AI-Webinars-1-5-Dated

WMA Webinar #1 – Introduction to AI in
Medicine – Key Points
Table of Contents
Introduction​ 2
Foundations of AI in Healthcare​ 2
A. Background on Digital Health 2
B. Key Terminology & Definitions 3
C. Historical Evolution of AI in Healthcare 3
D. Key Points on Safety and Trust in Healthcare AI 3
Generative AI and the Future of Medical Intelligence​ 5
A. Overview of Generative AI in Healthcare 5
B. Large Language Models (LLMs) in Healthcare 5
Will AI Replace Doctors? The Evolving Role of Physicians​ 6
A. WMA’s Statement on Augmented Intelligence 6
B. AI Use Cases 6
C. The Future of Doctor-Patient Relationships 6
Questions and Answers​ 7
Closing Remarks & Next Webinar Announcement​ 8
1
30 January 2025
Introduction
●​ Overview of the role of AI in modern healthcare.
●​ Key concerns regarding AI adoption:
○​ Appropriate use cases for AI in medicine.
○​ Safety, ethical implications, and regulatory considerations.
○​ The capabilities and limitations of generative AI and large language models.
○​ The potential for AI to complement or replace physician roles.
●​ Introduction of the five-part AI in Medicine webinar series.
Foundations of AI in Healthcare
Artificial Intelligence (AI) is transforming healthcare by improving diagnostic accuracy,
streamlining workflows, and enabling personalized treatment. As AI continues to integrate into
healthcare systems, understanding its foundational principles is essential. This section explores
the background, key terminology, historical advancements, and considerations related to safety
and trust in AI-driven healthcare.
A. Background on Digital Health
●​ Adoption Trends:
○​ 93% of physicians believe digital health tools are helpful and provide some
advantage.
○​ The average physician uses almost four digital tools, double the number from a
few years ago.
○​ The COVID-19 pandemic accelerated the adoption of digital health tools,
particularly telemedicine.
●​ Impact on Healthcare Delivery:
○​ AI tools are being used for remote patient monitoring and chronic disease
management.
○​ Personalized treatment plans are being enhanced through AI-driven data
analysis.
○​ AI is streamlining administrative tasks, such as prior authorization, scheduling,
and supply chain management.
●​ Data-Driven Decision-Making:
○​ AI is enabling predictive analytics to support early disease identification and risk
assessment.
○​ Real-time data access is improving clinical decision-making at the point of care.
2
B. Key Terminology & Definitions
●​ Digital Health: The use of technology to enhance healthcare services, efficiency, and
patient outcomes.
●​ AI: The application of advanced algorithms and machine learning models to replicate
human cognitive functions.
●​ Digital Health Components:
○​ EHRs: Electronic health records that centralize and streamline patient
information.
○​ Telemedicine: Virtual delivery of medical care through video or audio
consultations.
○​ Telehealth: Broader digital health services, including remote monitoring and
patient education.
○​ mHealth: The integration of mobile devices and wearable technology into
healthcare.
●​ AI-Specific Terms:
○​ Augmented Intelligence: AI designed to support, rather than replace, human
expertise in healthcare.
○​ Machine Learning: AI systems that continuously learn from vast datasets to
improve decision-making.
○​ Deep Learning: Advanced neural networks capable of complex pattern
recognition in medical imaging and diagnostics.
○​ Natural Language Processing (NLP): AI-driven interpretation of human language
in medical records and documentation.
○​ Foundational Models: AI models trained on large datasets, adaptable for
healthcare applications such as diagnostics and treatment planning.
C. Historical Evolution of AI in Healthcare
●​ Early Foundations (1950s-1960s): Introduction of AI concepts in computing and their
initial application in medicine.
●​ Expert Systems (1970s-1980s): Development of rule-based AI models such as MYCIN
for bacterial infection treatment.
●​ Machine Learning Advances (1990s-2000s): Early applications of data-driven AI in
radiology, pathology, and predictive modeling.
●​ Deep Learning and Big Data (2010s-Present): Integration of large datasets, deep
learning models, and real-time AI applications in clinical practice.
D. Key Points on Safety and Trust in Healthcare AI
●​ Physician Oversight: AI should assist, not replace, physicians, maintaining human
expertise as a critical component of patient care.
3
●​ Patient Safety as a Fundamental Priority:
○​ AI tools must be rigorously validated to ensure they do not harm patients.
○​ Physicians must maintain oversight of AI tools to ensure they are used safely and
effectively.
○​ The concept of “human in the loop” is critical, meaning that AI should assist, not
replace, physician judgment.
●​ Transparency and Explainability:
○​ AI systems must be transparent, meaning that physicians and patients should
understand that a decision is impacted by AI.
○​ The output of AI tools should be explainable, allowing clinicians to trace back
the reasoning behind recommendations.
○​ The “black box” problem (where AI decisions are not understandable) must be
avoided to maintain trust in AI systems.
●​ Regulatory and Ethical Considerations:
○​ Regulatory frameworks for AI in healthcare are still evolving, and there is a need
for global standards to ensure safety and efficacy.
○​ Ethical concerns, such as bias in AI algorithms and the potential for misuse of
patient data, must be addressed.
○​ Physicians should be involved in the design, development, and validation of AI
tools to ensure they align with clinical needs and ethical standards.
●​ Data Privacy and Security:
○​ Patient data used to train AI models must be protected, and patients should
have the opportunity to opt out of data sharing.
○​ Compliance with data protection regulations is essential to maintain patient
trust.
○​ There is a need for clear consent processes to ensure patients understand how
their data is being used.
●​ Physician Oversight and Accountability:
○​ Physicians remain ultimately responsible for patient care, even when using AI
tools.
○​ AI should augment, not replace, physician judgment, especially in complex or
uncertain cases.
○​ Liability frameworks must be clarified to determine responsibility when AI tools
are involved in decision-making.
●​ Global Collaboration and Standards:
○​ There is a need for international collaboration to develop consistent standards
for AI in healthcare.
○​ The World Medical Association (WMA) and other organizations play a key role in
advocating for ethical and safe AI practices globally.
4
Generative AI and the Future of Medical Intelligence
Generative AI distinguishes itself from traditional AI applications by its ability to create new
data rather than merely analyzing existing datasets. Unlike predictive analytics and standard
machine learning models that enhance decision-making, generative AI produces clinical
reports, synthesizes patient histories, and even formulates research hypotheses, making it
highly valuable in medical documentation, research synthesis, and patient communication.
A. Overview of Generative AI in Healthcare
●​ AI models that generate novel medical content, including diagnostic reports, treatment
recommendations, and research insights.
●​ Enhances medical research by synthesizing vast amounts of data into actionable
insights.
●​ Supports AI-driven clinical decision-making and automates documentation workflows.
B. Large Language Models (LLMs) in Healthcare
●​ AI systems trained on extensive medical literature to provide contextually accurate
recommendations.
●​ Automates administrative documentation, improving physician efficiency.
●​ Challenges include ensuring model accuracy, addressing biases, and maintaining
patient privacy.
5
Will AI Replace Doctors? The Evolving Role of Physicians
Despite concerns that AI will replace physicians, its primary function is to augment medical
practice by increasing efficiency and improving decision-making. While AI excels in automation
and data analysis, essential aspects of healthcare—such as patient interaction, ethical
decision-making, and complex case management—remain irreplaceable.
A. WMA’s Statement on Augmented Intelligence
●​ AI is designed to enhance medical expertise, not replace human judgment.
●​ AI systems should be developed in alignment with ethical, regulatory, and clinical best
practices.
●​ Continuous professional education is essential for integrating AI into medical practice
effectively.
B. AI Use Cases
●​ Administrative automation: AI assists with scheduling, billing, and documentation.
●​ Medical imaging diagnostics: AI enhances the accuracy of radiology, pathology, and
dermatology assessments.
●​ Predictive analytics: AI improves early disease detection and risk stratification for
preventive care.
C. The Future of Doctor-Patient Relationships
●​ AI enables enhanced personalized treatment, optimizing medical decisions based on
patient data.
●​ Physicians may increasingly act as AI interpreters, guiding patient care using AI-driven
insights.
●​ Ethical considerations must ensure AI integration does not compromise patient trust
and transparency.
6
Questions and Answers
1. Who is responsible for AI-driven medical decisions?
●​ This is still an evolving area with multiple factors influencing accountability.​
Countries are implementing processes to establish clear and consistent guidelines.​
Efforts are underway to define legal responsibility among physicians, AI developers, and
healthcare institutions.​
Regulatory frameworks are evolving to address accountability concerns.
2. What ethical considerations are involved in AI use?
●​ AI must be developed and used transparently to maintain patient trust.
●​ Ethical concerns include bias in training data, patient privacy, and accountability.
●​ Informed consent is crucial when AI influences diagnosis or treatment decisions.
3. How can AI improve patient care without replacing doctors?
●​ AI can analyze vast amounts of data to support evidence-based decision-making.
●​ It enhances workflow efficiency by automating administrative and diagnostic tasks.
●​ The human element—empathy, ethical reasoning, and nuanced
decision-making—remains irreplaceable.
4. What are the biggest challenges in implementing AI in healthcare?
●​ Ensuring AI reliability and avoiding biased or incorrect recommendations.
●​ Integrating AI into existing clinical workflows without disrupting care.
●​ Compliance with regulatory standards such as GDPR and HIPAA to protect patient data.
5. How can physicians stay informed about AI advancements?
●​ Engaging in continuous education programs on AI in medicine.
●​ Collaborating with AI developers from the very beginning to ensure practical, ethical
applications.
7
Closing Remarks & Next Webinar Announcement
●​ Key Takeaways:
○​ AI’s transformative role in healthcare, including its limitations and ethical
considerations.
○​ The importance of physician oversight in AI-driven decision-making.
○​ Generative AI’s role in improving efficiency, research, and personalized
medicine.
●​ Next Webinar:
○​ Topic: Ethics, Legal, and Regulatory Aspects of AI in Healthcare
○​ Date: February 27th
○​ Why Attend? Learn about the evolving regulatory landscape and ethical
frameworks guiding AI adoption in medicine.
We invite all participants to continue the discussion in the upcoming session and play an active
role in shaping the future of AI-driven healthcare.
8
Summary Document:
Ethics, Legal, and Regulatory Aspects of
AI in Healthcare
WMA Educational Webinar – Medical Technologies
Workgroup – 27 February 2025
Introduction
This document provides a summary of key points discussed in the second installment of the
WMA Educational Webinars on Artificial Intelligence in Medicine. The session focused on
ethical, legal, and regulatory aspects of AI in healthcare, exploring patient rights,
accountability, and risk mitigation strategies.
Key Topics Discussed
1. Ethical Considerations in AI-Powered Healthcare
●​ Autonomy & the Doctor-Patient Relationship
○​ AI can either empower patients by enhancing autonomy or reduce it if decisions
are overly dependent on AI-generated recommendations.
○​ Different global perspectives influence ethical views (e.g., Western consumerism
vs. paternalistic models in China).
●​ Informed Consent Challenges
○​ Traditional informed consent involves doctor-patient discussions.
○​ AI’s “black-box” nature makes it difficult for physicians to explain how AI arrived
at a decision, raising concerns about whether true informed consent is possible.
○​ Some legal frameworks (e.g., EU regulations) mandate disclosure when AI is
used in patient care, while others (e.g., US law) do not explicitly require it.
○​ The level of AI intervention significantly impacts informed consent.
○​ It’s essential to consider whether patients should be informed about AI use in
their care, balancing transparency with the risk of overwhelming them with
information.
1
27 February 2025
○​ The “explainability” of AI is crucial when AI significantly influences healthcare
decisions.
●​ Loss of Human Touch in Medicine
○​ AI-driven healthcare may reduce direct physician-patient interaction, leading to
concerns about dehumanization.
○​ There is potential for over-reliance on AI, which can introduce confirmation bias.
2. Legal and Regulatory Challenges for AI in Healthcare
●​ Liability & Accountability
○​ Three primary perspectives on liability:
■​ Physician Liability – Doctors are traditionally held accountable if AI-driven
decisions harm patients.
■​ Institutional Liability – Hospitals may be responsible if they implement AI
tools in ways that contribute to harm.
■​ Developer Liability – AI manufacturers and software developers could be
held accountable, though legal precedents are still evolving.
○​ In the future, AI could potentially influence the standard of care, shifting liability
from doctors to institutions or developers.
●​ Regulating AI Across Its Lifecycle
○​ Three key stages where regulation is needed:
■​ Research & Development: Ensuring ethical AI design, data bias
mitigation, and proper validation.
■​ Market Approval: Regulatory approvals vary across countries (e.g., FDA
in the US, MDR in the EU, SFDA in Saudi Arabia).
■​ Post-Market Oversight: Mechanisms to hold AI accountable after
deployment (e.g., WHO global governance frameworks).
●​ Public Sector vs. Private Sector Regulation
○​ Different countries adopt different approaches:
■​ Free-market-driven approach with minimal AI regulation to encourage
innovation.
■​ Strong regulatory frameworks prioritizing patient protection.
■​ AI governance is largely centralized, with government oversight playing a
key role.
●​ Documentation: Hospitals may need subcommittees to set internal documentation
standards for AI use.
●​ Best practices: Until clear guidelines emerge, physicians must exercise clinical judgment
and understand the standard of care, as relying solely on AI may not be a legal defense.
2
Privacy and Rights in AI-Assisted Healthcare
●​ Data privacy laws: Legal frameworks, such as GDPR, impact AI development and
deployment, but they also have gaps, such as exceptions for research and public
benefit.
●​ Data security: Blockchain technology may offer a path to securing health data, with
systems allowing patients to approve data access requests.
●​ Patient data ownership: Patients should maintain ownership over their data, raising
questions about data removal from AI systems and models. Balancing the desire to
create comprehensive AI systems with the need for patient autonomy and control is a
real tension.
●​ Intellectual property: AI developers may protect their algorithms as trade secrets,
complicating transparency and regulation.
●​ Global variability: There is tremendous variation in what developers do to ensure data is
unbiased and patients have recourse if harm is caused.
3. Risk Mitigation Strategies for AI in Clinical Practice
●​ Best Practices for Physicians Using AI Tools
○​ Physicians should document AI-assisted decision-making in patient records.
○​ Human-in-the-loop (HITL) models should ensure physician oversight in
AI-generated recommendations.
○​ AI tools should augment, not replace clinical judgment.
●​ Patient Data Protection & Security
○​ AI relies heavily on patient data, raising privacy concerns.
○​ Key data risks:
■​ Bias in Training Data – AI can perpetuate biases if trained on
unrepresentative datasets.
■​ Data Privacy – Regulations in different regions attempt to protect patient
rights, but gaps remain.
■​ Security Threats – Blockchain technology is being explored for securing
patient data.
○​ Should patients have the right to remove their data from AI models?
■​ While AI requires large datasets, patients should maintain ownership and
be informed about how their data is used.
●​ Human Oversight & AI Transparency
○​ AI should not function as an independent actor in patient care. Instead:
■​ Explainability: AI systems must provide reasoning for decisions where
possible.
■​ Regulatory Alignment: Healthcare institutions should align AI usage with
existing medical guidelines.
3
■​ Ethical Oversight: AI should be reviewed by hospital ethics committees
before clinical deployment.
Closing Thoughts & Next Steps
●​ AI in healthcare presents unprecedented opportunities and challenges.
●​ Legal and ethical frameworks are still evolving, with no single global standard yet.
●​ Clinicians must remain engaged in discussions about AI governance to ensure safe,
ethical, and effective AI integration.
Next Webinar in the Series
📅March 27, 2025 – “Current and Future Applications of AI in Medicine”
Speaker Introduction: TBD
Topics will include:
●​ Breakthrough AI applications in diagnostics and treatment
●​ AI’s role in precision medicine and drug discovery
●​ Challenges in AI adoption across healthcare settings
Feedback & Further Reading
📩Feedback Form: Please share your thoughts via the link.
📖Recommended Reading:
●​ Research Handbook on Health, AI, and the Law – Available online (Open Access).
●​ Informed Consent & AI in Healthcare – Article by Prof. Glenn Cohen.
For further inquiries, contact WMA at wma@wma.net.
4
FAQs:
Ethical, Legal, and Regulatory Aspects of AI in Healthcare
1. What are the primary ethical considerations when using AI in medicine?
Core ethical principles like autonomy, beneficence, non-maleficence, and justice are crucial.
AI’s impact on the doctor-patient relationship is a key concern. Does AI increase patient
autonomy by giving them more control over their data and care, or does it lead to a more
paternalistic dynamic where doctors overly rely on AI-driven insights? There is also a risk of
ceding too much authority to AI developers who set the standards for AI device usage if
regulations are inadequate. The potential loss of the human touch is also a major consideration
and must be addressed to maintain patients’ trust and the therapeutic alliance with their
doctors.
2. How does AI impact informed consent in healthcare?
The traditional informed consent process involves a doctor explaining treatment options and
the reasoning behind their recommendations to a patient. When AI is involved, especially in
decision-making, the explainability of the AI’s reasoning becomes crucial. If the AI’s
decision-making process is too complex to understand (even for its developers), how can a
doctor adequately explain the basis of the recommendation to a patient, fulfilling the
requirements for informed consent? The level of AI involvement in the medical decision greatly
impacts informed consent. Transparency about the AI’s role and limitations is essential.
3. Should patients be informed about the use of AI in their clinical care, and to what extent?
There are varying opinions. Some argue that informing patients about every instance of AI use
(e.g., in radiology for dose reduction) would be overwhelming and ineffective. Others believe
transparency is necessary, especially when AI significantly influences treatment
recommendations.
4. Who is liable if a patient is harmed by AI in healthcare?
Liability is a complex issue with several potential actors: the doctor, the healthcare institution
(hospital/clinic), the AI developer, and even (controversially) the patient. Currently, existing tort
law dictates that doctors are responsible if their actions fall below the standard of care,
regardless of whether AI was used. Hospitals may also be liable under vicarious liability if
doctors are hospital employees. The liability of AI developers is less clear, but is something that
is beginning to garner more attention. It is likely to increase in prominence in coming years.
The implementer, who connects the system, validates the data, and turns things on, also carries
some shared liability.
5. What steps can healthcare professionals take to mitigate liability risks associated with AI?
5
As of right now, clinicians must take responsibility for medical decisions even when relying on
AI, but in the future this may not always be the case. AI setting the standards of care could shift
liability elsewhere. Clinical judgment is paramount. Careful documentation of AI’s role in
decision-making is crucial. Hospitals should form subcommittees to establish internal
principles, guidelines, and standards for AI implementation and documentation.
6. What aspects of AI in healthcare should be regulated, and why?
Regulation is needed across AI’s entire life cycle, from research and development to
deployment and post-market surveillance.
Areas for regulation include:
• Research and Development: Ensuring developers adhere to best practices, such as engaging
with patient groups and using appropriate data sources to mitigate against biases.
• Market Approval: Establishing clear requirements for regulators to assess AI devices, not just
regarding risk, but also issues like medical liability and informed consent.
• Clinical Practice: Creating mechanisms to ensure patients can exercise their medical law
rights (informed consent, liability) when AI is used in their care.
7. What are the potential risks of AI to the physician-patient relationship?
AI could potentially dehumanize healthcare by leading to over-reliance on technology and a
loss of human touch. Doctors may become overly reliant on AI, exhibiting confirmation bias
and potentially overlooking crucial information. There is concern AI may lead doctors to
become “lazy” and simply rely on AI decisions without question. It is critical to consciously
guard against that.
8. How can patient data be secured and ethically used in AI systems?
Data bias, privacy concerns, and security vulnerabilities must be addressed.
• Data Bias: Careful data selection in the research phase is critical to avoid building AI systems
that are ineffective or discriminatory.
• Data Privacy: Existing data protection laws (e.g., GDPR) provide a framework, but local
authorities should supplement these with healthcare-specific guidelines.
• Data Security: Technologies like blockchain can enhance data security by giving patients
greater control over who accesses their data. The issue of intellectual property rights
surrounding algorithms can complicate data privacy regulations.
6
Notes and Takeaways from WMA Webinar:
CURRENT AND FUTURE APPLICATIONS OF
ARTIFICIAL INTELLIGENCE IN MEDICINE
Overview
This webinar explored the current and emerging applications of Artificial Intelligence (AI) in
medicine, including diagnostics, decision support, patient engagement, drug development,
and data governance. The conversation also addressed regulatory and ethical considerations
surrounding AI adoption in healthcare.
The session provided both a technical and strategic look at how AI is transforming clinical
workflows, supporting physicians, and potentially redefining care delivery through automation,
personalization, and augmentation of medical decision-making.
Current Medical Applications
1. Diagnostic Support and Accuracy
●​ AI systems can now emulate expert reasoning, such as that of a seasoned radiologist.
●​ FDA-approved AI tools are widely available to assist physicians in radiology,
dermatology, and pathology, with proven gains in diagnostic accuracy when combined
with human oversight.
●​ Example: In rheumatology, AI classified capillary images with performance nearly equal
to expert rheumatologists.​
2. Predictive Analytics and Clinical Decision Support
●​ AI models developed in Zurich predict clinical outcomes such as ICU delirium, lung
function decline, and the need for medication adjustments in chronic conditions.
●​ LLMs (Large Language Models) like ChatGPT have demonstrated >90% accuracy on US
medical licensing exam questions, surpassing average student performance.​
3. Conversational Agents in Clinical Workflows
1
27 March 2025
●​ AI-powered assistants can support physicians in creating radiology reports, translating
text, and navigating imaging data, increasing efficiency and reducing screen time.
●​ Example: A radiologist assistant prototype allows interactive queries such as “show me
pathology” or “write a report.”​
4. Operational and Logistical Applications
●​ AI tools help optimize workforce allocation (e.g., predicting nursing demand) and
reduce no-show rates (e.g., MRI appointments), although human behavior remains a
limiting factor.​
5. Education and Support Tools for Physicians
●​ AI is being integrated into medical education curricula to support clinical reasoning,
though only after traditional learning to ensure foundational knowledge is established.​
Future Applications & Trends
1. Personalized and Precision Medicine
●​ AI enables personalized care by identifying patient cohorts with similar genetic or
clinical profiles.
●​ Example: In oncology, AI helps match patients beyond standard treatment guidelines
using data-driven similarity searches.​
2. Drug Development and Discovery
●​ Tools like AlphaFold are revolutionizing protein structure prediction, expediting target
identification in drug development.
●​ AI also assists in designing CRISPR-based gene editors tailored to individual mutations.​
3. Closed-loop Systems and Medication Delivery
●​ AI-controlled drug delivery systems (e.g., for anesthesia or hypertension) have potential,
though regulatory and trust barriers remain high.​
Voice-based diagnostics (e.g., for schizophrenia) and wearable-guided medication
2
management are emerging.​
4. Patient Engagement and Remote Monitoring
●​ AI-powered virtual assistants and chatbots offer triage, education, and behavioral
nudges, though long-term engagement remains challenging.
●​ Embedded passive sensors may offer a solution for compliance without active user
interaction.​
5. Digital Twins and Future Care Models
●​ While digital twins are conceptually valuable, they are currently limited by data
availability and system integration challenges.
●​ The future may include AI-guided care for low-risk pathways, with human oversight
reserved for complex cases.​
Key Quotes or Insights
●​ “We’re trying to emulate expert thinking and put that into computers.”
●​ “AI and physicians both make mistakes. Together, they can reduce them.”
●​ “We expect AI to be unbiased, but we forget that humans are biased too.”
●​ “AI will support, not replace physicians—especially by taking over simpler, repetitive
tasks.”
●​ “Patients are gaining access to expert-level knowledge through AI—this changes
everything.”
●​ “Trust and transparency are critical. Physicians must know when AI is being used and
why.”​
3
Summary of Takeaways
●​ AI Augments, Not Replaces: AI excels at supporting physicians in diagnostics,
documentation, and decision-making but does not eliminate the need for human
judgment.
●​ LLMs Show Promise: Tools like ChatGPT achieve high accuracy on medical exams and
can support clinical reasoning if properly validated.
●​ Workload Relief: Automation of repetitive tasks improves efficiency and may help
address workforce shortages and burnout.
●​ Precision Medicine at Scale: AI facilitates personalized treatment by analyzing large
patient datasets and predicting outcomes.
●​ Operational Value: AI improves hospital efficiency through tools for staffing, scheduling,
and predictive logistics.
●​ Behavioral Engagement Remains a Challenge: AI tools are less effective for long-term
behavior change and must be designed for seamless integration into daily life.
●​ Digital Ethics and Trust: Transparency, bias mitigation, and human oversight are key to
ethical AI deployment.
●​ Cybersecurity and Data Privacy: Strong safeguards, anonymization, and secure
platforms are essential, especially when using external cloud services.
●​ Digital Twins and Closed-loop Systems: These are on the horizon but require significant
advances in data quality and regulatory clarity.
●​ AI in Education: Integrating AI into curricula must be balanced to preserve critical
thinking skills and clinical reasoning.
4
Q&A from the Webinar:
Question:​
Do LLMs have the capacity of quantifying the degree of certainty in the answers they give?​
Answer:​
Yes, certain methods and semantic-embedding techniques enable LLMs to estimate
confidence scores by analyzing response variation or semantic coherence, helping identify
unreliable answers. These techniques are being refined to improve trust and interpretability in
medical contexts.
Question:​
Will AlphaFold and new models help find cures for genetic diseases like Alpha-1 antitrypsin
deficiency?​
Answer:​
Yes, models like AlphaFold and AlphaMissense are advancing the identification of pathogenic
mutations, while CRISPR-based methods have shown potential in correcting DNA mutations in
alpha-1 antitrypsin deficiency, accelerating the development of curative therapies.
Question:​
Could future models trained on AI notes instead of physician notes eliminate bias?​
Answer:​
Unclear based on available sources. While AI-generated notes might reduce individual human
biases, they risk reproducing and amplifying systemic biases unless rigorously audited and
corrected during training.
Question:​
Will understanding AI’s “black box” improve diagnostic learning?​
Answer:​
Yes, interpretable AI approaches like semantic uncertainty quantification can provide insights
into reasoning patterns, potentially enhancing clinician learning and diagnostic
decision-making frameworks.
Question:​
How can AI integration address LMIC-specific medical challenges?​
Answer:​
AI in LMICs must be adapted for limited resources, using low-compute models, mobile-first
tools, and localized datasets to improve diagnostics and access. Shared benefits include better
triaging and telemedicine support, but infrastructure gaps remain a key barrier.
Question:​
What hidden biases affect AI fairness in medicine?​
Answer:​
Biases can stem from non-representative training data, clinician labeling errors, and failure to
5
include social determinants of health. WMA can help by promoting global standards for AI
fairness, transparent auditing, and equity-centered model development.
Question:​
Do health insurers use AI to override physicians and deny coverage?​
Answer:​
Yes, AI tools are increasingly used by insurers for prior authorization, with reports indicating
denial rates significantly higher than manual reviews and lawsuits alleging AI-led systematic
coverage denials.
6
Best Practices in Medical AI Development
WMA Educational Webinar Series on Artificial Intelligence in Medicine April 30, 2025
1. From Concept to Clinic: Understanding the AI
Development Journey
●​ Start with real clinical problems, not technology hype. Clinical problems are similar
worldwide, though workflows and administrative processes differ between institutions.
Use design thinking: interview and observe stakeholders to thoroughly understand the
real problem before developing solutions.​
●​ Validate the problem before seeking technological solutions. Example: What appeared
to be a hospitalist-physician communication issue was discovered to be specifically
about discharge processes. Creating follow-up appointments from the hospital to
outpatient services led to 93% patient attendance and 12% reduction in readmissions.​
●​ Consider the appropriate technology approach. Early AI applications focused on
imaging as data was already digital. Before generative AI (pre-2023), machine learning
required thousands of data points to prove a single algorithm. Generative AI has
transformed healthcare technology implementation, requiring less data and offering
more flexibility.​
●​ Build minimum viable products by adapting existing technologies when possible.
Example: Defense industry image analysis technology designed to detect changes in
maps was successfully adapted for mammogram analysis to identify year-over-year
changes, flagging differences for physician review rather than attempting specific
diagnoses.​
●​ Test thoroughly before implementation. Always obtain IRB approval and conduct
comparative studies with and without the technology. Successful integration depends
on embedding tools directly into clinical workflow, such as AI algorithms for brain
hemorrhage detection that prioritize abnormal CT scans with color-coding.​
●​ Focus on user adoption and workflow integration. Technology should integrate
smoothly into existing clinical processes. Success is evident when clinicians actively
request the technology, as happened when radiologists wanted the AI tool installed on
their personal computers after seeing its value.​
30 April 2025
2. The Clinician’s Role in a Health Tech Team
●​ Physicians are essential for meaningful healthcare technology development.
Technology specialists often don’t understand the complexity of medical care, ethical
considerations, and patient relationships. Example: A CEO without healthcare
background assumed a UTI consultation could be reduced to under one minute, not
understanding that even simple cases involve complex patient contexts.​
●​ Clinicians serve multiple critical roles in technology development: domain experts for
clinical functionality, gatekeepers for patient safety, evaluators of technology fit, and
advocates for ethical implementation. Their insight is crucial from concept through
deployment for successful integration.​
●​ Create effective multidisciplinary teams by bringing together medical specialists, data
scientists, engineers, and IT experts. Develop structured evaluation systems to assess
technologies from multiple perspectives (clinical value, integration feasibility, workflow
impact).​
●​ Bridge the gap between startup speed and healthcare institution caution by identifying
and modifying bureaucratic barriers. Example: Reducing contract length from 60+
pages to 3 pages facilitated startup relationships and reduced legal costs. Provide
project managers to help companies navigate hospital systems and regulatory
processes.​
●​ Clinicians must be assertive about their value in technology development. Companies
that fail to incorporate physician perspectives often struggle to create successful
healthcare solutions. Physician involvement is essential for understanding clinical
thinking and workflow integration.​
3. Ensuring Safety, Ethics, and Trust
●​ Patient safety is non-negotiable. Always follow proper regulatory processes (IRB, FDA)
for any technology involving patient care. Run comparative studies to demonstrate
effectiveness and safety. Example: AI tools for detecting lung nodules in ER chest X-rays
reduced missed diagnoses and related lawsuits by automatically flagging concerning
images.​
●​ Maintain the human-in-the-loop approach where AI makes suggestions but clinicians
maintain control and oversight. Physicians must remain responsible for final decisions in
all patient care situations. Key evaluation question: “Would I use this AI tool for a family
member? If not, it’s not ready for implementation.”​
●​ Ensure transparency and explainability in AI systems. Understand what data was used to
train algorithms and be alert to “black box” solutions that can’t be explained.
Technology must be continuously monitored for algorithmic drift (when algorithms
change behavior over time) and bias.​
●​ Address accountability considerations. AI can both reduce liability (by catching issues
humans might miss) and create new concerns. Technology must be continuously
monitored, and institutions should establish clear lines of responsibility for AI-assisted
decisions.​
●​ Design for equity and ethical implementation by vigilantly monitoring for biases in data
and algorithms, ensuring technology works for diverse populations, and implementing
proper data handling and privacy protections.​
Future Directions
●​ Medical education must evolve to include technology evaluation skills for future
healthcare professionals. The goal is not to teach every technology (impossible given
rapid changes) but to develop critical thinking about technology implementation.​
●​ AI can enhance patient-centered care by reducing administrative documentation
burden, transcribing and summarizing information, enabling pre-visit symptom
checking, and supporting personalized medicine approaches through better data
analysis.​
Conclusion
AI in medicine is not something being done to clinicians but must be built with their active
involvement. Healthcare professional expertise, caution, and vision are critical for ensuring
these technologies enhance patient care rather than compromise it. The successful integration
of AI into healthcare requires ongoing clinician involvement, careful validation, and a
commitment to maintaining human judgment and oversight throughout the process.
WMA Webinar #5 – AI for Health Equity:
Bridging the Global Divide – Key Notes and
Takeaways
Overview
This session focused on AI for global health and equity, exploring how technology can bridge
the gap between high and low-resource settings. The discussion, led by Professor Annie
Hartley, highlighted the use of open-source, participatory frameworks to make clinical AI
inclusive, adaptable, and globally relevant.
Current Challenges & Opportunities
●​ The AI Opportunity in Low-Resource Settings: While often seen as difficult
environments for technology, low-resource settings are arguably the best places to
introduce AI because there are often no alternatives for care. The goal is to provide
access where it currently does not exist.
●​ Data Inequity: Global data is inequitably available; for example, less than 3% of
PubMed content is pertinent to Africa. This leads to models that are inaccurate for
underrepresented populations.
●​ Imperfect Models: We cannot wait for perfect data digitization; instead, we must accept
imperfect models and adapt them through continuous use and validation.
Core Solutions: Meditron & MOOVE
●​ Meditron (The Model): An evolving suite of open-source medical Large Language
Models (LLMs) adapted specifically for low-resource settings.
○​ These models are designed to be efficient, with smaller versions capable of
running on local hardware (e.g., phones) without internet or heavy infrastructure.
○​ Despite their smaller size, they perform competitively on medical licensing
exams.
●​ The MOOVE Platform (The Method): Stands for Massive Open Online Validation and
Evaluation.
○​ It allows clinicians to “nudge” or tune global models to their local context by
rating answers and providing feedback.
○​ This process promotes local ownership and co-development, ensuring the AI
reflects local medical standards and language.
22 May 2025
Implementation & Ethical Considerations
●​ Vigilance over Replacement: The narrative of AI replacing doctors is irrelevant in many
global health contexts where millions never see a doctor at all. Training should focus on
“vigilance”—teaching physicians to know when the tool works and when it fails.
●​ Privacy and Sovereignty: Local ownership is key to privacy. By using open-source
models hosted locally, institutions can maintain data sovereignty and do not need to
send patient data to external commercial servers.
●​ Validation Approach: Rather than waiting for new global regulations, AI deployment
should follow existing Good Clinical Practice (GCP) standards, treating deployment like
a clinical trial to monitor efficacy and safety.
Summary of Takeaways
●​ Bridging the Divide: AI can provide essential medical knowledge to regions with severe
physician shortages, serving as a critical access tool rather than just an efficiency tool.
●​ Local Adaptation: Tools like MOOVE allow clinicians to adapt global models to local
contexts without expensive retraining, addressing data bias.
●​ Open Source Security: Open-source models allow for local hosting, which is superior for
patient privacy and prevents dependence on changing commercial APIs.
●​ Physician Responsibility: Liability remains with the physician, similar to using a medical
textbook; the human must interpret the information responsibly.
●​ Actionable Step: Clinicians are encouraged to participate in validating and tuning
models for their specific regions to ensure fair representation.