The Role of Artificial Intelligence in Nephrology Clinical Trials

Introduction

Clinical trials are the gold standard for testing interventions, however they are limited by multiple challenges. Artificial intelligence (AI) has the potential to transform trials. We provide an overview of the application AI to clinical trial conduct.

Brief Overview of AI

AI is the simulation of human processes by algorithms. Machine learning (ML) is a subset using statistical models to infer patterns. Deep learning is a ML subset using artificial neural networks. Natural language processing (NLP) is an application focusing on understanding human language. The introduction of the transformer model, which learns relationships across trillions of words taking the context into account, has transformed this field. Transformers power large language models (LLMs) such as OpenAI's ChatGPT. Below we provide an overview of how AI can facilitate all clinical trial stages (Table 1).

Table 1.

Potential uses of artificial intelligence in clinical trials with associated considerations

Trial Stage	Use Case	Considerations
Design	Drafting protocol documents	• Need to determine time for protocol drafting and edits compared with current standard • Improvements in critical thinking, logic, and transparency • Unclear whether the inclusion of editing the document will save time
	Drafting informed consent documents	• Prior studies tested against unedited consent documents, which is unlikely to occur in real practice • May not improve readability • Lack of availability in other languages • Unclear whether the inclusion of the time needed to edit the document will save time
Pretrial	Synthetic patient data for testing	• Need to ensure availability of multimodal patient data for model development • Need for a diverse and large source population for digital twin development • Ensuring privacy protections for patient information
Recruitment	Cohort identification	• More diverse training data to improve model generalizability • Concerns when users use unclear cohort definitions • Not able to extract initial event cohorts
	Obtaining informed consent	• Current models need testing in more diverse populations • Lack of availability in other languages • Unclear if time to supervise chatbot will save time • Dehumanization of clinical research
	Patient retention	• Digital inequities, which may limit access to certain groups
Outcomes and analysis	Ascertain outcomes	• May have limited use because of expertise required and computing cost
	Digital end points	• Regulatory approval for use of digital end points • Low physician utilization of digital tools

Study Planning and Drafting

Trials require extensive documentation, and protocol drafting is a bottleneck. Markey et al. tested LLMs for drafting 140 clinical trial protocol sections and found that they performed well on clinical terminology and content relevance but poorly on clinical thinking and logic and were unable to assess transparency and references.¹ The addition of retrieval-augmented generation (RAG) improved performance. A RAG model is composed of both a retrieval and generation component. It allows users to specifically reference a resource (e.g., Kidney Disease Improving Global Outcomes [KDIGO] guidelines) when the LLM generates the response. By adding an external knowledge base, RAG models produce accurate and contextually relevant results. Therefore, a RAG-augmented LLM may be useful for first drafts of trial protocols. However, human input is critical to refine protocols for accuracy.

Clinical Trial Emulation

AI can leverage electronic health records (EHRs) to create digital twins, which simulate patient trajectories and model clinical trial outcomes, demonstrated with TWIN-GPT.² Using this technology, researchers can conduct virtual clinical trials with digital twins and predict trial outcomes and adverse events. This may facilitate needed drug development targeting kidney disease.

Recruitment

Patient recruitment and identification is challenging. While EHRs contain information, identifying patients requires technical expertise to build queries, which is costly and time consuming. Common data models, such as Observational Medical Outcomes Partnership, lead to better data harmonization; however, users must still be able to craft cohort definition queries, which requires an understanding of programming and knowledge of clinical data representation. Researchers have developed NLP tools, such as Criteria2Query, which converts user-entered free-text clinical criteria for cohort definition into Observational Medical Outcomes Partnership cohort queries.³ These tools can allow researchers to identify patients at scale and can be particularly helpful in identifying patients for trials that are highly selective (e.g., many comorbidity exclusions with specific proteinuria and eGFR values).

Recruitment of patients at higher risk of the outcome can lead to efficient trials, and accurate risk prediction models can facilitate appropriate recruitment. ML models can handle missing data points, varying follow-up intervals, and can model nonlinear and complex relationships. There are studies demonstrating the superior performance of ML models over more traditional models for progression of kidney disease, and they can be used as trial enrichment tools.⁴

Informed Consent

Obtaining appropriate and informed consent is critical, but drafting consent documents at the appropriate reading level is difficult. Decker et al. compared an LLM with surgeon-generated consent documents.⁵ While there was no difference in readability between the two methods, the LLM resulted in better descriptions of the benefits and alternatives to surgery. It has yet to be determined whether LLM-drafted consents can improve understanding in nephrology clinical trials.

Another potential use is to have an LLM-based chatbot to obtain informed consent from potential participants. This has benefits including reduced time and effort and freedom for potential participants to provide informed consent at their convenience. This has been studied in genomic research, and it resulted in shorter consent time, a similar number of affirmative consents, good knowledge transfer, and no negative experiences.⁶ These chatbots could be used to explain study protocols, especially adverse events in higher-risk trials (e.g., bleed risk with a kidney biopsy). One concern with this approach is the dehumanization of clinical research. Research staff use the informed consent time to build rapport with patients, which leads to more patient engagement and less loss to follow-up.

Patient Retention

Minimizing loss to follow-up is a critical part of clinical trial management. AI-powered assistants can automate routine monitoring tasks, such as notifying patients for follow-up appointments. In addition, the use of wearable devices may allow for home monitoring, decreasing participant burden. AI models can also be built to predict patients who may drop out of the study and allow for researchers to intervene earlier on these patients.

Outcome Ascertainment

Many trials may require imaging or pathology review for outcome ascertainment (i.e., change in total kidney volume), which is time consuming, subjective, and burdensome. AI models have been developed for assessing radiology images and pathology reports, which can streamline outcome ascertainment. Several models are able to segment renal biopsy samples.⁷ While transformers were originally developed for NLP tasks, they have been adapted for imaging processing with success. They have been used with accuracy for various tasks on computed tomography scans.⁸

In addition, digital end points, often collected through wearables and mobile health apps, provide continuous, real-time data, which AI efficiently manages and analyzes.⁹ Moreover, AI automates data collection and analysis of these end points, reducing manual errors. AI accelerates the integration of digital end points in clinical trials by optimizing end point selection, enhancing data management efficiency, and ensuring reliable trial outcomes. AI's ability to process vast datasets enables the identification of the most predictive end points. Although, digital end points could transform trials, there are implementation barriers.

Finally, AI can be used for real-time data analysis, which can be used to monitor ongoing results of trials. This will allow immediate identification of patients who meet outcomes, such as changes in albuminuria or eGFR slope. AI can also be used to automate processes such as data transfer, data cleaning, and analysis from EHR or clinical databases into research databases.

Nephrology Utilization of AI in Clinical Trials

While AI can potentially facilitate many aspects of clinical trials, to our knowledge, there are no nephrology clinical trials that have used AI in their development or conduct. Currently, the field is focused on validating various models, such as those for outcome ascertainment (e.g., AKI or progression of kidney disease), and testing the performance of LLMs. Unfortunately, our search is limited to published literature and protocols. We are unable to assess AI integration in companies that do not publish protocols or have proprietary workflows.

Limitations

While AI can streamline various steps of the clinical trial process, there are limitations (Table 1). AI systems are often black box in nature in that how the model arrived at the decision is not transparent. The lack of explainability may cause lack of trust in the results by physicians and patients. Performance of AI models depends on the amount and quality of underlying data. Many clinical models are trained on EHR data, which is messy and biased, resulting in poor performing and biased models. In addition, models trained at single institutions may not generalize to patients in other locations because of differences in patient demographics and practice patterns. In addition, models built using biased data will be biased. LLMs are trained from a large corpus of data, and these data may also contain misinformation, which the model will perpetuate and may lead potential participants to make decisions based on inaccurate information. LLMs are also prone to making hallucinations, a response that is factually incorrect. Unfortunately, many LLMs are still cloud based and therefore not compliant with standards to protect patient privacy and security. Lastly, it is the responsibility of the principal investigator to oversee all aspects of the research. This will entail review of AI-generated protocols, consents, and review logs of chatbot consents by researchers. Therefore, it is unclear whether implementation of AI into the various steps of the clinical trial process will be time saving but should be tested rigorously.

Conclusion

Clinical trials require substantial time and effort, and AI has the potential to significantly improve the clinical trial process, particularly in nephrology. However, careful consideration of AI's limitations is crucial for successful implementation.

Disclosures

Disclosure forms, as provided by each author, are available with the online version of the article at http://links.lww.com/JSN/E814.

Funding

G.N. Nadkarni: National Institute of Diabetes and Digestive and Kidney Diseases (R01DK133539 and U01DK133093) and National Heart, Lung, and Blood Institute (R01HL168897 and R01HL167050). L. Chan: Division of Diabetes, Endocrinology, and Metabolic Diseases (K23DK124645 and U01DK13725) and National Institute of Diabetes and Digestive and Kidney Diseases (U01DK137259).

Author Contributions

Investigation: Lili Chan, Girish N. Nadkarni.

Resources: Girish N. Nadkarni.

Supervision: Girish N. Nadkarni.

Writing – original draft: Lili Chan, Girish N. Nadkarni.

Writing – review & editing: Lili Chan, Girish N. Nadkarni.