In an era where artificial intelligence is reshaping countless facets of healthcare, a groundbreaking study published in the journal Machine Learning: Health unveils a promising development in the realm of clinical trial recruitment. Researchers from UT Southwestern Medical Center have demonstrated that ChatGPT, a leading large language model (LLM), can significantly accelerate the patient screening process for clinical trials. This innovation addresses one of the most persistent challenges in medical research—identifying eligible patients swiftly and accurately to enhance trial efficiency and success rates.
Clinical trials remain the cornerstone of medical advancement, responsible for validating new drugs, therapies, and medical procedures before they become widely accessible. However, a chronic bottleneck in these trials is patient enrollment. Many trials fail to recruit enough participants in a timely manner, with some estimates indicating that up to 20% of National Cancer Institute (NCI)-affiliated trials do not reach necessary enrollment thresholds. This shortfall compromises the reliability of trial outcomes, inflates costs, and delays the introduction of potentially life-saving treatments to the broader population.
Traditional patient screening for trials is painstakingly manual, requiring healthcare staff to meticulously comb through extensive electronic health records (EHRs). These records often contain crucial information buried in unstructured formats such as physicians’ notes, making it challenging for legacy systems to parse relevant data efficiently. The process can take an average of 40 minutes per patient, a demanding pace for already burdened clinical teams. This inefficiency inevitably leads to missed opportunities, where eligible patients remain unidentified, hindering trial enrollment and progression.
Recognizing this challenge, the team led by Dr. Mike Dohopolski, a radiation oncologist and AI researcher at UT Southwestern, turned to the capabilities of cutting-edge language models like GPT-3.5 and GPT-4. These models excel at understanding and processing natural language, even in complex and unstructured datasets such as EHRs. The researchers sought to determine whether these large language models could parse patient records and evaluate their eligibility for clinical trials much faster than traditional methods without compromising accuracy.
Their study involved analyzing data from 74 patients being considered for a clinical trial focused on head and neck cancer. By leveraging three distinct prompting techniques—Structured Output (SO), Chain of Thought (CoT), and Self-Discover (SD)—they explored how best to elicit accurate and interpretable eligibility assessments from the AI. Structured Output required the AI to conform responses to a predetermined format, Chain of Thought prompted the model to explain its reasoning process, and Self-Discover allowed the model to independently determine the pertinent information for decision-making.
The findings were striking. GPT-4 outperformed its predecessor GPT-3.5 in accuracy, albeit at the cost of marginally increased computational expense and longer processing times. Screening per patient ranged broadly from approximately 1.4 minutes to over 12 minutes, substantially faster than the conventional 40-minute manual review. The associated costs per screening varied from $0.02 to $0.27, making this an economically viable tool for healthcare providers seeking to scale up trial recruitment efforts.
Despite the encouraging results, Dr. Dohopolski cautions that large language models are not infallible. “While GPT-4 and similar models provide significant time savings and support more flexible criteria screening, their accuracy diminishes when strict, inflexible eligibility rules must be applied,” he explained. The AI should be viewed as a powerful adjunct to human reviewers rather than a complete replacement, offering scalability and consistency that can enhance clinical workflow and decision-making.
Beyond accelerating individual patient screenings, this integration of AI has broader implications for the entire clinical trial ecosystem. Faster, more robust candidate identification can reduce the time and cost of trials, improve the diversity and representativeness of participant pools, and ultimately hasten the introduction of novel therapies to market. This is particularly crucial in oncology and other fields where rapid access to cutting-edge treatments can dramatically impact patient outcomes.
The success of ChatGPT in this context also underscores the transformative potential of natural language processing technologies wherever complex medical data exists in unstructured formats. The inherent flexibility of LLMs enables a nuanced understanding of clinical narratives, medical jargon, and subtle data cues beyond the reach of conventional algorithms. This opens avenues for AI-assisted workflows in diagnostics, treatment planning, and personalized medicine, all of which depend on extracting meaning from vast and intricate data sources.
This study arrives as the inaugural article in IOP Publishing’s Machine Learning series™, the world’s first open access journal series dedicated to machine learning and AI applications in the sciences. The series aims to bridge the gap between rapidly evolving AI technologies and their practical deployment in scientific research and clinical practice, fostering innovation that is both impactful and transparent.
Parallel advances by the same research team include the development of GeoDL, a deep learning system that enables real-time adjustment of radiation therapy while patients remain on the treatment table. This AI can process CT scans and treatment data within milliseconds to provide precise 3D dose estimates, exemplifying how deep learning can enhance the accuracy and efficiency of adaptive radiotherapy in clinical settings. Together, these projects reflect a paradigm shift in oncology, where AI is not only expediting research but directly improving patient care in real time.
Looking forward, this pioneering use of ChatGPT for clinical trial screening is poised to inspire new research, prompting integration with electronic health record platforms and the creation of hybrid AI-human workflows. Further studies will be essential to validate and refine these models across diverse patient populations and trial types, ensuring that the benefits observed in pioneering work at UT Southwestern can be generalized broadly within medicine.
As AI continues to evolve, its role in healthcare is set to expand, pushing the boundaries of what’s possible in diagnostics, treatment, and research. The demonstrated ability of large language models to handle complex, unstructured medical data rapidly and accurately signals a future where clinicians are empowered by intelligent tools that enhance their expertise and workflows, ultimately transforming patient outcomes on a global scale.
Subject of Research: Not applicable
Article Title: ChatGPT Augmented Clinical Trial Screening
News Publication Date: 31-Jul-2025
Web References:
- https://iopscience.iop.org/article/10.1088/3049-477X/adbd47
- https://iopscience.iop.org/journal/3049-477X
- https://onlinelibrary.wiley.com/doi/10.1002/cam4.5276
References:
Dohopolski, M. et al. “ChatGPT Augmented Clinical Trial Screening,” Machine Learning: Health, IOP Publishing, 2025.
Image Credits: IOP Publishing
Keywords: Artificial intelligence, Health and medicine
