Practical and ethical challenges

This webinar (which took place on 12th February 2025) features experts in data quality and related ethical challenges. The talks focus on addressing common challenges associated with the use of AI in clinical research including data bias, missing data, and ethical concerns.

Talks include:

• Data Bias in Healthcare AI: Strategies for detection and mitigation, Mr Cato Pauling, PhD Student (Healthcare AI), UCL Great Ormond Street Institute of Child Health.
• The ethics of data quality in healthcare AI: What is it, and why does it matter?, Dr Alex McKeown, Head of Data Ethics. Information Governance Services Limited.
• Dealing with missingness when analysing electronic health records: Patient-level observational analyses, Professor Evan Kontopantelis, Professor in Data Science and Health Services, Research Faculty of Biology, Medicine and Health The University of Manchester.
• AI in health care: What kind of consent is appropriate and what should patients know about data and decision quality?, Professor Søren Holm, Professor of Bioethics Organisation, The Centre for Social Ethics and Policy, University of Manchester.

Key takeaways from this video:

1. Every dataset will have some level of bias, there are many different types of bias and possible causes.
2. We can detect bias through a detailed understanding of the data, using statistical analysis, model performance and fairness metrics, and developing explainable artificial intelligence.
3. We can mitigate data bias through using appropriate techniques to curate large and diverse datasets, ensuring quality ground truth labelling (for supervised machine learning), pre-processing the data, and addressing class imbalance.
4. Bias drift means that the bias of a dataset will change over time, requiring constant monitoring and adjustments.

 

Key takeaways from this video:

1. What data quality means in healthcare AI and why it matters: ensuring accurate, complete, and unbiased data for ensuring AI models are reliable and trustworthy
2. Understanding ethical considerations: providing a brief overview of some central challenges of ethical relevance in this context, addressing fairness, trust, and accountability in AI-driven musculoskeletal healthcare
3. Presenting challenges and risks: explaining the importance of identifying biases, gaps, and potential harm caused by poor data quality
4. Promoting best practices in data and AI governance: outlining strategies for implementing ethical frameworks and reasoning to enhance AI transparency and patient trust

 

This presentation focuses on the challenges and methodologies for handling missing data in patient-level observational analyses using Electronic Health Records (EHRs). It outlines key data resources, such as UK Primary Care Databases, and discusses considerations like longitudinal data nature, bias, and issues with recording variables like smoking and ethnicity. The presentation delves into missing data mechanisms (MCAR, MAR, MNAR) and emphasizes multiple imputation (MICE) as a robust method for addressing missing data, despite its computational complexity. It also explores advanced techniques like longitudinal multiple imputation, Bayesian modelling, and machine learning, while highlighting the need for sensitivity analyses and future improvements in handling missing data under MNAR assumptions.

 

Uses of AI in Orthopaedics: The presentation outlines various applications of AI in orthopaedics, including image analysis, diagnostic advice, therapeutic advice, triage/waiting list prioritization, summaries of patient notes, and drafting of discharge letters.

Risks of AI Use: It discusses several risks associated with AI, such as issues with data sources, transparency, explainability, AI bias, AI error, and automation bias/overreliance. These risks can be influenced by the context in which AI is implemented in clinical workflows.

Patient Expectations: The presentation emphasizes that patients expect AI to be as accurate as a good doctor and to serve as an advisory tool rather than making definitive decisions. It also highlights the importance of explainability, testing for bias, and considering multiple opinions.

Consent and Information: It stresses the importance of informing patients about the use of AI in their care and ensuring that they understand what they are consenting to. The presentation also notes that consent requirements may change over time as patients become more informed about AI’s role in healthcare.

Contestation: The presentation discusses the need for patients to be able to contest AI outputs, the basis for those outputs, and the clinical decisions made using AI advice. It provides a framework for evaluating the transparency and accountability of AI-based decisions in orthopaedics.