Artificial Intelligence and Real-World Evidence in Oncology Research: A Q&A with Eric Wu

Researchers and practitioners hoping to develop deeper insights about diseases and treatment outcomes have more methods and tools at their disposal to extract information from data than ever before. Artificial intelligence (AI) can provide innovative methods to process and analyze vast datasets, model complex health economics phenomena, and enhance decision-making processes.

Many Analysis Group consultants, including Managing Principal Eric Wu, have explored ways to leverage AI in health economics and outcomes research (HEOR) projects. Dr. Wu and Vice President Rajeev Ayyagari were among those who presented on this topic at ISPOR 2024. Dr. Ayyagari spoke to Dr. Wu about applying AI to the analysis of real-world data (RWD) and the generation of real-world evidence (RWE), both of which can advance our understanding of diseases and improve health care outcomes.

What is the value of RWD?

Eric Q. Wu: Managing Principal, Analysis Group

The interaction of patients with health care systems worldwide produces tremendous amounts of rich and complex RWD. The benefits of RWD include large sample sizes, multi-dimensionality, representativeness and inclusivity of patient populations, timeliness and rapidity, and longitudinal perspectives.

However, obstacles such as heterogeneity, the lack of interoperability, missing or erroneous data, and varying contextual conditions under which the data were collected make it extremely difficult to harvest research-grade RWD that provide the full picture of patient characteristics, treatment patterns, and clinical and economic outcomes.

What are the key challenges presented by RWD?

To translate large volumes of multi-source data into clinically meaningful, high-quality datasets, three key challenges need to be addressed.

The first is the temporality of the data – that is, when they were collected. The state of medical consensus at the time of collection must factor into the translation of point-in-time information, such as lab values, to form medical judgements – for example, diagnoses or prognoses.

The second is the spatiality of the data – that is, where they were generated. The state of medical practice, research methodologies, and data privacy regulations can vary from region to region and country to country. All these differences may influence the content and interpretation of the data. 

Third, the quality of the data may be affected by their lack of completeness and accuracy. There may be missing data elements, for example, related to disease staging, progression, or lines of therapy, and there is also the risk of miscoding or data errors. For example, clinician notes may document only the initial treatment choice but not the entire patient journey throughout subsequent treatment choices and outcomes. Or it may be unclear how diagnoses of exclusion or suspected diagnoses that are mentioned differ from the final diagnosis.

Many studies to date have not fully addressed these issues with RWE generation and interpretation, but it is critical to do so to ensure data relevance in the fast-evolving health care field. This is especially important for complex disease areas, such as oncology or rare diseases, that involve multiple diagnostic criteria and small or isolated clinical trials.

What methods and tools have been developed to address these issues and overcome RWD challenges?

One major innovation has been the development of dynamic disease models (DDMs). These AI-powered models harness the power of RWD by enhancing the integration of data from various sources, establishing consistency and relevancy to any specific research question, and maximizing the usability of historical information. They use automated data processors to process information from multiple sources or sites while resolving data conflicts and unifying standards and definitions. DDMs provide a fact-based lens to review data based on traceable records, and the algorithms that they employ are continuously updated by and implemented in scientific research.

A second innovation that could be helpful in overcoming the challenges posed by RWD is the use of generative AI (GenAI) platforms that are powered by large-language models (LLMs). An example of this technology is Analysis Group’s own proprietary GenAI platform – AGHealth.ai™ – which extracts insights from an array of data sources and excels in text classification, research summarization, and automated data analysis. By learning from the data that they process, GenAI platforms can help improve the algorithms that power DDMs to deliver research-relevant data. Both of these innovations can streamline complex analyses, enhance research accuracy, facilitate evidence-based decision making, and deepen overall understanding of complex diseases and treatments.

Rajeev Ayyagari: Vice President, Analysis Group

How do you anticipate that DDMs may shape the future of RWE?

The benefits of DDMs go beyond their ability to generate fit-for-purpose RWE capable of addressing various local and current research questions. They are also designed to incorporate adequate information for addressing future questions and those from other regions or settings. Such efforts can help to eventually bridge fragmented and inconsistent data to create a truly coherent and dynamically evolving system with broad, rigorous applications.

Thus, the thoughtful design of RWD synthesis and the perpetual inclusion of information to interpret clinical consensus via AI algorithms can be valuable for advancing research and informing medical decision making.

How could DDMs affect health inequality and disparities in health care?

DDMs can be used to improve RWE about underrepresented or underserved populations. RWD sources can capture more diverse populations than clinical trials that have strict criteria for enrollment and that often underrepresent elderly, minority, or low-income populations. Therefore, DDMs can contribute to better understanding of patient characteristics, treatment patterns, and clinical outcomes in underrepresented or underserved populations. ■