Artificial Intelligence in Healthcare: Considerations for Adoption and Adaptation in Academic Medical Settings

1.1 The economic value proposition of artificial intelligence in healthcare

Before analyzing the potential use cases and pitfalls of AI in academic medicine, it is important to first understand the underlying factors driving AI adoption in healthcare. The US healthcare market presents a unique opportunity for AI integration as a means of value creation through increased efficiency, safety, and productivity [6, 10, 11]. According to the Centers for Medicare and Medicaid Services (CMS), National Healthcare Expenditure (NHE) in the US totaled nearly $4.3 trillion in 2021, accounting for approximately 18.3% of US gross domestic product (GDP). Looking ahead, NHE is projected to grow annually at 5.6%, outpacing anticipated 4.6% annual GDP growth over the period of 2022–2031 [12]. Moreover, the market for healthcare products and services faces many challenges and inefficiencies, including, but not limited to a geographically imbalanced distribution of medical facilities and providers, workforce shortages and associated provider burnout, and increasing costs of medical education [13, 14]. In this context, the functional capability of AI to perform tasks that are perceived as menial and/or repetitive tasks may help address many of the aforementioned challenges [3]. In fact, widespread adoption of AI is estimated to be associated with NHE savings of 5-10% annually [6]. Consequently, significant opportunities exist to realize untapped value within the existing healthcare paradigm, with AI being among the most promising and paradigm-changing developments [15].

The overall value proposition presented by the integration of AI into academic healthcare has drawn an influx of dedicated resources, spurring significant new development and investment [16, 17]. This led to the creation of AI platforms that are academic-centric, with emphasis on streamlining processes and improving efficiencies related to academic medicine environments [7, 18, 19, 20]. Specific domains included in this innovation push include medical education, clinical productivity, scholarly publishing, and other related platforms [18, 19, 20, 21, 22]. The corresponding amount of private investment into AI has been especially generous in the healthcare sector when compared to other industry verticals [Figure 1] [16]. Still, important questions persist, including the question of AI’s ability to provide levels of transparency inherently required across our healthcare settings [23].

2.1 Artificial intelligence, machine learning, and large language models: Intersections and distinctions

Since the public launch of Open AI’s ChatGPT (San Francisco, California, USA) in late 2022, and the widespread public exposure to AI, Google ™ (Alphabet, Inc., Mountain View, California, USA) searches for the term “artificial intelligence” increased by nearly 250% from ChatGPT’s launch to April 16, 2023 [24]. Likewise, the term “large language model” increased by nearly 592% during the same time period [25]. Thus, the introduction of the end-user-oriented large language models (LLMs) coincided with notable increases in public interest around AI topics. In parallel, other important actors entered the market, including open-source platforms such as Ollama (Toronto, Ontario, Canada) and H₂O AI (Mountain View, California, USA). These platforms offer subscription-free and local LLM functionalities to end-users, thus enabling truly autonomous and private use of their models for the general public free of charge.

The term “artificial intelligence” was first coined in 1956 by computer scientist John McCarthy [26]. Artificial intelligence broadly refers to the ability of a computer to perform tasks typically associated with intelligent beings [27]. As per its definition, AI encompasses any non-naturally occurring system that displays capabilities typically associated with intelligence: communication, self-awareness, world knowledge, intentionality, creativity, and learning [28]. Artificial intelligence can be represented through a variety of end-user technological manifestations and implementations, as outlined in previous sections of this chapter.

Machine learning represents a distinct facet of the larger AI umbrella. Machine learning (ML) is a branch of computer science that allows computers to learn from, and derive outputs using data-driven relationships autonomously [29]. Given its functionality, ML is often leveraged as a prediction and classification tool with applications including natural language processing, computer vision, and omics learning models [30, 31, 32]. Provided with a large set of data, ML algorithms attempt to identify relationships between variables within a dataset, and the machine can then make predictions about the values of an outcome variable based on a given input or set of inputs. For example, ML may be able to train itself on the relationship between body mass index (BMI) and the probability that a patient has type 2 diabetes mellitus [33]. When provided with sufficient data, ML algorithms will be more likely to identify a meaningful relationship between input variables [34]. The algorithm may then be used to predict, for example, the probability that a future patient has type 2 diabetes mellitus given information about his or her BMI [29]. Ultimately, ML is a subset of AI that centers on a computer’s ability to autonomously learn associations between pieces of data.

Large language models (LLMs), also often referred to as natural language processors (NLPs), are text-specific applications of ML algorithms that have become a popular end-user product and consequently warrant dedicated discussion. Large language models, such as generative pre-trained transformers (GPTs) or Bidirectional Encoder Representations from Transformers (BERT), can generate coherent text in response to a given end-user prompt [35]. As a form of ML, large language models are trained on large amounts of text that are fed into algorithmic architecture to guide the machine in developing associations between words. By understanding the association between words, pertinent grammar and other language rules, the machine can provide a coherent response to a verbal prompt by predicting which words may be most appropriate in a response based on the LLM’s trained associations. All to say, large language models are a specific application of machine learning, which are trained specifically on text-based data (Figure 2).

Understanding the connections between AI, ML, and LLMs is important to understanding the scope of applications for which each may be employed in the healthcare field, including medical academics and related domains. The applicability of AI technologies can vary widely across the spectrum of healthcare and academic medicine. Therefore, a framework may be useful to physician executives, hospital administrators, medical researchers, educators, and practitioners for guiding considerations around the implementation of AI in healthcare settings.

3.1 Considering applications and pitfalls of artificial intelligence in healthcare: What? Where? And Why? Framework

Artificial intelligence has potential applicability across many niche functions of healthcare delivery and medical academics [3, 8, 36]. Consequently, a framework that organizes the functionality of the AI technology, the context in which it is being used, and the end goal (or intended goal) of each use-case helps guide our considerations for AI implementation. In other words, when thinking about AI integration in healthcare and academic medicine, we propose a framework to consider what technology is being used, where it is being implemented, and why it is being used.

3.1.1 What technology is being used?

Artificial intelligence in healthcare and medical academics has many potential applications based on the functionality of AI technology being deployed. Particularly, ML technology in the forms of LLMs, computer vision models, and omics-specific learning models have demonstrated potential use cases in academic medicine (Figure 3) [37, 38]. These forms of ML technology function as pillars upon which many healthcare end-user AI applications are built, and therefore, understanding the functionality of each can be beneficial in identifying their potential use cases and inherent flaws [3].

Fundamentally, LLMs are able to interact with written text including the ability to help edit sentences, aid in translation, and respond to prompts by a third-party user [39]. Large language models such as ChatGPT, LaMDA, and BERT produce text by predicting the appropriateness of every “next word” based on associations the model makes when trained on millions of lines of human-produced text [40, 41]. Functionally, this leads to the “context problem,” where the model does not actually “comprehend” the text in terms of its semantic context; it merely produces probabilistically “optimal responses” based upon the data on which it was trained [42]. Therefore, the architectural development of LLMs lends itself to a few generalized pitfalls of such models including reliability concerns (the propensity of the model to provide inaccurate or non-factual information), consistency issues (the ability of the model to produce different responses to the same prompt), prompt-interpretation complications (the sensitivity of the model to respond differently to minor variations in prompt phrasing), response bias (the tendency of a model to produce biased responses as a reflection of its training data), and black-box explainability barriers (the inability to concretely explain why the model responds the way it does) [42, 43, 44]. What follows is the realization that any implementations of LLMs in academic medicine, especially when it comes to the requirement for high-fidelity, quantifiable, reliable, and ethical data processing, may be inherently flawed without the a priori definition of strict curbs to be utilized in the academic research process [45, 46, 47]. After all, the LLM itself may not be “aware” of the boundary between “true” and “confabulated” reality [47, 48, 49].

Computer vision (CV) serves as a technological tool to many end-user AI applications in the medical field, especially within the academic, educational, and research domains [50]. Computer vision refers to the ability of a computer to process and identify images and video, much like humans do, and hence, CV can be leveraged for object detection, facial recognition, image classification, reverse image searches, and more [51, 52, 53]. Computer vision is an attempt to develop computer-based perception, which is relatively rigorous compared to the development of computer-based logic, as in the case of LLMs. In order to develop a CV model, computers must be trained on thousands of labeled images or videos, which allows the machine to develop an association between the image and the label [51]. This lends itself to a few potential pitfalls. First, the reliability and accuracy of a CV model are, in part, a function of the size of the training dataset. However, obtaining images, particularly those of human subjects, with appropriate consent can be both logistically difficult and costly. This presents a quandary (and an important trade-off) between improving model accuracy and developing a model in a timely, cost-effective, and ethical manner. All to say, CV models can be error-prone. Additionally, CV models are particularly vulnerable to identification bias. Even with sufficient training data, each image must receive a specific label. Incongruent with the adage that “a picture is worth a thousand words,” images in CV training receive a single label. Consequently, labels can be subject to the biased perspective of the labeler. Furthermore, the task of labeling, despite its notable downstream influence on the model’s functionality, is often delegated to third-party contributors outside of the primary development pipeline of the model [52]. Inaccurate or biased labeling may propagate and be reinforced through the use of the model, creating potential hazards.

Omics-targeted AI technologies aid in the characterization of biological products. The field of omics refers to an array of biological subfields that study the molecular processes occurring within the various and diverse constituents and components of living cells [54, 55]. Omics fields include genomics, proteomics, metabolomics, and more. Omics-targeted AI models have shown the ability to facilitate omics research by efficiently organizing, processing, and deriving relationships within large datasets necessary to analyze omics processes [56]. The primary challenges inherent to omics models are statistical in nature. This is because omics models are often built upon a large number of variables/parameters that must be inherently accounted for and accurately represented within the very complex and intertwined matrix of biomolecular processes. Given the large number of parameters that need to be addressed, developing models that are sufficiently robust and reproducible over multiple independent studies is difficult [57]. In statistical practice, training a model in which the number of samples (n) used in training is less than the number of parameters (p) in the model creates a problem with respect to the reliability of the model. This relationship between n and p should be considered critical for omics models, given the significant number of parameters used and the challenges associated with obtaining clinical and experimental data. Lastly, a key challenge inherent to omics learning models, given their size and complexity, lies within interpretability. The utility of a model can be hindered if it is difficult to understand and interpret the relevance of the results of the model [58, 59].

Ultimately, understanding what technology (LLM, CV, or omics model) is being used can allow individuals to assess generalized pitfalls and use cases for the technology, and is hence, a starting point for the overall framework.

4.1 Overview of applications and pitfalls around artificial intelligence in clinical practice

The integration of AI into both clinical and academic domains of medicine has the potential to transform many aspects of clinical practice, education, and research, while simultaneously addressing some of the challenges in healthcare delivery. Applications vary widely in their utility and potential benefits, including clinical documentation optimization, improvements in educational experience, more streamlined and effective research, and augmented diagnostic imaging capabilities, among other areas [3]. However, regardless of the specific technology at hand, integrating AI into relevant settings is a major endeavor that comes with inherent risks that should be considered, managed, and mitigated to prioritize any potential downstream consequences, from patient safety issues to research biases [60].

Both LLMs and natural language processors (NLPs) have the potential to provide numerous administrative and clinical benefits to diverse groups of users, including medical students, medical practitioners, researchers, allied health professionals, and patients [77, 78, 79]. Given the importance of fluid communication between various stakeholders, LLMs may serve to bridge communication gaps. For example, LLM-based paradigm may help improve patient understanding of their medical record by simplifying complex medical language [62]. In addition, LLMs may assist healthcare providers in drafting clinical documentation, summarizing dense medical text, and aiding in diagnosis and triage [80, 81]. Additionally, large language models have the ability to facilitate administrative clinical tasks, such as the extraction of pertinent information for medical billing and insurance authorization approvals, as demonstrated in Figure 5 [82, 83]. Prior authorization and clinical documentation are major contributors to physician burnout, and automating parts of these processes may lead to a reduction in administrative burden while also reducing costs and improving the patient-provider experience [84].

However, it is important to understand the pitfalls associated with the implementation of large language models in clinical settings. First, the general ability of LLMs to “fabricate” information and promote trained biases presents a consequential concern that should be ultimately mitigated through expert oversight and hardwired quality control tools [76, 85]. Furthermore, while these tools may have the ability to augment patient-provider interactions, they cannot replace practitioners in building trust with patients [76]. Additionally, over-reliance on LLMs to guide diagnosis and medical triage may impair a provider’s critical thinking skills over time [86].

Computer vision technology within clinical medicine has demonstrated potential to aid in radiologic, pathologic, and dermatologic image recognition [3]. Such image recognition tools may help physicians detect abnormalities sooner or with more efficiency [87]. Through potential productivity enhancement, these tools may help expand care in the face of physician shortage across specialties including radiology and pathology [88]. However, the consequences of CV error and propagated bias are heightened in the clinical setting, particularly if such missteps compromise patient safety. In the event of misdiagnosis by a CV model, it is currently unclear to whom or what liability would fall [89]. Ultimately, potentiality for error, concerns of patient-provider trust, and regulatory hurdles, prevent a fully autonomous employment of diagnostic imaging tools in the present [90].

Additional uses of CV within clinical practice include the development of immersive experiences for patients and providers using augmented reality (AR) and virtual reality (VR). Augmented reality may be used to guide individuals through surgical procedures or emergency response techniques through the use of lenses that provide direction and feedback to the user [3]. Such technology has the potential to further advance opportunities in minimally invasive surgery. Minimally invasive surgery has become a focus of medical research as it presents an opportunity to mitigate operative trauma and expedite postoperative recovery. However, the consequence of achieving these improvement outcomes is a more difficult surgical task with limited vision and restricted vision. Augmented reality tools have the potential to relieve this tradeoff by increasing viewing area [91]. Augmented reality technologies have already been employed in certain intraoperative settings in the field of spine surgery and will most likely continue to expand initially in orthopedic and neurosurgery due to the immobility of the skeletal system and the spine [92]. However, with the rapid advancement and increased use of peri- and intraoperative imaging in thoracic and abdominal surgeries, AR could potentially impact a much broader range of surgeries [92, 93, 94]. Besides aiding physicians in providing clinical care, AR and VR tools can directly impact patient experiences. For instance, by providing an escape to virtual worlds and environments, VR immersive experiences have shown benefits in improving therapeutic outcomes in patients experiencing debilitating or painful illnesses such as chronic stroke. Furthermore, immersive AI technology used as an important adjunct has demonstrated partial relief to late-stage cancer patients (Table 1) [3].

Applicability of omics AI models within a clinical setting lies primarily within the field of precision medicine—the ability to tailor healthcare interventions to an individual based on genomic composition, age, gender, geography, race, family history, immunity profile, metabolic panel, microbiome, and environmental sensitivities [3]. While precision medicine may boast a number of envisioned benefits including improved drug response, minimized complications, and reduced costs, the ability to develop sufficiently reliable omics models under social and economic constraints remains difficult [95].

4.2 Overview of applications and pitfalls around artificial intelligence in modern medical research

Artificial intelligence can enhance the process of medical research by streamlining clerical tasks and presenting new tools for data analysis. Large language models in medical research can assist in many clerical tasks associated with the research process including enhancing vocabulary, rectifying grammar errors, annotating documents, and scraping text-based information [76, 96]. A particularly powerful use case of LLMs is the ability to summarize and condense medical text. This can be of benefit to researchers when developing annotated bibliographies, performing integrative reviews, polishing language/grammar, and more [76]. A notable concern in the use of LLMs in medical research is the ability of the model to “fabricate” or overgeneralize research insights, resulting in inappropriate conclusions and pointing to a need for continued critical review of the overall research process by humans. Perhaps transparent disclosure of the use of AI in the research process can aid in identifying potential instances of fabrication or bias [76]. The ability to track and correlate the potential effects and biases of AI-based implementations on research quality and fidelity will be an important first step in ensuring that robust error-reduction frameworks can be operationalized in the future (Table 2).

Besides aiding in streamlining various research tasks, AI may be able to improve research study design. For instance, patient selection is a vital component of clinical trials and presents one of the main bottlenecks when it comes to trial completion [97]. Various barriers exist, including overwhelming medical jargon and the non-centralized distribution of information on clinical trials [98]. To aid patients in finding and understanding available clinical trials, NLP programs can be leveraged to mine available clinical trial databases for eligible studies and to translate these studies to improve patient understanding [98]. Other potential uses for AI in research design include determining the novelty of research questions and assisting in statistical analysis.

Artificial intelligence can also directly influence medical research in potentially powerful ways. Large machine learning omics models can be leveraged in medical research to enhance the study of biomolecular processes. Given their functionality of identifying relationships between parameters, especially in multi-parametric and highly complex systems, ML omics algorithms can be leveraged to identify relationships between gene expression and other EMR data [3, 99]. Another notable area of applicability for omics ML models falls into drug discovery and development [100]. For instance, researchers may leverage such models to predict drug activity, target interaction, and other properties, which may aid in the molecular design process [3, 100]. The key challenge in this context consists of developing cost-effective models that are sufficiently reliable to meet safety requirements for human research subjects and, ultimately, patients receiving the resultant therapeutic agent [58].

4.3 Overview of utility and limitations of artificial intelligence applications in today’s medical education

There has been significant development in education focused AI technologies over the past two decades influencing the methodologies by which students learn and curricula are designed [101, 102]. The introduction of “teacher bots,” which are effectively education-focused chat bots that serve as individualized tutors, has provided a mechanism for students to address learning needs at their convenience [103]. This area of development is bound to be highly controversial for a number of well-founded reasons [104]. Education underpins the foundation of modern medicine, from undergraduate coursework to post-graduate studies and lifelong continuing medical education (CME). Given the educational nature of the field, medicine presents a unique opportunity to introduce educational AI tools across the entire learner continuum, and realization of this opportunity has been reflected by increased enthusiasm and research around the topic of AI in medical education [44, 105].

Artificial intelligence tools in education can take on a variety of forms including intelligent tutoring systems, integrative learning environments, assistant chatbots, grading and assessment tools, diagnostic assistants, analytics monitors, and medical simulations [106, 107, 108]. In addition, when properly and safely implemented, with appropriate checks and balances that focus on content validity and model fidelity, AI-based educational tools may allow entire regions of the globe (especially in low-resource environments) to make an “educational leap” not dissimilar to the leap made when highly restricted traditional telephony gave way to omnipresent cellular telephony across the world [109, 110, 111, 112].

Intelligent tutoring systems and integrated learning environments are learning tools that adaptively expose students to learning content based on the students’ level of mastery as perceived by the AI model overseeing the learner’s progress. Perhaps the most readily apparent use of artificial intelligence in medical education is the use of AI in generation of practice or assessment questions. Some studies have already investigated the use of generative AI in producing multiple-choice, board-style questions for undergraduate medical education, and although AI-created questions generally still required further human editing to reach satisfactory accuracy and comprehensiveness, this process was still more efficient than traditional question generation [113]. In the future, these use cases may be extended to include generation of open-ended questions. Consequently, these tools have the potential to aid in the delivery of lesson plans and provide the opportunity for a “flipped classroom” in which each student has the opportunity to learn at their own speed, thereby potentially enhancing student outcomes and improving the productivity of time in-class [74, 114]. However, notable pitfalls inherent to AI-based tutoring systems include the hindrance of social interaction, detriment to human communication skills, and inability to monitor students’ body language [114]. Finally, over-reliance on intelligent tutoring/learning systems may result in both quantitative and qualitative declines in the educational workforce. In this context, intelligent tutoring systems and integrated learning environments are best suited as supplementary educational tools, and appropriate mechanisms should be utilized to ensure that the AI educational revolution does not result in irreversible brain drain and talent erosion.

Educational chatbots can serve to field questions from educational stakeholders. Chatbots are end-user-focused LLMs that can aid medical students in the individual learning process. This on-demand access to receive answers to questions may improve the productivity and the experience of medical students, while reducing the burden on education providers; however, there are pitfalls to consider [74]. Leveraging, the What, Where, Why? framework, we must remember that currently, available LLMs have the ability to “fabricate” information, which can be further compounded by a lack of transparency regarding the generative process itself [115, 116, 117]. In addition, the constant and ever-accelerating growth of the existing medical information available to LLMs for processing and use, would require a great deal of oversight and real-time model updates, inclusive of purging of information that is no longer valid and/or known to be harmful [118, 119]. Further, concerns linked generally to ML algorithms may present students with a gap in clinical reasoning. In other words, it must be ensured that intelligent tutoring systems in medical education are capable of explaining the reasoning behind the answers they generate in order to provide students with proper cognitive intuition [44]. The potential for the provision of misinformation coupled with potential barriers thus requires another reference point for students to engage [115].

Another potential benefit of AI-based tools, particularly LLMs, in medical education is within the area of grading and assessment of students [120]. Artificial intelligence can provide a mechanism for grading assessments without the influence of a human grader for questions beyond multiple choice and other more traditional standardized testing. This, in theory, presents the opportunity for educators to assess students in a more open-ended fashion (i.e., free response questions and essays) without the concern for ease of grading and/or standardization of the assessment [74]. At the same time, given concerns inherent to LLMs, understanding why the grading tool assigns a specific grade to a student may be problematic, and therefore, all final grades may still need to be reviewed by a human grader [121]. More so, the nature of assessments in medical education is varied, and thus, AI-based grading systems may not be useful in all contexts of student evaluation, such as in the assessment of a student’s performance on a physical examination [44]. Nonetheless, such grading tools may be implemented to provide a first-pass review of a student’s work, thus improving the overall systemic efficiency.

Artificial intelligence also poses benefits to simulation-based medical education [122, 123]. Tools including augmented reality resuscitation simulators and AI-simulated patients, in theory, present the opportunity for medical and surgical trainees to practice skills related to clinical evaluation and surgical procedures in a seemingly realistic context without the potential harm to an actual patient [86]. In the future, surgical simulation software may also be able to track how well students are performing or making progress, highlighting areas for improvement and even predicting the number of repetitions needed to master certain skills [124]. These simulations could be further supplemented with CV algorithms to track learner movements, leading to the possibility of providing feedback on surgical dexterity and motion economy [125]. However, the key challenge in this dynamic is training future doctors, who will be responsible for the care of human beings, using a training platform or an educational entity that is intrinsically non-human. This may create both potential biases and “unintended standardizations” of future doctors with less exposure to the real-life diversity of physical and social differences among patients – something that healthcare providers are bound to experience throughout their careers. Despite the above risks, some researchers have proposed ways that AI could be effectively leveraged to enhance empathy and compassion among both trainees and future physicians, such as through immersive experiential simulations of conditions like old age or disabilities [126]. Still, these technologies must be further improved before widespread implementation in medical education. In aggregate, our current knowledge supports the use of simulated/robotic patients primarily as supplementary tools, but by no means a replacement for real-life bedside or clinic experience (Table 3) [127].

In summary, exposure to AI during medical education, if implemented with appropriate guardrails, may also serve to equip the next generation of practitioners with the technical and cognitive skills, and the understanding needed to use AI-based technologies in future practice. The overall comfort levels and understanding of AI technology by providers is key to integrating AI into clinical practice, and this is optimally established during medical training [128].