Using ethical AI to streamline HEOR

Artificial Intelligence (AI) can assist human decision making in many aspects of clinical development, including the evaluation of evidence to inform health technology assessments (HTA) and market access decisions. As AI transforms clinical development, it also creates new complexities, including the risk of bias and the need for transparency. Generative AI, including large language models (LLMs), can drive efficiencies in health economics and outcomes research (HEOR). Those of us who use the technology, however, must be committed to its responsible use. 

At the ISPOR Europe conference in November ‘23, we moderated and participated in an educational symposium on the AI revolution in HTA. Along with industry colleagues Dalia Dawoud from the National Institute for Health and Care Excellence (NICE), Microsoft’s Denise Meade, and Raquel Aguiar-Ibáñez, a Senior Director in oncology-related outcomes research, we discussed the ethics, benefits, and acceptance of AI in this space, as well as earning the acceptance of all stakeholders – including patients. 

Defining the role of AI and its ethical use

Machine learning (ML), a subset of AI algorithms, is a powerful tool for analyzing vast datasets far beyond human capability. But interpreting and implementing AI-informed insights will require the nuanced understanding and creative thinking of a human reviewer to navigate HEOR’s complex ethical and emotional landscapes. AI is a complement to human intelligence, not a replacement.

As we deploy AI for decision-making support, every developer and user must ensure that this technology ethically serves us. Denise offered the following principles to guide the design and deployment of algorithms that power AI solutions.

• Fairness. AI systems should treat similar people in similar ways. Fairness also means ensuring that among different demographic groups, all groups are treated equitably.
• Privacy. This is particularly important when using sensitive, health-related data.
• Safety and reliability. AI systems should behave in expected ways, even under unexpected conditions.
• Accessibility. AI systems should be designed using inclusive design practices that address potential barriers for people with disabilities or who have limited technology knowledge.
• Transparency. Users should understand how AI systems function so they can identify potential performance issues, safety and privacy concerns, biases, exclusionary practices, or unintended outcomes. Users also have a responsibility to be transparent about when, why, and how they deploy AI.
• Accountability. The people who design and deploy AI systems should be accountable for how their systems operate. 

In pursuing ethical AI use, we must also recognize and address areas of potential bias within AI systems. These can include:

• Bias in programming. The behavior of an AI system is determined in large part by its algorithms, including choice of models and parameters. Bias can be inadvertently introduced into algorithms by the developers’ assumptions, perspectives, and decisions. It is essential, then, that programming teams are diverse and inclusive of professionals from different demographic groups, disciplines, and paths of experience. This helps ensure a holistic approach to AI development. Teams should also use development best practices designed to minimize bias.
• Bias in training data sets. If the data used to train AI algorithms doesn’t represent the diversity of the real world, the system will likely develop skewed understandings and predictions. To combat this, developers must ensure the diversity of training sets and regularly audit training data for bias.
• Bias in goals for AI algorithms. Development teams should define clear, ethical, and inclusive goals for every AI tool, thinking through the potential impacts of any work performed. These goals should be reviewed and refined in response to feedback and evolving societal values.

By acknowledging and actively working to mitigate these biases, AI technologies can be developed to serve the diverse needs of all stakeholders.

Streamlining HEOR with AI 

To best understand a therapy’s impact, we need evergreen evidence — continuously integrated new data that reflects real-world healthcare outcomes over time. By creating comparators for mature data, such evidence gives us the power to make dynamic decisions.

One key component of evergreen evidence: systematic literature reviews (SLRs), through which researchers comprehensively identify evidence related to clinical, economic, and humanistic outcomes. SLRs are considered the gold standard by HTA agencies and decision makers.

Aided by AI, researchers can now conduct real-time SLRs in which data can be updated as soon as a citation is published. Such speed is nearly impossible when human reviewers work unassisted, but AI enables a living SLR process that makes critical data quickly available to the experts who need it. AI assistance also reduces human errors (such as erroneous data capture and transfer), standardizes output, and enhances visualizations and data representation.

Of course, the possible biases discussed earlier can impact AI systems used for SLRs. To use AI in a fully ethical way, researchers need to understand the algorithm used by AI for decision making so they can evaluate its quality and the validity of its output. AI users also need to be confident that the algorithm has been trained on high-quality, unbiased data sets.

Stakeholder acceptance of AI-identified evidence

In Raquel’s experience, AI solutions offer strong potential for supporting large SLR data visualization and reporting and for performing ongoing search updates within six months of HTA submission. Users and HTA professionals, however, are less confident about deploying AI in its current evolution for complex tasks such data extraction or quality assurance and interpretation.

While NICE is developing guidance on how sponsors should use AI or report its use in submissions, Dalia said the organization expects sponsors to maintain human oversight over any AI applications and to address ethical considerations in an ongoing way. When considering AI-identified data, she and her colleagues at NICE consider:

• The rigor, transparency, explainability, and reproducibility of data. Sponsors should be prepared to address how AI systems arrive at their recommendations. 
• Possible bias or discrimination. This can include issues with data representativeness, health discrimination within real-world evidence datasets, or human choices in the design, development, and deployment of AI models.
• Possible hallucinations or confabulation in which AI systems can generate inaccurate outputs based on the model’s limited or incomplete knowledge, rather than real-world data.

It’s reasonable to assume that other HTA bodies and market-access decision makers will share similar concerns and expectations. 

In addition to HTA stakeholders, drug developers and their partners should consider patient opinions on AI use. According to recent research, approximately 70 percent of adults in the U.S. are concerned about the increased use of AI in health care, with respondents most frequently expressing fear that systems will make diagnostic errors or recommend incorrect treatments. 

We cannot earn widespread support for AI without addressing stakeholders’ legitimate interests. Our first step must be an industry-wide commitment to standards for ethical AI use, a foundation for which could be the principles of fairness, privacy, safety, reliability, accessibility, transparency, and accountability. We must demonstrate to stakeholders that we will use AI in an ethical way and that we will be forthcoming about its deployment, allowing interested parties to examine and challenge algorithms as needed.

We must also be clear that AI will never supplant or surpass human decision making—a fact that may grow AI’s acceptance among patients. Rather, we need to reinforce AI’s rightful role: a precise, powerful tool for streamlining HEOR for the good of researchers, the wider research community, and the patients we all serve.

Return to Insights Center