Artikkel
Reshaping patient and treatment insights with RWE and AI
-
Derek Swiger, PharmD, MS
-
Ryan Fiano, Ph.D., MPH
The role of real-world evidence (RWE) to inform healthcare decision-making is expanding. With its potential to support a broader evidence base, RWE offers insights that complement traditional clinical trials by capturing patient experiences, treatment patterns, and outcomes in real-world settings.1,2 As healthcare stakeholders demand more robust, timely, and actionable evidence, artificial intelligence (AI) is helping to transform how RWE is generated, analyzed, and applied. Already, AI is being leveraged to support RWE in several ways and shows huge promise for broader applications in the future. Nevertheless, caution and an understanding of its limitations should be applied.
Current applications of AI in RWE
Enhancing coding and algorithm development
One area where AI is revolutionizing RWE studies is with the coding workflows required for complex algorithms. We are finding that tasks that once took months can now be completed in weeks, thanks to AI’s ability to assist in developing, refining, and validating code.
Ryan Fiano, PhD., a researcher at Cencora, explained that by leveraging AI in a recent project, the time involved in building complex algorithms for treatment patterns dropped from several weeks to a week or less. "The efficiency comes from knowing how to prompt the AI and prior experience with turning outcomes into actionable steps,” he said.
This specific expertise is vital because the pre-work required to define common retrospective outcomes—such as Lines of Therapy (LOT), persistence, or treatment switching—is rarely a standardized process; definitions often vary significantly across studies. By combining deep domain knowledge with focused prompting techniques, analysts can equip AI with validated internal code patterns, effectively "teaching" it to apply standard “coding” approaches across studies. For instance, by "showing" the model exactly how the team handles typical rules, such as medication overlaps and gap contingencies to define LOT, analysts can help to ensure that more complex edge cases require only minimal refactoring.
This process, we have found, dramatically reduces the "debug loop"—transforming what was once a multi-day cycle of syntax troubleshooting into a streamlined validation process where analysts focus on verifying clinical intent rather than correcting code structure.
Ryan Fiano, PhD., a researcher at Cencora, explained that by leveraging AI in a recent project, the time involved in building complex algorithms for treatment patterns dropped from several weeks to a week or less. "The efficiency comes from knowing how to prompt the AI and prior experience with turning outcomes into actionable steps,” he said.
This specific expertise is vital because the pre-work required to define common retrospective outcomes—such as Lines of Therapy (LOT), persistence, or treatment switching—is rarely a standardized process; definitions often vary significantly across studies. By combining deep domain knowledge with focused prompting techniques, analysts can equip AI with validated internal code patterns, effectively "teaching" it to apply standard “coding” approaches across studies. For instance, by "showing" the model exactly how the team handles typical rules, such as medication overlaps and gap contingencies to define LOT, analysts can help to ensure that more complex edge cases require only minimal refactoring.
This process, we have found, dramatically reduces the "debug loop"—transforming what was once a multi-day cycle of syntax troubleshooting into a streamlined validation process where analysts focus on verifying clinical intent rather than correcting code structure.
Improving data validation and quality
Ensuring the validity and quality of real-world data is critical for generating credible evidence. AI plays a key role in this process by identifying gaps, inconsistencies, and errors in datasets.
For example, Dr. Fiano noted that his team was able to leverage prompt templates derived from established quality control practices to systematically generate quality control checkpoints that validate outputs while ensuring consistency.
“This approach transforms manual quality assurance into a scalable, repeatable process integrated seamlessly within AI workflows,” Dr. Fiano noted.
By automating these validation processes, AI allows researchers to focus on higher-value tasks, such as study design and interpretation, while maintaining the rigor required for regulatory and payer submissions.
For example, Dr. Fiano noted that his team was able to leverage prompt templates derived from established quality control practices to systematically generate quality control checkpoints that validate outputs while ensuring consistency.
“This approach transforms manual quality assurance into a scalable, repeatable process integrated seamlessly within AI workflows,” Dr. Fiano noted.
By automating these validation processes, AI allows researchers to focus on higher-value tasks, such as study design and interpretation, while maintaining the rigor required for regulatory and payer submissions.
Challenges and limitations of AI in RWE
Ensuring validity and transparency
While AI offers significant efficiencies, it also introduces challenges related to validity and transparency. AI models, particularly large language models (LLMs), function as black-box systems, making it difficult to understand how certain outputs are generated.3
Derek Swiger, PharmD, MS, an expert in digital innovation at Cencora, highlighted the importance of building safeguards into workflows:
“At the end of the day, Generative AI tools are black boxes, so we need to be creative with building checks into the process that ensure the outputs are valid and aligned with client needs.”
For example, researchers can validate AI-generated data by comparing it against known benchmarks or using self-checking mechanisms to ensure consistency. These safeguards are essential for maintaining trust in AI-driven RWE.
Derek Swiger, PharmD, MS, an expert in digital innovation at Cencora, highlighted the importance of building safeguards into workflows:
“At the end of the day, Generative AI tools are black boxes, so we need to be creative with building checks into the process that ensure the outputs are valid and aligned with client needs.”
For example, researchers can validate AI-generated data by comparing it against known benchmarks or using self-checking mechanisms to ensure consistency. These safeguards are essential for maintaining trust in AI-driven RWE.
From human expertise to AI effectiveness
AI’s effectiveness in RWE depends heavily on the expertise of its users. Crafting accurate prompts and interpreting AI-generated outputs requires a deep understanding of both the data and the research context.
“The value of AI right now lies in combining content expertise with the ability to build effective prompts,” Dr. Fiano explained. “You need to know how to guide the AI to get meaningful results and ensure that those results align with the study objectives.”
This human-in-the-loop approach underscores the importance of retaining critical thinking and domain expertise, even as AI automates many aspects of RWE generation.4,2
While human expertise is integral, so too is accessing high-quality data – a persistent problem given how often data is siloed across institutions or locked behind proprietary systems. Federated learning models, which allow data to remain decentralized while enabling collaborative analysis, offer a potential solution. However, implementing these models requires overcoming technical and regulatory barriers.
“The value of AI right now lies in combining content expertise with the ability to build effective prompts,” Dr. Fiano explained. “You need to know how to guide the AI to get meaningful results and ensure that those results align with the study objectives.”
This human-in-the-loop approach underscores the importance of retaining critical thinking and domain expertise, even as AI automates many aspects of RWE generation.4,2
While human expertise is integral, so too is accessing high-quality data – a persistent problem given how often data is siloed across institutions or locked behind proprietary systems. Federated learning models, which allow data to remain decentralized while enabling collaborative analysis, offer a potential solution. However, implementing these models requires overcoming technical and regulatory barriers.
Future trends and opportunities
Federated learning—and, more broadly, federated analytics—is already reshaping RWE by enabling multi-institutional collaboration without requiring the centralization of patient-level data.
Dr. Fiano shared his perspective on the potential of federated learning. “Federated frameworks enable the analysis of dispersed datasets as if they were coalesced in one place, creating a unified virtual cohort that overcomes single-institution limitations without ever physically centralizing sensitive patient records.”
Rather than pooling identifiable records, federated architectures execute standardized analytic queries locally at each data partner, returning only aggregate results or summary statistics. Europe's DARWIN EU network exemplifies this at scale. As of December 2025, it has initiated over 100 studies across 32 data partners covering approximately 188 million patients,6 all coordinated through a federated approach built on the OMOP Common Data Model (CDM).
Dr. Fiano shared his perspective on the potential of federated learning. “Federated frameworks enable the analysis of dispersed datasets as if they were coalesced in one place, creating a unified virtual cohort that overcomes single-institution limitations without ever physically centralizing sensitive patient records.”
Rather than pooling identifiable records, federated architectures execute standardized analytic queries locally at each data partner, returning only aggregate results or summary statistics. Europe's DARWIN EU network exemplifies this at scale. As of December 2025, it has initiated over 100 studies across 32 data partners covering approximately 188 million patients,6 all coordinated through a federated approach built on the OMOP Common Data Model (CDM).
This model is directly enabled by the AI-driven harmonization via privacy-preserving tokenization which enables patient journey linkages across disparate datasets (claims, EHR, mortality, genomics) without exposing identifiable information. Additionally, semantic schema alignment and automated CDM mapping reduce the traditional multi-month extraction, transform, and load (ETL) burden to weeks.
An earlier initiative was the European Health Data & Evidence Network (EHDEN), which was focused on gathering broad insights and evidence from real-world clinical data. The project has since transitioned to the Netherlands-based EHDEN Foundation, which is focused on building the infrastructure for a federated network across Europe.7
In the United States, the Food and Drug Administration’s (FDA) recent policy shift to accept de-identified RWE for certain regulatory submissions further validates federated evidence as a credible alternative to centralized data warehouses.8
For rare diseases or niche populations where single institutions lack statistical power, federated analytics enables researchers to generate robust comparative effectiveness evidence while maintaining strict data governance—each institution retains custody of its data, and re-identification risk is mitigated through architectural controls that separate tokens from quasi-identifiers.9 For example, federated queries can now link genomic data, claims, and mortality registries across multiple institutions to study treatment patterns and long-term outcomes in ultra-rare diseases where individual sites may have fewer than 50 patients, but the federated network collectively reaches sample sizes sufficient for causal inference and regulatory-grade evidence.10
An earlier initiative was the European Health Data & Evidence Network (EHDEN), which was focused on gathering broad insights and evidence from real-world clinical data. The project has since transitioned to the Netherlands-based EHDEN Foundation, which is focused on building the infrastructure for a federated network across Europe.7
In the United States, the Food and Drug Administration’s (FDA) recent policy shift to accept de-identified RWE for certain regulatory submissions further validates federated evidence as a credible alternative to centralized data warehouses.8
For rare diseases or niche populations where single institutions lack statistical power, federated analytics enables researchers to generate robust comparative effectiveness evidence while maintaining strict data governance—each institution retains custody of its data, and re-identification risk is mitigated through architectural controls that separate tokens from quasi-identifiers.9 For example, federated queries can now link genomic data, claims, and mortality registries across multiple institutions to study treatment patterns and long-term outcomes in ultra-rare diseases where individual sites may have fewer than 50 patients, but the federated network collectively reaches sample sizes sufficient for causal inference and regulatory-grade evidence.10
Conclusion
AI is fundamentally reshaping the landscape of RWE, offering unprecedented efficiencies in data consolidation, coding, validation, and study design. While challenges related to data validity, transparency, and access remain, the opportunities presented by federated learning and AI-enabled study design are immense.
AI will play an increasingly important role in the development of RWE solutions as companies seek to navigate the complexities of evidence generation and achieve their market access goals.
AI will play an increasingly important role in the development of RWE solutions as companies seek to navigate the complexities of evidence generation and achieve their market access goals.
*Kildene fortsetter nedenfor
Ansvarserklæring:
Informasjonen i denne artikkelen utgjør ikke juridisk rådgivning. Cencora, Inc. oppfordrer leserne på det sterkeste til å gjennomgå tilgjengelig informasjon relatert til emnene som diskuteres og stole på sin egen erfaring og ekspertise når de tar beslutninger relatert til dette.
Ta kontakt med teamet vårt
Vårt team av ledende verdieksperter arbeider iherdig med å omsette bevis, policy-innsikt og markedsinformasjon til effektive tilgangsstrategier for det globale markedet. Vi hjelper deg med å finne frem i dagens komplekse helsetjenester. Kontakt oss, så hjelper vi deg med å nå målene dine.
Sources
1. Towards responsible artificial intelligence in healthcare-getting real about real-world data and evidence. J Am Med Inform Assoc., Nov 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12626219/
2. Opportunities and Challenges for AI-Based Analysis of RWD in Pharmaceutical R&D: A Practical Perspective, Künstliche Intelligenz, Oct 2023. https://link.springer.com/article/10.1007/s13218-023-00809-6#:~:text=analysis%20approach,illustrate%20challenges%20and%20methodological%20considerations
3. What is black box AI? IBM, Oct 2024. https://www.ibm.com/think/topics/black-box-ai
4. Towards responsible artificial intelligence in healthcare—getting real about real-world data and evidence, J Am Med Inform Assoc., Nov 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12626219/
5. Darwin EU. https://www.darwin-eu.org/
6. Big data highlights, EMA, Dec 2025. https://ec.europa.eu/newsroom/ema/newsletter-archives/70356
7. European Health Data and Evidence Network (EHDEN): Shaping the future of health data in Europe, European Union. https://data.europa.eu/en/news-events/news/european-health-data-and-evidence-network-ehden-shaping-future-health-data-europe
8. FDA Eliminates Major Barrier to Using Real-World Evidence in Drug and Device Application Reviews, FDA, Dec 2025. https://www.fda.gov/news-events/press-announcements/fda-eliminates-major-barrier-using-real-world-evidence-drug-and-device-application-reviews
9. Tokenization techniques for privacy-preserving healthcare data: tokenization nuts and bolts., Front Drug Saf Regul., Dec 2025. https://pubmed.ncbi.nlm.nih.gov/41488932/
10. Privacy-by-Design with Federated Learning will drive future Rare Disease Research, Journal of Neuromuscular Diseases, Dec 2024. https://journals.sagepub.com/doi/10.1177/22143602241296276
