Advancing methodological development of artificial intelligence in patient-centered comparative clinical effectiveness research: Patient-Centered Outcomes Research Institute’s unique contribution to research done differently

Abstract

Background

Recent advancements of Artificial Intelligence (AI) are rapidly transforming clinical research. While this technology offers exciting opportunities, it amplifies existing concerns regarding the need for transparent methodology that fosters patient engagement, and introduces new challenges. PCORI’s Improving Methods portfolio has invested in methodological research to enhance rigor and transparency via patient-centered approaches in AI.

Objective

This commentary outlines PCORI’s approach to funding and promoting a portfolio of methodological research that aims to improve the conduct of patient-centered comparative clinical effectiveness research (CER), with a focus on AI methods. The paper highlights a growing portfolio of over 40 AI related projects, including a recent cohort leveraging large language models to augment research processes in CER.

Discussion

PCORI’s current portfolio of methods projects in AI illustrate timely opportunities for the clinical research informatics community to develop and assess AI applications that will further advance a robust, interoperable and ethical infrastructure for patient-centered CER. PCORI’s requirement for ongoing, meaningful engagement of patients throughout the research lifecycle provides a blueprint for patient-centered AI by developing and applying models and methods designed to create value for patients and other healthcare partners.

Introduction

Recent advancements of artificial intelligence (AI), particularly the rise of generative large language models (LLMs), are rapidly transforming key aspects of health and healthcare, including clinical research informatics (CRI). These generative models have demonstrated unprecedented capabilities to process, understand, and generate human-like text.^1,2 AI applications directly intersect with the key domains in CRI, including clinical research, data standards, and information science that support the discovery and management of new knowledge relating to health and disease.³ Many of these information-intensive activities leverage real-world data (RWD) from electronic health records (EHRs), claims, clinical registries and patient-reported information. They span clinical trials, secondary use of clinical data, as well as translational research, all of which rely on an increasingly sophisticated informatics infrastructure. Advances in generative AI have set the stage for applications that can integrate the ever-growing volumes of RWD, generating novel insights for biomedicine and clinical research, including patient-centered comparative clinical effectiveness research (CER). Such approaches have the potential to facilitate the development of an integrated learning health system (LHS) that develops and incorporates new knowledge in clinical care to support informed health and healthcare decision-making by patients and clinicians.

While the technology offers exciting opportunities for CRI and LHS, it amplifies existing concerns regarding the need for transparent methodology and patient engagement, and introduces new challenges related to addressing public distrust of AI.⁴ As a funder of patient-centered CER in the United States, the Patient Centered Outcomes Research Institute (PCORI) is uniquely positioned to support methods development in CER as 1 strategy to bridge the trust gap between AI technologies and the patients and stakeholders they serve. This commentary introduces PCORI’s holistic approach to funding and promoting methods research to advance the use of AI methodologies that can improve the conduct of patient-centered CER, and outlines potential opportunities for collaboration between the CRI and CER communities to further this aim.

The need for transparent methods to support patient-centered AI

The imperative for methods development that incorporates patient-centered approaches in AI stems from both long-standing and emerging concerns related to fairness, transparency and reliability. Involving patients and stakeholders in a co-design process of model development, optimization, and interpretation can build trust and help alleviate these concerns.

First, the rise of large-scale generative models has exacerbated existing concerns in AI, particularly fairness and transparency. It is well understood that algorithms trained on biased datasets can lead to incorrect research conclusions that may disproportionately affect certain populations.⁵ Generative models, which are primarily trained on datasets scraped from the internet, may contain subtle, hard-to-detect societal biases. As these models scale, ensuring that the data they use are high-quality, fair, and representative becomes increasingly difficult. Additionally, many current AI algorithms function as “black boxes” with complex decision-making processes that are not transparent or easily interpretable. This issue of transparency/interpretability is even more pronounced with some of the powerful generative models, where technical details, model weights, and datasets are proprietary and undisclosed.⁶

Second, generative AI models raise new challenges related to reliability and the risk of unintended outcomes. Despite their ability to create novel and customized content, these models are prone to hallucinations⁷ and can produce factually incorrect yet plausible outputs that may mislead users. Scaling these models can enhance performance, but the highly complex architectures introduce unpredictability and can lead to behaviors that even developers did not anticipate.^1,8 A further challenge lies in integrating AI methods into clinical research workflows with fidelity. While these models offer powerful foundational capabilities, adapting them to specific tasks and datasets without losing robustness is a significant hurdle, often resulting in significant variability in accuracy across different settings.^7,9,10

Although the number of studies comparing different AI methods in clinical research is growing, comprehensive evaluations that employ a patient-centered approach remain limited and are needed to fully assess their relative utility.^11,12 This gap in evidence underscores the need to augment existing CRI priorities with an emphasis on designing and assessing patient-centered uses of AI applications that strengthen the existing clinical research infrastructure for patient-centered CER.

PCORI is well positioned to support efforts to advance the patient-centeredness of AI methods used in CRI, with a particular focus on patient-centered CER. Per PCORI’s mission, “PCORI helps people make informed healthcare decisions, and improves healthcare delivery and outcomes, by producing and promoting high-integrity, evidence-based information that comes from research guided by patients, caregivers, and the broader healthcare community.” The commitment to patient-centeredness is reflected in PCORI's Foundational Expectations for Partnership in Research,¹³ engaging patients and community members throughout the research process, from design to execution and dissemination.

PCORI requires a patient-centered approach in all research it funds. These requirements, evaluated through PCORI’s letters of intent, applications and programmatic oversight of the research, ensure that research questions, conduct, and findings are informed by patient and stakeholder perspectives. By design, this approach can help address key limitations in AI methods, such as transparency/interpretability, fairness, and reliability. Similarly, PCORI’s commitment to transparency and public dissemination of research results is a cornerstone of the mission, and a critical need as AI applications and their use in research and care evolves. Together, these approaches fulfill PCORI’s commitment to Research Done Differently.¹⁴

With PCORI’s requirements in mind, 1 approach to conceptualizing patient-centered AI that PCORI introduced at the 2024 PCORI Annual Meeting¹⁵ is to “engage patients as partners throughout the entire AI life cycle, from problem formulation to design and implementation of AI services that enable patients to make informed healthcare decisions and contribute to improvements in healthcare delivery and health outcomes.

Achieving this aim requires that AI solutions are designed, used, and optimized in a way that:

• Is fair, secure, transparent, and reliable;

• Demonstrates responsiveness to real-world needs of engaged patients and communities; and

• Is aligned with patient values and preferences.”

PCORI’s contributions to improving use of AI methods in CER

As AI in clinical research is advancing, contributions to transparent methodologies that promote the goals of patient-centered AI are critical to promoting high-quality research findings. A proportion of PCORI’s funding portfolio is dedicated to improving the methodology of patient-centered CER, as well as the data and technology infrastructure for patient-centered CER, including notable investments such as the National Patient-Centered Clinical Research Network, or PCORnet.¹⁶ As of the spring of 2025, PCORI has funded more than 40 research projects through its methods research portfolio that improve use of AI and Machine Learning (ML) methods in clinical research, including 6 projects led by informaticians. In addition, in the spring of 2024, PCORI funded supplements to fifteen projects that aim to advance methodological research using AI, with a particular emphasis on LLMs in patient-centered CER. Collectively, this portfolio of more than 40 projects focused on AI and ML will provide useful insights into the range of methodological considerations needed to promote transparency and patient-centeredness. With this framework, the outcomes of PCORI’s methods portfolio have the potential to advance the integration of patient-centered approaches into clinical research using LLMs.

PCORI’s methodology projects that leverage LLMs do so in a variety of ways, from data processing and information extraction to model interpretation (Figure 1). Many of the awards seek to improve the efficiency of the research process or address the initial challenge of data collection and integration. These projects propose new approaches on data processing/information extraction to reduce the burden for research participants, research staff, or investigators by processing unstructured data from a variety of sources including EHRs, clinical notes, patient-provider text messages and claims data to extract useful information.^17–20 Projects include streamlining approaches to develop computable phenotypes²¹ to identify potential research participants; identifying clinical covariates,²² disease activities,²³ and mental health outcomes²⁴; facilitating identification of non-medical drivers of health²⁵; or identifying household or family relationships.²⁶

Figure 1.

Key stages in research process utilizing AI/ML methods in CER, with focus areas addressed by Methodological projects funded by PCORI. This figure outlines 5 key stages in the research process where AI/ML methods are applied: data processing/information extraction, research consent, predictive modelling, model evaluation, and model interpretability.

As 1 example, Wu et al.²⁵ developed and disseminated clinical LLMs such as GatorTron²⁷ and GatorTronGPT²⁸ to address limitations of traditional rule-based machine learning methods, which often lack scalability and transportability across data networks. These models demonstrate measurable improvements by accurately extracting 19 distinct categories of non-medical drivers of health (eg, transportation access) from unstructured clinical narratives, achieving high performance in real-world deployments. These models have enabled broader implementation and validation of the LLMs across different EHR systems, enabling external researchers and institutions to adapt these methods to their own datasets and data workflows. To further strengthen methodological rigor, the research team will continue to address key challenges in AI applications, particularly model transportability and evaluation.²⁹ They aim to enhance the transportability of AI methods for cross-institution applications by developing LLMs with strong transfer learning ability via multitask prompt tuning. Additionally, they will systematically evaluate the pretrained model weights to identify potential biases, and implement a patient-in-the-loop mechanism to engage patients, healthcare providers, and researchers in assessing bias impact and developing mitigation strategies, a critical step toward increasing trust and transparency in AI applications for patient-centered CER.

A subset of funded projects focuses on predictive modelling, adapting LLMs to specific clinical cases to develop models for adverse outcomes in diabetes,³⁰ risk in lung cancer,³¹ and treatment effects in depression.²⁴ Importantly, outputs of the LLMs/predictive models can go through rigorous model evaluation, examining factors such as model bias,^30–33 prediction errors,³⁴ model hallucinations,^23,33 and alignment with human/expert responses.^33,35 For example, Diaz et al. propose to develop mediation methods to assess fairness and mitigate bias in LLMs through deliberate discussions with a panel of diverse stakeholders. These stakeholders collectively determine which model outcomes are perceived as fair or unfair.³²

Funded studies will also contribute to patient-centered LLM methods through rigorous evaluation of the ethical aspects of LLMs. For example, Nebeker et al. propose to evaluate how commercial LLMs collect and manage user data and convey data sharing/privacy changes in the terms of service agreements. The project explores how these terms can be best communicated during the research consent for studies that use digital health, which has been developed in partnership with relevant stakeholders, including patients, Institutional Review Board (IRB) members, and research coordinators via design workshops and surveys.³⁶ Another study by Flory et al. addresses model interpretability to assess clinical equipoise for the purpose of designing natural experiments in real-world settings. Through intensive focus groups of patients and stakeholders, this project seeks to understand how LLMs may be used to define, identify, and report clinical uncertainty.³⁵ This type of continuous improvement process aligns with PCORI’s National Priority for Health “Accelerating Progress toward an Integrated Learning Health System,”³⁷ where best practices are seamlessly embedded and new knowledge is captured to continuously improve practice and health outcomes. Such considerations are important to the design of LHS approaches and have potential to inform the design, deployment, and use of LLMs to better align with patients’ needs and preferences.

The range of AI/ML methods projects PCORI is funding continues to grow and is directly responsive to the findings and recommendations of PCORI’s Emerging Technologies and Therapeutics report—AI in Health Care: Strategies to Improve the Impact of AI on Health Equity.^38,39 By ensuring patient partnership as well as participation throughout the development process of LLM and AI methods more broadly, PCORI aims to facilitate a transition from the traditional focus of AI methods from performance optimization to approaches that focus on the real-world needs and objectives of patients and communities. PCORI's methods funding initiatives aim to address these key factors by empowering patients to shape the development of patient-centered AI methodologies that truly serve patient interests, enabling patients, caregivers, and the broader health and healthcare community to make informed care decisions that improve healthcare delivery and outcomes.

Areas for further explorations

As the AI landscape continues to evolve, important questions related to the development of robust methods to evaluate and foster patient-centeredness of AI will require further exploration by the clinical research community, with several areas of particular relevance to CRI. Potential areas for further exploration of the current portfolio and framework of methods for patient-centered AI include:

• defining and implementing patient-centered metrics that measure LLM optimization;

• evaluating the comparative clinical effectiveness of AI methods for specific clinical use cases, considering factors such as the influence of training data, and the trade-offs of different LLM development approaches in terms of transparency, privacy, and security;

• assessing and improving AI model generalizability across different patient populations and clinical settings;

• enhancing data linkage and integration across disparate sources to create more fair and representative datasets;

• integrating multi-modal data sources (eg, structured EHR data, unstructured clinical notes, medical imaging) including patient-generated data from digital health technologies; and

• advancing the science of meaningful patient engagement throughout the AI development lifecycle, as well as dissemination and implementation of AI in patient-centered CER and evidence-based practice.

PCORI’s Improving Methods for Conducting Patient-Centered CER funding announcement⁴⁰ supports innovative methodological research on these topics and is open to interested investigators over 3 funding cycles per year.⁴¹ The CRI community is well-positioned to develop and assess AI applications and methodologies to advance a robust, interoperable and ethical infrastructure for patient-centered CER.

PCORI sees important opportunities for collaboration between the informatics and CER communities to support research and develop innovative methodologies that reflect patient values and preferences. As the clinical research paradigm shifts to incorporate AI in an integrated LHS, it is essential that CRI and CER communities continue to work together as part of a learning system that generates knowledge and evidence for patients and community members. Collaboration to advance an approach to patient-centered AI will ensure patients and the public have the information they need to make informed decisions that align with their desired health outcomes. PCORI’s requirement for ongoing, meaningful engagement of patients throughout the research lifecycle provides a blueprint for patient-centered AI by developing and applying models and methods designed to create value for patients and other healthcare partners.

Author contributions

Jinghua Ou conceptualized the manuscript and developed the initial manuscript. Erin Holve made substantial contributions to the framework and critically revised intellectual content. All authors read and approved the final manuscript.

Funding

None declared.

Conflicts of interest

None declared.

Data availability

No new data were generated or analyzed in support of this research.