This article is a summary of a Masterclass webinar supported by McGraw Hill, which took place on Monday 29 September 2025.

Speakers:

Joseph Loscalzo, MD, PhD

Samuel A Levine Professor of Medicine, Hersey Distinguished Professor of the Theory and Practice of Medicine, Harvard Medical School; Physician-in-Chief Emeritus, Brigham and Women’s Hospital, Boston, MA, United States.

Carol A Langford, MD, MHS

Harold C Schott Endowed Chair in Rheumatic and Immunologic Diseases, Cleveland Clinic; Professor of Medicine, Cleveland Clinic Lerner College of Medicine of Case Western Reserve University of Medicine, OH, United States.

Masterclass Summary

Artificial (or augmented) Intelligence (AI) is reshaping how clinicians care for patients and how they themselves are trained. This article explores what that means for practising physicians and educators, and how Harrison’s Principles of Internal Medicine (HPIM) is positioning itself in this landscape.

Part I

Framing AI as Augmented Intelligence

For this educational webinar, AI is best described as augmented intelligence, emphasizing its role in extending human clinical reasoning rather than replacing it. AI aims to mimic human intelligence and sometimes challenge clinician intuition by uncovering non linear relationships in complex, high dimensional data that are not obvious on routine review. AI has already been shown to be effective in several key domains, including communication, knowledge representation, reasoning based on prior data, learning (including machine and deep learning), computer vision, and robotics. Each of these domains already has, or may soon have, direct implications for clinical practice, from diagnostics and therapeutics to workflow and health system design. 

A useful way to appreciate AI’s potential is to consider how computational capacity has expanded. Over recent decades, computing power measured in floating point operations per second (FLOPS) has increased from mainframes shared across universities (3.4 x 104 FLOPS) to modern supercomputers approaching human brain scale estimates (supercomputer: 1.25 x 1017 FLOPS; human brain: ~3 x 1017 FLOPS), and more recently, emerging quantum devices (~1012 faster than a high-performance system), although these are still in a nascent state for practical application This increase in computing power is what now makes training these models possible with potential relevance to the practice of medicine.

AI and Conventional Machine Learning

Conventional machine learning typically focuses on predefined features: the user specifies variables of interest, and the algorithms learn correlations between them. Logistic regression is a classic example in clinical work where a binary outcome is predicted (based on patient data), such as the presence of absence of disease.  In contrast, modern AI systems can process thousands of variables simultaneously, discovering patterns, clusters, and associations in very high dimensional spaces without the user pre specifying which relationships to test. A simple way to compare these approaches is to consider logical association versus intuition. Intuitive clinical reasoning draws on experience to spot subtle or complex patterns, much as AI systems do when analyzing large datasets.

Clinical Machine Learning

Supervised learning remains the most familiar form of AI in medicine and is one of classification (sorting x from y) and regression (x predicts y). In this approach, inputs such as image pixels or laboratory values are paired with labelled outputs, for example a diagnosis or outcome, with models trained to predict those labels.(1) Convolutional neural network analysis exemplifies this methodology, for instance, in the  histopathological image analysis and diagnosis of lymph node metastases in breast cancer.(2)

Unsupervised learning focuses primarily on clustering of variables (identifying similarities) and includes principal components analysis (PCA), a contemporary example of which is its application (by dimensionality reduction) in the pre-training of large language models (LLMs) where the model recognizes the sequence regularity in words, tokens, or other units to predict “what comes next.” An exemplary clinical application includes learning medical knowledge from large-scale bodies of text, or visualizing structure in gene-expression data.

Between supervised and unsupervised machine learning lies semi-supervised learning and self-training approaches. A contemporary example of machine learning is consistency regularization, seen in the clinical application of cardiac magnetic resonance imaging and clinical datasets for diagnosing hypertrophic cardiomyopathy (HCM). Deep reinforcement learning, originally illustrated by systems such as AlphaGo Zero (the first machine learning strategy that learned how to play the game ‘Go’ by playing against itself), explores multiple outputs best exemplified classically by optimal control, and has been proposed for the management of sepsis. 

Current Clinical Applications

The clearest example of AI in medicine is in image analysis, where models recognize patterns on  X-rays, histological slides, retinal scans, and electrocardiograms among others. 

Other established or emerging applications include:

  • Construction of differential diagnoses from complex constellations of symptoms, signs, and features.
  • Drug discovery and optimisation of therapeutic regimens.
  • Analysis of temporal trajectories of disease and treatment response, sometimes using digital twins or synthetic patients in clinical trials.
  • Real time support for procedural decisions, such as robotically guided biopsies or stent placement.
  • Integration and interpretation of genomic and multi omic data in the investigation of undiagnosed conditions.

AI in ICU Outcome Prediction

A 2023 report(3) illustrates how AI was used in ICU decision making. Bedside monitoring data and serial laboratory results create extremely high dimensional, time dependent datasets. In the study, the question asked was whether patterns in the data could predict the likelihood of 3-day survival or death at the time of admission. In this dataset, the predictive algorithm performed well, but the model initially produced only a simple ‘thumbs up/thumbs down’ output. It did not explain which features were driving the prediction or how they might be modified, so its clinical value was limited. Through application of the Shapley Additive exPlanations (SHAP) test to the same dataset, it was then used to identify directionality of the features primarily responsible for the outcome, along with their rank order. The SHAP analysis is the subject of this 2023 report and nicely illustrates a methodology by which to circumvent the typical ‘black box’ output of earlier AI analyses, providing recognizable and actionable dataset features or parameters that are responsible for the outcome of interest.

AI can be used for decision making in a variety of settings. There needs to be an adequately rich dataset from which the AI engine can learn and, as a clinician, ensure that those specific outcomes might provide some insight into remediation.

Digital Twins in Health Care

Another area of growth in the application of AI to medicine is the notion of digital twins which originated in aerospace and manufacturing  in the 1960s, when scenarios that informed real-time responses to situations in space were developed for astronauts.(4) The same principles, now with much faster approaches that go beyond conventional machine learning, is being adapted for healthcare. 

In medicine, a digital twin links an individual patient’s real world data to a virtual representation derived from large populations of patients with similar characteristics. The twin can be used to explore likely disease trajectories, responses to therapy, surgical planning, device performance, or alternative clinical trial designs. A major aspiration is to use deeply phenotyped and, in some contexts, genotyped virtual cohorts to reduce the need for large real world trial enrolment, thereby limiting risk and cost.

However, for digital twins to be clinically credible, the underlying virtual populations must be accurately and comprehensively characterized. Limitations or biases in the source data will propagate into any inferences about a given patient.

Multi-Omics and Disease Redefinition

Modern genomic and multi omic programmes, including those that are part of contemporary clinical trials, now gather transcriptomic, proteomic, metabolomic, epigenomic, and deep phenotypic data for individual patients. These datasets are vast and high-dimensional, and their true value lies in understanding both within layer and cross layer relationships.(5)

This field is in its infancy, but one emerging outcome includes redefining diseases. Rather than accepting histopathological categories to identify or define a cancer from a supervised approach, AI driven clustering across multi omic and phenotypic layers can identify groupings that may better predict prognosis and treatment response and, over time, may fundamentally change how medicine is practiced in the future. 

AI in Health Care Delivery

Beyond diagnostics and biology, AI is being used in more practical strategies to improve healthcare delivery and operations.(6) Applications include:

  • Optimizing point-of-care and referral pathways.
  • Improving the performance of value-based care models.
  • Automating and streamlining administrative functions.

Uptake varies widely across healthcare systems and countries, and from a public health perspective, AI has already been used for the identification and surveillance of outbreaks, tracking cases and outcomes, optimization of clinical trial performance, retrieval and analysis of medical information, image analysis, and operational organization.7 In medical practice, AI is being actively applied to facilitate documentation during patient visits. In the future, more deeply integrated AI approaches may support real-time differential diagnoses and patient-clinician-AI interactions. One provocative concept is the “virtual clinical partner,” an AI agent that continuously learns from a clinician’s practice patterns of how they take histories, interpret data, and make decisions. The goal is for the AI agent to handle routine work when the clinician is otherwise occupied with emergencies or time-consuming complex cases, for example. Although the idea is technically attractive, it does raise questions and concerns surrounding patient safety and accountability, with the added loss of the human touch in patient care.(7)

Benefit, Limitations, and Bias

From a clinician’s standpoint, potential benefits of AI include improved efficiency, mitigation of some cognitive biases in differential diagnosis (such as framing), and the possibility of reallocating time toward direct patient interaction rather than routine data handling. If implemented thoughtfully, AI could help restore aspects of humanism in medicine unless it is used to move medicine toward more transactional, reflexive, rapid approaches to diagnostics and therapeutics, which is a real concern.

Several limitations are important:

  • An a priori assumption that disease definitions are correct which may change as multi-omic and phenotypic data evolve.
  • Standards for application and testing of AI tools  remain immature.
  • Rare diseases and atypical presentations often lack sufficiently large, high-quality training datasets, and will continually require human curation. 
  • Algorithmic biases that enter machine learning strategies despite best efforts to minimize them.
  • Regulatory, legal, and ethical frameworks are still in early development and lag the technological capabilities.

In an example of bias in a learning set, an early published study of using image analysis to diagnose melanomas demonstrated that the initial algorithm was highly reliable in predicting lesions that were ultimately demonstrated to be invasive melanoma.(1) However, when the feature analysis was performed, the most powerful driver of that correct diagnosis appeared to be the presence of a ruler as a predictive feature (rather than any visual features of the melanoma itself), effectively using clinician concern and the act of lesion measurement as an indirect signal that correlates with a diagnosis of melanoma. This example underscores the responsibility of those curating training data to exclude spurious cues and ensure that clinically meaningful features drive model performance.

Ongoing challenges include managing continuous algorithm updates. Users need to be aware of the advantages and disadvantages of each update and be able to intervene as illustrated in the Shapley analysis for the ICU dataset. We also need to be alert to risk failures, adverse events, and data reliability issues familiar from conventional statistical approaches. Acceptable thresholds of accuracy, the risk of over-interpretation,  “hallucination,” and questions of liability and user education will all require shared decision-making among stakeholders.(8)

Part II

Medical education is an area of great interest and excitement for innovation with AI, opening the door to opportunities for learning and personal development, from curriculum to competency assessment.

Core Opportunities

Curriculum Redesign and Personalization

Traditional-based teaching methods of  “lecture and listen” will remain part of the educational toolkit, but AI broadens the range of available approaches. We have already seen the development of the flipped classroom, and AI offers another option for educators to tailor medical education content to learners’ needs and preferences. 

Diagnostic Assistance 

AI could diagnose and address knowledge gaps through analyzing learner performance across assessments and clinical encounters. In addition, AI systems can introduce the range of diseases (particularly rare disorders and uncommon presentations) early in training, improving recognition and differential diagnosis skills later in clinical practice. 

Competency Assessment and Trainer Development

AI-supported assessment tools can evaluate what a learner understands and how well they can explain and teach that knowledge to others. This “train the trainer” model, arising from the concept of watch, do, and teach, leverages the principle that explaining to others deepens an individual’s understanding and reveals gaps in knowledge. 

Multi-Media Enhancement

Lecture recordings or expert-authored content can be transcribed and repackaged by AI into formats optimized for individual learners such as summaries or expanded text or presented in video or infographic formats. 

Administrative Streamlining

Beyond pedagogy, AI can assume burdensome administrative tasks, freeing educators to focus on mentorship. 

Precision Medical Education

AI can also enrich educational tools including virtual reality and point-of-care ultrasound training as examples. These opportunities offer a guide for precision medical education whereby the right content is delivered, at the right level, at the right time, and in the right format for each learner. The goal is to support knowledge integration and retention while maintaining learner engagement throughout their training journey.(9)

Critical Challenges

Several important concerns must be acknowledged:(10)

Data Ownership and Governance

As new learning tools are developed using AI, clarity about data ownership and control is essential. How will data ownership be defined and assigned?

System Integration

Effective implementation across different institutions and educational infrastructure is critical, requiring careful attention to technical compatibility and workflow integration so that AI tools enhance rather than disrupt existing systems. 

Source Bias and Population Alignment

The medical information underlying an AI learning tool reflects the databases and range of literature sources on which it was trained. If the body of training is too narrow or unrepresentative, biases can propagate. 

Learner and Patient Privacy

Privacy concerns exist for both learners and patients. Learner data, including knowledge gaps and performance, must be protected against inappropriate disclosure, while patient materials need robust safeguards to ensure that anonymity is maintained.  

Content Synthesis and Retrieval

The mechanisms by which AI accesses, processes, and synthesizes information for medical education use require transparency. 

Preserving the Clinician-Patient Relationship

Perhaps the most sensitive concern is that the communication between provider and patient represents one of the core foundations of medical practice, and any AI-supported educational approach must be designed to enhance, rather than undermine a learner’s ability to communicate and empathize with patients.

Responsiveness to Individual Learner Needs

The calibration of content generation and personalization must be robust for an AI system to consistently detect a specific learner’s needs and to be a reliable and effective learning tool. 

Advantages and Strategic Positioning

Despite these challenges, AI offers substantive advantages for medical education through rapid, scalable content delivery. AI can quickly deliver high quality, relevant,  personalized content aligned with the educational goals of both the institution in which the learner is enrolled, and the learner’s personal goals.(10)

While it will be important to have a very broad education, by tailoring content to learner performance and interests, AI-enhanced systems can enhance learner engagement.

Precision medical education tailors the learning to an individual’s needs based upon performance progression, and personal interests. AI can also process unstructured data and personalize that content to ensure the learner remains engaged throughout the educational process. However, care must be taken to maintain comprehensive, factually correct, up-to-date information that ensures optimal learning.

Harrison’s Principles of Internal Medicine (HPIM) and AI

A key question emerges in the age of AI regarding the role of HPIM. Comprehensive, factually correct, and up-to-date information will always remain the foundation of this authoritative medical education resource, which HPIM will ensure as it develops its own enterprise-based AI system. 

HPIM will continue to adapt to region-specific educational needs and clinical priorities, ensuring relevance as a key resource throughout the world. Additionally, with the many years of experience of HPIM as the definitive source of information for medical educators at all levels, HPIM is well poised to adapt to this rapidly evolving field for the benefit of learners and for society.

Moving Forward

AI in medical education evokes both excitement and apprehension, and a reflection from Seneca is apt:

“It is not because things are difficult that we do not dare; it is because we do not dare that they become difficult.”

By confronting the challenges head-on, we can harness AI as a powerful educational tool. The key is judiciousness, applying AI as a tool to support learning and administrative efficiency while safeguarding privacy and ensuring factual accuracy. Most critically, preserving  the vital relationship that exists between providers and patients which must always remain the foundation of medical practice.

Key Themes from the Q&A

Integrating AI into Curricula Without Losing the Human Element

Participants raised concerns about whether AI in medical education might erode traditional training that centers on human interaction. 

Panel response:

AI and human-led education must be designed to work together, not in competition. AI tools should support program goals and content design while keeping human supervision at the core. Educators must remain responsible for monitoring learner progress, curating the educational materials used by the learner, and developing the training skills that machines cannot teach, such as interpreting subtle verbal and non-verbal cues of anxiety or misunderstanding during contact with patients. AI can enhance learning, but guiding the importance of empathy and professional judgement must continue to be a human-led priority in medical education. 

Everyday Clinical Use of AI Now and in the Near Future

Participants asked how AI is already being used in routine care and how this might change over the next year. 

Panel response:

Currently, common applications include image-based diagnostic support and analysis of potential therapeutic options for different malignancies. Both areas, in addition to operational applications, are likely to continue to evolve rapidly.  Also, using an AI partner for recording and documenting information from a medical interview can reduce the administrative burden of note-writing for clinicians, provided the transcription and summarization are accurate and clinically reliable. The emphasis here is to free clinicians to spend more time with patients, listening, examining, and maintaining the human touch. A major risk, however, is that institutions might respond by increasing patient throughput at the expense of utilizing that time for thoughtful medical practice.

Using AI to Enhance Rather Than Erode Learner Skills

Another concern was whether learners might become over-dependent on AI which may weaken their own diagnostic reasoning.

Panel response:

When we think about over-reliance on AI, we consider whether the machine is doing too much of the learner’s thinking. AI-enabled education needs to be deliberately designed to support active integration and problem-solving which is particularly important in rare diseases where the clinician’s ability to recognise subtle, atypical constellations of findings remains critical and datasets for training AI tools may be sparse. It remains essential in medicine to use assessments that show how well learners independently integrate, process, and apply knowledge when approaching clinical problems.

Examples from other disciplines demonstrate how this can be done. At one institution, Boston University, an extraordinarily successful strategy has been the development of the “Socratic” AI system. This system does not simply give the answer to a problem submitted by a learner, (e.g. a student enrolled on a math course), but instead guides the learner through the steps required to solve the problem. Students gain access to a non-judgemental tutor, available on demand, while undertaking the cognitive work themselves. A similar approach would work well for medical students where they could be walked through differential diagnoses or a treatment plan and aligns well with the educational mission of resources such as HPIM. 

Digital Twins and Potential Impact on Patient Outcomes

Questions about digital twins focused on concrete clinical benefits and future directions.

Panel response:

There are several good examples of using digital twins emerging, including:

  • Predicting responses to cancer therapies, especially in patients with advanced or complex malignancies
  • Optimizing surgical planning 
  • Personalizing chronic disease management
  • Improving hospital operations. 

By integrating multimodal patient data, these virtual representations move clinical practice closer to personalized medicine. In drug development, such models may help identify and prioritise therapeutic targets using a more systematic, model driven approach than traditional semi empirical methods. For clinicians, they offer decision support. The key questions are how transparent and trustworthy these virtual models will be, and how closely their predictions match real world patient trajectories.

Institutional Safeguards for Trustworthy AI in Practice

Participants also asked what safeguards institutions should prioritize to ensure AI improves accuracy and efficiency without undermining patient care, trust, or team dynamics.

Panel response: 

Several areas of discussion emerged:

  • Human-curated oversight of AI outputs, including systematic monitoring of outcomes is essential to make sure that systems do not behave in an uncontrolled or unacceptable way.
  • More data are not automatically better, as quality matters more than volume. If large proportions of incoming data are inaccurate or poorly curated, model performance will reflect these inaccuracies which will increase over time.
  • Institutions should strive for partially curated systems built on validated, authoritative sources. For example, the intention for HPIM and related McGraw Hill resources to underpin AI tools with a vetted, evolving knowledge base to minimize hallucinations and errors.
  • Privacy protections must be robust and keep pace with broader digital regulations. Experiences in regions with more advanced digital-privacy frameworks, such as the EU, may provide useful models. Ultimately, some degree of international harmonization will be needed because data and models increasingly cross borders.

From the clinician’s side, how we communicate AI-derived information to patients is critical for maintaining trust. AI can generate lists of diagnostic possibilities or treatment options, but clinicians remain responsible for understanding those outputs and contextualizing them for the patient, addressing questions in a language that supports shared decision making. In this sense, AI should be a tool, not a crutch, and clinicians must continue to deepen their knowledge so they can translate AI suggestions into meaningful conversations tailored to each patient.

Conclusion

Artificial or augmented intelligence is already reshaping both clinical care and medical education, but its impact will depend on how thoughtfully it is designed and implemented. AI can process vast, multimodal datasets to enhance diagnosis, drug development, health system operations and personalised learning, but these gains are dependent on high quality data, rigorous curation, and transparent, interpretable models. 

For clinicians and educators, AI should function as a powerful adjunct. It should support decision making, give learners access to a wider range of complex, varied, and realistic clinical cases, and ease the administrative burden, rather than be a substitute for clinical reasoning or human connection. Realising this potential requires clear governance around data ownership and privacy, active management of bias, robust assessment of learner competence, and safeguards to maintain patient trust. 

HPIM is positioning its curated, globally adapted knowledge base at the center of this ecosystem, aiming to anchor emerging tools in authoritative content so that AI strengthens, rather than dilutes, the science and humanism of medicine.

Get your copy of Harrison’s Principles of Internal Medicine, 22E

Footnote

The 22nd edition of HPIM is now available on AccessMedicine, McGraw Hill’s interactive digital platform, with updated self‑assessment questions, related clinical case studies, podcasts, videos, and thousands of additional resources to support medical education and clinical practice.