Skip to main content
NCBI home page
NIHPA Author Manuscripts logoLink to NIHPA Author Manuscripts
. Author manuscript; available in PMC: 2025 Aug 22.
Published in final edited form as: Otolaryngol Head Neck Surg. 2024 Dec 12;172(2):734–743. doi: 10.1002/ohn.1080

American Academy of Otolaryngology–Head and Neck Surgery (AAO-HNS) Report on Artificial Intelligence

Noel F Ayoub 1,2, Anaïs Rameau 3, Michael J Brenner 4, Andrés M Bur 5, Gregory A Ator 5, Selena E Briggs 6, Masayoshi Takashima 7, Konstantina M Stankovic 2; AAO-HNS Artificial Intelligence Task Force
PMCID: PMC12369755  NIHMSID: NIHMS2103112  PMID: 39666770

Abstract

This report synthesizes the American Academy of Otolaryngology–Head and Neck Surgery (AAO-HNS) Task Force’s guidance on the integration of artificial intelligence (AI) in otolaryngology–head and neck surgery (OHNS). A comprehensive literature review was conducted, focusing on the applications, benefits, and challenges of AI in OHNS, alongside ethical, legal, and social implications. The Task Force, formulated by otolaryngologist experts in AI, used an iterative approach, adapted from the Delphi method, to prioritize topics for inclusion and to reach a consensus on guiding principles. The Task Force’s findings highlight AI’s transformative potential for OHNS, offering potential advancements in precision medicine, clinical decision support, operational efficiency, research, and education. However, challenges such as data quality, health equity, privacy concerns, transparency, regulatory gaps, and ethical dilemmas necessitate careful navigation. Incorporating AI into otolaryngology practice in a safe, equitable, and patient-centered manner requires clinician judgment, transparent AI systems, and adherence to ethical and legal standards. The Task Force principles underscore the importance of otolaryngologists’ involvement in AI’s ethical development, implementation, and regulation to harness benefits while mitigating risks. The proposed principles inform the integration of AI in otolaryngology, aiming to enhance patient outcomes, clinician well-being, and efficiency of health care delivery.

Keywords: artificial intelligence, bias, ethics, large language model, otolaryngology–head and neck surgery

Recent years have witnessed a meteoric rise in artificial intelligence (AI) across numerous industries.1 The widespread dissemination of generative AI and other forms of large language models (LLMs) has accelerated the use of AI in health care. As is often the case with innovation, health care has lagged behind other industries.2 However, patients, professionals, and policymakers stand to benefit from AI through improved patient outcomes, clinician well-being and lowered costs by automating repetitive tasks and complex analyses. Otolaryngology employs state-of-the-art technology, and the field will inevitably be shaped by advancements in AI in coming years.3 Given the pervasive influence of AI on otolaryngology, there is a vital role for otolaryngologists to play in understanding, developing, and implementing AI-based technologies.

National policies and governance structures for AI remain limited, with minimal guidance for developing, adopting, and maintaining AI. To address this gap, medical organizations have developed guidelines to promote the safe development and use of AI within health care. American Medical Association and World Health Organization (WHO) guidelines emphasize ethical, safe, and purposeful AI use in health care, with the goal of mitigating potential risks.4,5 No such resources have been developed in otolaryngology in the United States. Given the evolving role of AI in health care and implications for the specialty, the American Academy of Otolaryngology–Head and Neck Surgery (AAO-HNS) AI Task Force was created. This report, authored by the AAO-HNS AI Task Force, provides guidance and recommendations for the safe, appropriate, and practical use of AI in otolaryngology. Because this technology is moving at a rapid pace, the Task Force’s recommendations will require periodic updates.

Methods

A comprehensive literature review was performed to gather information on AI within otolaryngology. The search focused on the potential benefits, harms, and needs, in addition to ethical, legal, and social ramifications of implementing AI models in this field. The PubMed/MEDLINE database was searched in February 2024 for relevant publications on AI within otolaryngology, including primary research articles, systematic reviews, and other reviews (e.g., scoping and state of the art) investigating the development or validation of AI applications (eg, neural networks, LLMs, reinforcement learning, random forest trees, and support vector machine algorithms).6 These articles were reviewed to provide the Task Force with an understanding of the existing AI applications within otolaryngology and help guide this report, rather than to analyze the original articles or perform a systematic review. A second PubMed/MEDLINE search was performed to gather existing position statements from other professional health care societies within and outside the United States. These statements were reviewed for recommendations that would be similarly applicable to otolaryngology.

Based on the information synthesized from the literature review, the Task Force’s recommendations were drafted. A modified Delphi method was followed to achieve consensus on the topics for inclusion and recommendations. The draft was distributed to all members of the Task Force, and the process included multiple rounds: (1) An initial round of asynchronous review where all Task Force members provided written feedback on the draft recommendations, (2) a synchronous virtual meeting where the Task Force discussed the feedback and reached consensus on various points, and (3) a final asynchronous review to refine and approve the final recommendations.

Considerations of the Task Force

Based on the literature review, a select number of topics were chosen for this report, focusing on issues that are clinically relevant, timely, and impactful. The discussion is not meant to be all-encompassing or exhaustive; instead, it aims to provide a broad overview that encourages further discussion, research, exploration, and interest.

Briefly, AI is an overarching term encompassing various fields, most of which have potential applications within otolaryngology. At its core, AI refers to computational systems designed to mimic human cognitive functions. Generative AI describes models that can create new content. Machine learning (ML) is a field of AI where algorithms identify patterns and relationships within data, allowing the models to make evidence-based conclusions without continuous human oversight. Within ML, deep learning (DL) uses artificial neural networks to analyze complex patterns, execute tasks, and continuously learn relatively autonomously. Natural language processing (NLP) reads and interprets text and speech to permit text analysis, language translation, and speech recognition. Computer vision uses data to analyze visual information, allowing for the identification and classification of objects.

Generative AI

The release of OpenAI’s ChatGPT-3 (generative pre-trained transformer) in November 2022 represented a landmark moment in AI and catalyzed widespread use of AI within health care.7 This trend accelerated with the subsequent release of ChatGPT 4 in March 2023 and other LLMs. Trained with Reinforcement Learning from Human Feedback, these GPT models undergo extensive training and utilize pattern recognition to generate new information (text, images, etc) and respond to prompts accordingly in a humanlike chatbot format.8 They are trained on tremendous amounts of data, primarily obtained from online resources. ChatGPT is but 1 of many different publicly available generative AI programs, some of which (eg, Med-PaLM 2) are geared toward medical professionals.9

Within otolaryngology, generative AI permits a variety of different inputs, including history, physical, laboratory, and imaging data. These models can additionally provide different outputs, such as a differential diagnosis, predictive modeling of disease course, recommended diagnostic tests, or a management plan. Generative AI can also be used to produce responses to patients’ queries about symptoms, request insurance preauthorization, and perform myriad other functions relevant to clinical practice. AI also has broad implications for education and research. Collectively, these features potentially allow the buildout costs of these models to be amortized across vast data sets rather than across datapoints available to individual patients, learners, or researchers. For these reasons, generative AI has provoked great interest in AI use cases in otolaryngology and health care as a whole.

Impact on Patient Care

The primary goal of AI systems in health care should be to enhance safe, fair, and equitable patient care. All health care-associated AI interventions should primarily aim to support and improve the health of patients relative to the standard of care without disenfranchising patient rights. To this end, the creation and use of AI should follow the quintuple aim for health care improvement: (1) The pursuit of better health for all (2) the attainment of improved outcomes, (3) the promotion of health equity, (4) health care cost reduction, and (5) enhanced clinician well-being.10

Importantly, the integration of AI within otolaryngology has the potential to truly usher in an era of precision medicine. Otolaryngologic disorders encompass a diverse array of conditions with various genetic, environmental, anatomic, and iatrogenic contributions. The complexity of these varied diagnoses necessitates individualized diagnostic and treatment approaches. AI’s ability to analyze patient data to detect subtle patterns and distinctions could help transform otolaryngology from a specialty that often utilizes generalized treatment protocols to one that truly follows the principles of precision medicine. When used appropriately, this technology in theory has the potential to improve patient outcomes, reduce the time to diagnosis, and decrease the use of unnecessary treatments.

Otolaryngologists can use these tools to update patient education materials, taking advantage of the ability to easily translate information into specific health literacy levels and languages. Additionally, otolaryngologists can serve as resources to help ensure that patients of varying education and health literacy levels benefit from this technology. “AI alignment” and “prompt engineering” refer to the concept that the quality of the input affects the response from generative AI programs, and optimizing the use of these tools requires skilled prompt development and refinement.11

Impact on Clinicians

AI systems should extend, and not replace, clinician capabilities, enabling the use of approved AI tools to augment clinical care safely and feasibly. This could take a variety of forms in clinical practice.

Clinical decision support (CDS) systems are 1 such area where AI can impact otolaryngologists.12 CDS systems assist clinicians by suggesting appropriate courses of action based on patient data. Various forms of CDS already exist and are widely used in clinical care, such as those that provide warnings about medication interactions or allergies.13 AI-based CDS can utilize algorithms and patient data to generate possible diagnoses, recommend management options, and predict future outcomes. Importantly, any direct clinical use of AI should involve continuous monitoring by the human counterpart to ensure that the model is acting appropriately.

Perhaps the most significant short-term impact of AI on clinicians will be on their administrative and operational duties. AI applications have the potential to improve hospital operational efficiency, optimize resource allocation, reduce administrative burden, automate repetitive tasks, increase patient health literacy, refine risk stratification, and reduce data fragmentation across electronic health records (EHRs). EHR interoperability limitations and data fragmentation alone contribute to over $250 billion wasted annually, and over $4 billion each year is attributed to clinician burnout.14,15 As a surgical field that is very clinic-driven, otolaryngology stands to benefit from these changes. A key use case is clinical documentation, the biggest time sink for most physicians in clinic.16 These systems have the potential to, in near real-time, generate a medically appropriate summary as a byproduct of conversing with the patient. Additionally, real-time billing and after visit summaries, at a patient-friendly vocabulary level, can be generated.16 Increasingly, minimal clinician input is needed to produce highly relevant, accurate documentation while relieving clinicians of a clerical burden.

Impact on Education and Professional Skills

AI has the potential to transform education and modernize training opportunities for new generations of surgeons.17,18 From the learner’s perspective, the integration of AI into surgical training could help provide enhanced, real-time feedback through objective assessments of surgical performance. Similarly, AI-powered chatbots could help develop simulation-based training experiences, summarize research articles, create study guides, generate case-based or board-style questions, and practice difficult patient interactions. AI-enabled virtual reality and augmented reality can increase opportunities for surgical simulation. From the educator’s perspective, AI may help teachers provide enhanced critical feedback, better identify deficiencies in clinical knowledge and surgical skills, update teaching resources, and reduce the time and effort necessary to develop new educational resources.

Despite these benefits, educators must also evaluate potential adverse effects on training. An overreliance on technology may negatively impact how much trainees truly strive to learn and retain information. Clinicians and surgeons must continue to view AI as an extension, rather than a replacement, of their clinical knowledge and surgical abilities. Educators and universities must also review and clearly delineate their policies on the use of generative AI programs within the educational setting for learning, assignments, examinations, and content generation.

Outcomes and Monitoring

AI-powered software may be considered entirely different than traditional medical devices. However, they bear many similarities and should undergo the same scientific scrutiny as other health care interventions. Prior to the introduction of AI products into clinical practice, the safety and efficacy of the product must be adequately demonstrated. Similarly, the external validity of these products must be understood and tested. A model created in 1 health care system may not produce accurate and reliable results at other institutions.

While there are similarities between traditional medical interventions and AI systems, the differences must be noted as well. One significant benefit of AI-powered models is that they can continually evolve and learn from new data. This dynamic, continuous learning would serve to keep the model up to date, but it may have negative consequences as well. New data that is poor in quality may negatively impact a model that was previously approved, as the accuracy and reliability from the model’s initial testing may change over time as new data is introduced into the training sets. As a result, when implementing AI technology in health care settings, monitoring systems must be set in place to continuously surveil outputs. Manufacturers and developers should monitor the performance of continuous ML models over the long-term and be readily available to troubleshoot and assist health care providers should issues arise.

Research

AI also has a role in advancing scientific innovation in otolaryngology, including translational and clinical research, quality improvement initiatives, and disparities or health services research domains. Researchers from across the world have published numerous applications of AI across all subspecialties within otolaryngology, most notably in otology, laryngology, rhinology, and head and neck.19–21 AI-driven translational research in otolaryngology involves the integration of advanced computational techniques with biological and clinical data to elucidate disease mechanisms, identify biomarkers, and develop novel therapeutics. ML algorithms enable the analysis of data sets, such as genomics, proteomics, and imaging data, facilitating the discovery of molecular or clinical signatures associated with otolaryngology conditions. AI-driven drug repurposing approaches can also accelerate the identification of potential therapeutic agents for otolaryngology disorders, expediting the translation of basic science discoveries into clinical applications.

AI also has promise for clinical research, whether improving diagnostic accuracy, treatment planning, or prognostication. DL algorithms trained on large-scale imaging data sets can improve performance in detecting and characterizing lesions in radiological studies. Moreover, AI-based predictive models can use clinical and imaging data to personalize treatment strategies and forecast patient outcomes. AI-powered analytics may also enhance quality improvement initiatives within otolaryngology by enabling real-time data analysis, performance monitoring, and decision support. NLP algorithms can also extract valuable insights from unstructured clinical notes and EHRs, facilitating the identification of clinical patterns, adverse events, and opportunities for process optimization. AI can also have relevance to health disparities research, health services research, in addition to informing policy decisions and interventions to promote health equity in otolaryngology–head neck surgery. However, safeguards are necessary relating to the inclusion of human subjects, bias in data sets, privacy considerations, and disclosures regarding the use of AI in research. Additionally, the ethical standards of using AI to help develop research methodology, analyze results, and write and review manuscripts require further attention.3

As health care AI evolves, greater resources will need to be devoted to research, development, and monitoring programs. While clinicians should be at the forefront of these advancements, joint efforts among academic centers, larger health systems, and industry will be necessary. Otolaryngologists should also strive to collaborate with other medical specialties, as the multidisciplinary nature of health care implies that advances in other fields will directly and indirectly impact the specialty.

Data Quality, Security, and Privacy

LLMs require the collection of large amounts of data to ensure accurate, consistent, and valid outputs. Limited and error-prone clinical data—often absent, incomplete, or incorrect—compromise model reliability when used for training; thus, the output may be false, biased, or harmful to the patient. Researchers, clinicians, and developers should strive to ensure that high-quality data is used to train these models. Clinicians utilizing AI in medical diagnostics and treatment planning or research must remain vigilant regarding AI hallucinations or confabulations, where the AI generates false or misleading information not supported by its input data.22 This vigilance is crucial to ensure the reliability and safety of AI-assisted medical decisions, safeguarding patient outcomes against potential errors stemming from these AI-generated inaccuracies.

ML in clinical medicine involves the use of algorithms and statistical models to analyze and interpret complex medical data, allowing for the prediction of outcomes, personalization of treatment, and enhancement of diagnostic processes. This approach requires large, unbiased data sets to train these algorithms effectively, ensuring they learn from a comprehensive and representative sample of cases, thereby minimizing biases and enhancing their accuracy and reliability in real-world medical applications. Information security systems are increasingly important to reduce the risk of data breaches and inappropriate management of patient data. Sharing of data should follow strict privacy practices with appropriate deidentification and patient consent. Under current federal laws, personnel and institutions must follow the regulations from the Health Insurance Portability and Accountability Act (HIPAA) to ensure the safe storage and protection of patient data.23 AI-enabled health care platforms, however, introduce new challenges in the realm of patient privacy. Existing HIPAA regulations were largely developed in 1996 and lack safeguards in the realm of digital health.24 Clinician, regulators, and legislators must thus advocate for a modernization of modernize patient privacy regulations to keep pace with this evolving technology.

The legal ramifications of utilizing AI in health care are an evolving area that will have a major impact on health care providers, institutions, and payers. The legal guardrails for all parties remain opaque, and the potential for medical negligence and liability remains a concern. For example, class action lawsuits have been filed against insurance companies utilizing AI to make coverage decisions.25,26 Clinicians must also be aware of the potential legal ramifications of data sharing for AI development, as the legal risks and implications are in flux.27 Similarly, it remains undecided if the clinician must hold an informed discussion with the patient on the use of AI in their care, including the potential risks, benefits, and alternatives.28

Within otolaryngology specifically, the widespread utilization of potentially identifiable endoscopic and microscopic images and videos, radiography, and photography in clinical practice increases the available data for model training. However, the use of these images highlights the additional specific privacy concerns for AI model training within otolaryngology. Similarly, clinicians must be wary of inputting any patient data into non-HIPAA compliant chatbots.29

Federated learning (FL) is 1 paradigm that has the potential to reduce the constraint of limited access to high-quality data.30 To increase access to data, some have advocated for “Data Lakes,” where data is pooled from many different institutions at 1 central site. FL is an alternative approach that enables algorithms to be trained using data from different institutions, without protected patient data ever leaving each institution. This decentralized approach to algorithm development prevents the need to transfer patient data, as model parameters are generated at each institution, and then aggregated across institutions. With this strategy, the amount of data used to train each model would increase without amplifying the risks to patient privacy.

Regulation

Efforts are underway by government agencies and institutions to regulate the use of AI, but the potential applications of AI often emerge in advance of robust regulatory frameworks. This lack of regulation limits the ability to monitor the technology and ensure developers and users are acting in a responsible, ethical, and professional manner. The US government currently uses the U.S. Food and Drug Administration (FDA) Software as a Medical Device framework to regulate AI-based software.31 This is the same act that regulates cochlear implants, operating room microscopes, and endoscopes. However, the ability to create code from anyone’s garage or bedroom and release it to the world instantly makes AI very different and much harder to govern than traditional medical devices. As of February 2024, the FDA has approved fewer than 700 AI-enabled medical devices, 2 of which are in the field of otolaryngology.32 Additionally, generative AI tools used to produce clinical documentation are not yet regulated by the FDA. In March 2024, the FDA released a report titled, “Artificial Intelligence & Medical Products: How CBER, CDER, CDRH, and OCP are working together.”33 In this document, the FDA highlights its strategy to regulate AI in health care, with a focus on 4 major focus areas: “(1) foster collaboration to safeguard public health, (2) advance the development of regulatory approaches that support innovation, (3) promote the development of harmonized standards, guidelines, best practices, and tools, and (4) Support research related to the evaluation and monitoring of AI performance.” This collaborative effort remains a work in progress.

In 2022, the DoD signed the Responsible Artificial Intelligence Strategy and Implementation Pathway, which states that the development of AI should be (1) responsible, (2) equitable, (3) traceable, (4) reliable, and (5) governable.34 Similarly, the White House in 2022 published an AI bill of rights. While both are a step in the right direction, more extensive and comprehensive regulation is needed. More recently, in October 2023, the White House released an Executive Order that promoted the safe advancement of AI in health care, mandated that the Department of Health and Human Services35 establish a reporting program for unsafe uses of AI, and encouraged the development of AI-enabled educational resources.36 As a follow up, the White House in October 2024 published a National Security Memorandum and a companion document, the “Framework to Advance AI Governance and Risk Management in National Security.”37,38 Both provide a more comprehensive outline for the United States’ strategy to remain at the forefront of AI while addressing national security strategies and safety concerns. The memorandum guides federal agencies to develop plans for responsible and ethical development, deployment, and use of AI across various sectors. The companion document establishes 4 key pillars for governance and risk management strategies: (1) AI use Restrictions, which defines prohibited and “high-impact” use cases; (2) Minimum Risk Management Practices, which mandate assessments and safeguards for high-impact AI; (3) Cataloguing and Monitoring, which emphasizes the need for annual inventories of AI; and (4) Training and Accountability, requiring standardized training and oversight mechanisms. The documents also call for global cooperation and the establishment of internationally recognized AI standards and safeguards.

Globally, the WHO in October 2023 published 6 critical considerations for regulating AI in health care: (1) transparency and documentation, (2) risk management, (3) intended use and analytical and clinical validation, (4) data quality, (5) privacy and data protection, (6) engagement and collaboration.39 Using this framework, and until overarching regulation is passed, regulatory bodies and institutions can develop their regulatory principles and guidelines. The European Union (EU) Published the EU Artificial Intelligence Act in March 2024 to provide guidance and regulation for the development, use, and monitoring of AI.35 This Act notably classifies AI applications into 4 risk categories and adopts a risk-based approach to regulating this technology, based on the risk profile of each AI tool. The Act also has extraterritorial implications, as it applies to any developers outside of the EU who create technology that is used within the EU’s borders.

Despite these regulatory efforts, AI technology continues to evolve at a pace that outstrips the development of corresponding regulations, and there remains a notable absence of coordinated global initiatives to address this disparity.

Costs and Sustainability

Despite the potential financial benefits of implementing AI programs across health care, the upfront costs have posed a significant barrier to widespread adoption.40 The costs of AI encompass not only development, implementation, and maintenance but also extend to significant investments in graphics processing units (GPUs) for processing, cybersecurity measures to protect data, and cloud systems for scalable storage and computing resources, reflecting the multifaceted and resource-intensive nature of ensuring AI systems are accurate, secure, and up-to-date. GPUs have become a cornerstone of AI tasks like ML and DL but require high computational power and expenditure. The cost of moving an AI model from inception to patient care can be prohibitive for most centers. This high cost is especially restrictive if a separate model must be created for every clinical question the hospital seeks to answer, as was necessary with most traditional forms of AI. The ability to train generative AI programs to perform multiple tasks, however, may reduce the need to develop a new model for every clinical question. This benefit must be balanced against the higher cost of developing and maintaining generative AI programs.

Developers should also consider the carbon footprint of AI. The training and development of 1 NLP model GPU emits over 626,000 lb of CO2 equivalents, which is more than the lifetime emission of 5 average American cars.41 In 2023, AI already constitutes about 1% of global carbon emissions, and this number is expected to rise as the AI market grows.42

Transparency and Trust

There are numerous concerns regarding the implementation of AI in health care and across other industries, and the lack of transparency of AI models has contributed to some of this. AI models have the powerful ability to run complex algorithms and rapidly generate thought-provoking answers. However, the output these models provide is often opaque, potentially misleading, and sometimes false.43 Traditionally, when an AI model is fed a question, it provides an output without the ability to clearly delineate what data or reason led to that answer. The inability of AI applications to clearly explain their decisions to human clinicians represents a major hurdle, as a core tenet of health care delivery is the ability to review and learn from every clinical encounter and every adverse outcome. This “black box” problem can impact trust and acceptance of this technology.44 AI confabulations and hallucinations are 2 well-known phenomena that also plague these systems.

The widespread implementation of AI necessitates transparent models that can be interrogated, questioned, understood, and trusted. These systems must also adapt to different users and generate effective communication strategies for different actors, depending on whether the model is explaining a finding to a doctor, patient, or other member of the health care team. Concerns regarding the lack of clarity led to the development of the field of Explainable AI (XAI). XAI has 3 major principles, which highlight the need for AI programs to be (1) interpretable, (2) explainable, and (3) transparent.45 Since 2015, the Defense Advanced Research Projects Agency has put considerable effort into developing XAI that interfaces with humans and can clearly explain why specific decisions were made and how the model would act in future similar situations. Models that employ what-if counterfactuals could also demonstrate to clinicians and patients how the model’s predictions or recommendations would vary if specific actions or data points (eg, patient quit smoking) are modified.46

Ethics and Equity

All stakeholders are responsible for ensuring the ethical development, implementation, and regulation of AI. The Stanford Institute for Human-Centered Artificial Intelligence and other leading bodies argue that the development of AI should consider the human impact of this technology and work to augment, rather than replace, human capabilities and experience.47 Leading global organizations, including the United Nations, National Academy of Medicine, and WHO, have additionally published guidelines for the ethical use of AI.33–35 These bodies also argue that developers should strongly consider the societal and ethical implications on patients, providers, and society as a whole. However, the risk of ethical transgression within AI remains high. Ethical concerns in AI may occur in the creation, implementation, and regulation phases of development and impact multiple ethical pillars.

The training process for ML algorithms remains a critical concern due to the potential introduction of bias. It is well known that errors in a training data set are mirrored in the final model.3 Thus, the bias present in a training data set is similarly propagated by AI models, and this use of faulty data has the potential to compound existing biases. This is known as algorithmic bias, which may ultimately and unintentionally violate the principle of nonmaleficence. The algorithmic bias may reinforce stereotypes, perpetuate systemic racism, and hinder diagnostic and treatment outcomes. AI is particularly at risk of selection, confirmation, racial, gender, and skin color bias, among others. Similarly, if the socioeconomically disadvantaged do not equally benefit from AI innovations, then health care disparities will be exacerbated, and the ethical pillar of justice/fairness will be breached.

Every new technological advancement, including AI, has the potential to amplify health disparities and further exacerbate inequality. If the economically marginalized within the United States do not appropriately benefit from these innovations, or if the technology remains within the borders of the United States, then there is a risk for the augmentation of health disparities.21 The high cost of implementing AI systems, the significant memory and storage capacity required to store large data sets and train complex models, and the existing digital divide across populations will limit the impact of AI-based health care services among the marginalized. Existing data sets heavily biased toward specific groups or demographics may also limit the external validity of AI applications across diverse backgrounds.

The importance of ethical development and implementation necessitates that otolaryngologists must continue to raise ethical questions. How can we conceivably develop and advance these tools ethically, and what are the risks we should anticipate? How can we ensure that these tools are developed with the good of humanity in mind and that malicious actors don’t take advantage? What ethical frameworks should guide us, and how do we operationalize them? Providers should take part in exploring the potential biases in existing data and developing ethical frameworks that guide the conduct of all. These frameworks should promote fairness, mitigate bias, minimize harm, and ensure patient autonomy and fairness.

Deploying AI in the health care setting demands a delicate framework for managing accountability concerns. To work in health care, humans must complete training programs, pass board exams, and undergo recertification.48 Medical devices, in contrast, require premarket clearance and post market surveillance. AI does not yet have similar requirements. Furthermore, when humans or medical devices make errors, there can be major consequences, but AI systems currently lack a clear framework for accountability. Existing civil liability models do not clearly determine to what degree, in the continuum between software and a human, the AI system will be treated and held responsible.48 Rigorous oversight and a system that balances leveraging AI’s potential and ensuring trust and safety is necessary.

Technological Determinism

AI’s role in the patient’s care cycle must also be understood. While AI cannot replace clinicians, there is controversy regarding how AI will be integrated into otolaryngology and health care at large. Otolaryngologists play a role in defining the role for AI in supporting clinical activities, surgical training, and research applications. Otolaryngologists should ensure that AI serves as a tool to augment, not replace, their clinical judgment, knowledge, and technical expertise. It is crucial that otolaryngologists maintain skills and acumen, allowing AI to enhance patient care rather than acting as a stand-in for the nuanced, clinician-delivered care.

Conclusion

AI has the potential to shape the future of health care and has the promise of reducing health care costs and waste while improving outcomes. AI-powered health care applications should maximize value-based care initiatives, integrate across every aspect of a patient’s care cycle, encourage transparency and trust, and promote equity. All stakeholders have an enormous responsibility to ensure the safe, ethical, and fair utilization of this technology.

Principles

  1. AI can advance the science and practice of otolaryngology in patient care, education, and research but requires guardrails and governance to ensure safe, equitable, ethical, and environmentally conscious use.

  2. When clinicians use AI tools, the clinician is responsible for exercising sound clinical judgment according to standards of care.

  3. AI should not replace the need for mastery of clinical and surgical knowledge and skills of otolaryngology or the need for continuous learning across the career continuum.

  4. Regular monitoring of models throughout their life cycles and algorithm auditing for performance in different populations are encouraged.

  5. Developers, shareholders, and users of AI should disclose conflicts of interest related to AI products in clinical encounters, scientific meetings, publications, and educational activities.

  6. Clinicians should be prepared to discuss the use of AI, including privacy concerns, in clinical care during informed consent discussions. Clinicians should also advocate for continued safeguards and encourage new regulations to complement or supplant existing legislation.

  7. XAI should be a core feature of AI systems used in clinical otolaryngology. Otolaryngologists should encourage data sets and algorithms to be made publicly available.

  8. Developing and implementing AI requires broad stakeholder engagement from clinicians, patients, computer scientists, ethicists, social scientists, health care institutions, payors, industry stakeholders, and government regulators.

  9. Clinicians, payors, researchers, educators, and other parties should adhere to established and evolving ethical and legal standards in the use of AI.

Funding source:

Automated detection and classification of laryngeal diseases using deep neural networks—R03 CA253212 Project BUR, Andres Martin; Radiogenomic predictors of treatment response in head and neck squamous cell carcinoma—P20 GM130423 Project BUR, Andres Martin; Developing an app-based voice clinical decision support tool to augment the sensitivity of the bedside swallow evaluation in older adults—5Tf K76Actf AG079040 RAMEAU, ANAIS; Bridge2AI: Voice as a biomarker of health—building an ethically sourced, bioaccoustic database to understand disease like never before—1Tf OT2Actf OD032720 RAMEAU, ANAIS. Konstantina M. Stankovic acknowledges support from the Bertarelli Foundation endowed professorship.

Footnotes

Competing interests: None.

References

RESOURCES