algorithm to identify chief factors (i.e., flap ischemic
time, smoking) involved in complications for head and neck free flap tissue transfer,
which differed from factors emphasized by traditional statistical models (i.e., flap
type). However, it should be noted that one difficulty with developing models in facial
plastic surgery is the significant variation in facial anatomy between individuals,
which necessitates a large number of sample images to train AI-based algorithms.
Postoperative Outcomes
A major limitation of assessing postoperative outcomes in plastic surgery is the lack
of objective evaluation, especially considering the inherently subjective qualities
of aesthetic surgery.[28] Currently, patient satisfaction and surgeon-based perception assess outcomes. However,
because patient satisfaction is influenced not only by the actual surgical results,
but also by personal expectations, AI may assist in creating more objective methods
of evaluation. A study examined the accuracy of an AI algorithm to predict perceived
age reduction after facelift,[29] where neural networks were very accurate in predicting preoperative age. Such programs
can provide validated methods of determining postoperative results for plastic surgeons.
With the advent of generative AI (including generative adversarial networks and diffusion
models) such as DALL-E, one can create various types of synthetic images using AI.
This is very useful for simulating postsurgery result images virtually even before
the procedure. Surgeons can virtually try out different interventions and both surgeons
and patients can look at the AI-simulated result face images to evaluate and score
them. This process helps the surgeon fine tune the actual procedure preemptively and
aim for the best outcome.
Surgical Training and Research
Plastic surgeons require expertise in technical ability, anatomy, surgical planning,
and clinical knowledge. Oftentimes, surgical performance is difficult to assess objectively
and most metrics are not validated with evidence.[30] In this regard, AI-based simulation is another application that is gaining attention,
as algorithmic analyses of video-recorded surgeries can help trainees identify technical
weaknesses and predict outcomes of different surgeries.[30]
[31] For example, with the use of wearable sensors, an ML algorithm was able to identify
movement patterns that were associated with the skill level of a surgeon.[32] As these algorithms mature, there is great promise for these programs to be utilized
as tools that enhance the quality of surgical training. Finally, AI may also be utilized
as a supplemental resource when conducting research. Gupta et al[33]
[34] prompted ChatGPT to produce novel ideas for systematic reviews across various topics
in plastic surgery, and its results were subsequently compared to literature searches
of multiple databases. While the overall accuracy of proposing truly novel ideas for
systematic reviews was only 55%, this software may provide a useful starting point
for those wishing to contribute to research.
Future Applications
The horizons of AI implementation in the future are broad, where continuous advancements
in technology may make it possible to consider opportunities and applications that
cannot currently be imagined. While one group used AI to predict which patients may
be more likely to develop a postoperative surgical site infection,[35] further advances with image processing, interpretation of symptoms, and vital sign
assessment may eventually allow for patients to triage any concerns surrounding their
surgical site. AI tools can also help monitor and forecast postsurgery wound healing
times and potential issues with that. Moreover, computer vision AI-based systems can
also be used to coach and monitor patients and caregivers for optimal postoperative
behavior and care, such as exercises for muscle movement. Such software could assist
in decision making that may prevent unnecessary presentations to the emergency room
or give medical recommendations to patients who may not have access to timely healthcare
or are unsure of what care to seek. Additionally, future applications of cosmetic
surgeries may utilize predictive software; for example, a program may learn individual
surgeons' particular “styles” through before-and-after pictures and inform patients
wishing to undergo similar procedures, with the patients' outcome preferences driving
the program's output. In addition, AI could be used to develop patient-specific treatment
plans that combine surgical and nonsurgical procedures to improve reconstructive or
rejuvenation outcomes. For example, one patient may benefit from deep plane facelift
with fat grafting, while another may benefit from mini-face lift, laser, and filler.
Ethics
Proponents of AI in patient care agree that any use of AI should be informed by the
same ethical principles dictating traditional human-provided care; however, this is
complicated by the fact that the increasing involvement of advanced technology in
healthcare presents unique questions that may not be completely answered by traditional
ethical principles. Relatedly, there have been concerns regarding the possibility
that AI may eventually replace, rather than assist, healthcare practitioners. This
may be especially true in technologically-based subspecialties and when considering
the development of an AI-driven robotic surgical device that can successfully perform
basic surgical skills.[11] However, years of physician expertise and the sanctity of the patient-provider relationship,
along with other inherently human qualities like empathy and compassion, cannot be
digitally replicated and should not be undermined. When considering its additional
limitations, it does not appear that AI will overtake physicians anytime soon.[1]
[11]
[36]
[37] Indeed, Pinto Dos Santos et al[38] found that, in a survey of 260 medical students, a vast majority (96%) disagreed
that physicians could be replaced by AI and agreed that AI should be incorporated
into medical training (71%). Thus, the inclusion of AI in medicine should not be feared,
but approached with careful excitement and curiosity for the ways that AI may, when
used appropriately, contribute to efficient, accurate, and patient-centered care.
Consequently, there is a need to dissect the ethical intersection between human thought
and virtual processing when considering the proper implementation of AI as a tool
that may augment physicians' abilities to provide care, educate, and innovate.
Legal Liability
One important consideration is the liability of AI-based medical recommendations,
especially the point at which a physician's recommendations may be discordant with
AI. Loftus et al[39] used DL to evaluate the over-triaging and under-triaging of postoperative patients
and whether they should be treated in an ICU setting or a ward setting with regard
to resource utilization and acuity of care. Though this data was analyzed in a retrospective
setting, studies like this will be used to inform development of real-time decision-making
tools to aid surgeons in determining the most appropriate disposition for postoperative
management. As described by Morris et al[40] in the event of patient harm, there remains to be a discussion as to whom the liability
falls upon: the surgeon overriding a ML-based prediction or the developer of the technology.
Future considerations for implementation of such technology into patient care would
necessitate continued shared decision-making and an informed consent process in which
detailed risks and benefits to AI-augmented care should be discussed with patients.
Bias
Another important ethical consideration with the increased prevalence of AI in surgery
is the amplification of racial, socioeconomic, or gender biases. If the data used
to teach automated neural networks is based on preexisting documentation, then the
systems are inherently prone to the biases of the provider documenting such information.[41] Additionally, patients of lower socioeconomic status—who are more likely to seek
care at multiple institutions—may be perpetually disserviced if records of their care
are spread across multiple, noncommunicating EHRs, possibly resulting in over-testing
or undertreatment.[42] Parikh et al[41] proposed methods to reduce such bias including training AI models with populations
mirroring those most affected by a particular condition rather than from the general
population. They also suggested flagging patients for whom a predictive model has
limited EHR data to indicate that the algorithmic outputs should be reviewed in greater
detail. In order for these technologies to guide patient care, it would be crucial
to ensure aggregation of medical records for accurate information processing.
For supervised ML, different experts may label the same data point or image differently,
based on their experience. This causes interobserver variability, and the AI system
needs to be cognizant of and gracefully handle this issue with ground truth uncertainty.
Bias in data collection may cause degradation of the AI model when deployed in new
settings (a different hospital or location). Data to train AI may also be lacking
for certain races causing AI bias.
AI fairness is an important and growing topic. To mitigate bias, various metrics such
as group parity have been developed. Facial plastic surgery would need to use such
fair AI models, even at the possible cost of lower overall accuracy.
Specific to facial plastic surgery is the concept of beauty, the pursuit of which
drives many patients' decisions to pursue cosmetic procedures. ML programs have been
taught to interpret the attractiveness of faces from photographs and recommend surgical
plans for patients wishing to undergo aesthetic surgeries.[43] Computer systems that—by nature—rate beauty as an objective measurement and leave
no room for subjective preference would elevate certain qualities as superior to others;
in other words, they would define the “perfect face.” This is problematic when perfection
is based on a narrow set of beauty standards derived from specific cultures (i.e.,
Western countries), where such bias leads to the undervaluing of qualities that may
be considered beautiful in other races or ethnicities. Consequently, AI may contribute
to decreasing diversity in perceptions of beauty,[11]
[14]
[44] and would only serve to inform the goals of individuals who have beauty ideals similar
to those that are standardized within the algorithm. As such, surgeons using predictive
software during patient consultations for aesthetic procedures should bear this in
mind and minimize the risk of coercion in pursuing an outcome misaligned with the
patient's goals. Again, this emphasizes the role of AI as a tool that aids patient
decision making, not as a primary decision maker itself.
AI Model Explainability
AI models have become more complex over the years. However, it is crucial that the
AI models are interpretable, especially in a healthcare setting. One needs to understand
what is triggering the AI to make a certain decision. For this, explainability is
a growing and important topic. This process takes a complex black box model and helps
the user (doctor and/or patient here) gain insights into the AI decision making process.
One concrete example in medical imaging AI is when the AI calls a particular patient's
X-ray as suspicious (say, for cancer), then the explainability module will generate
a “heatmap” showing which regions of the image caused the AI to call cancer, most
likely because those image regions had lesion like objects. This is also crucial for
plastic surgery. Consider a scenario where the plastic surgeon is trying to evaluate
a surgical outcome. Using explainable AI, they can predict and identify specific parts
of the face that may be affected adversely due to the procedure (resulting in an overall
poor outcome) and prioritize these areas accordingly.
Data Security
Implementation of third-party AI software into patient care would require stringent
policies on patient data usage and storage held to the standards outlined by the Health
Insurance Portability and Accountability Act (HIPAA). One consideration unique to
surgery involves the use of video-based AI learning models. Prigoff et al[45] highlighted several ethical recommendations regarding the use of video recording
in operating rooms, including a clearly stated purpose for recording, informed consent,
the ability to opt-out of data retention, and for video recordings to be held with
the same security standards as other medical records. Further standards regarding
the timeline of data retention for applications such as preoperative-to-postoperative
predictive software would also need to be developed. Distributed storage and federated
ML can be used to train AI with the data still residing on the patient's phone or
computer, reducing the risk of data loss or breach. AI can also be used to identify
and mask out personal identifying information and other sensitive information before
patient data is stored on the cloud, to ensure patient privacy.
Conclusion
The applications of AI in healthcare are numerous and will likely continue to grow
as advancements are made in the quality and accuracy of AI-based models. AI shows
great promise in the field of plastic surgery, where collections of facial profiles
and the need for objective assessment of surgical outcomes serve as unique points
of application. When used carefully and in alignment with physician judgment, AI programs
may provide important contributions to efficient healthcare workflow and delivery
in facial plastic surgery. This review provides a glimpse into the current applications
and potential future directions of these intelligent technologies, as well as an important
discussion regarding the ethical implications of nonhuman, computer influences on
a field that, while based on scientific objectivity and biological fact, is equally
driven by inimitable human character.
