LLMs in Clinical Trial Screening: Can AI Find the Right Patients?
Explore how LLMs in clinical trial screening could improve patient matching, and why human oversight, data quality, and trust remain essential.
Large language models (LLMs) are changing how clinical trial screening could work, particularly in oncology, by helping interpret unstructured clinical records. But even the best model is unlikely to ever replace clinical judgement.
Why Does Clinical Trial Screening Need a New Approach?
Clinical trial recruitment remains complicated, especially in oncology. Eligibility criteria are becoming more complex as precision medicine, biomarker selection, targeted therapies, and combination treatments expand.
Clinical trial screening finds the right patients from fragmented clinical data.
Around one-quarter of oncology trials fail to reach their planned sample size, while 18% close after enrolling less than half their target. Further, only around three to five per cent of eligible cancer patients participate in clinical trials.
Manual screening makes this harder. Clinical research staff and oncologists may need to examine medical histories, clinic notes, laboratory results, pathology reports, treatment histories, and molecular information before deciding whether a patient could qualify.
Much of this information is not stored in a clean, searchable format.
A patient may have the relevant information in a narrative clinic note rather than a structured EHR field. A biomarker result could sit in a pathology report; previous treatment may be described differently across records, or a toxicity history may only become clear when several notes are reviewed together.
This is where AI large language models (LLMs) could help.
Traditional automated screening works well when eligibility information is structured. LLMs can potentially interpret the narrative information around those structured fields.
That’s important because clinical eligibility is not an easy database query.
LLMs Could Turn Clinical Text into Trial-Matching Intelligence
Their potential role of LLM-based approaches using models including GPT-3.5, GPT-4, GPT-4.0, o1, and Llama is to process clinical text and trial eligibility criteria, then generate structured outputs that can support patient pre-screening.
The important point is not that an LLM can “understand” a medical record in the same way as a clinician. Rather, an LLM can process large volumes of heterogeneous information quickly and consistently enough to support the screening workflow.
That could include extracting data on:
- Disease characteristics and stage
- Biomarker or genomic information
- Previous treatments
- Laboratory results
- Comorbidities
- Toxicity history
- Treatment exposure
- Relevant eligibility exclusions
The challenge is that different trials demand different types of reasoning.
A precision-medicine study may depend heavily on molecular variants and previous targeted therapy. A dose-escalation study may require close attention to organ function, laboratory trends, safety history, and protocol-defined criteria.
Randomised controlled trials can introduce further complexity through treatment history, washout periods, stratification factors, and arm-specific eligibility.
One generic LLM workflow therefore cannot be expected to perform equally well across every trial design.
As a result, hybrid approaches may be more usable. These combine LLM capabilities with rules-based logic. The LLM can help interpret the messy clinical information, while deterministic rules can enforce protocol-specific requirements.
That hybrid architecture is more realistic than treating the LLM as an autonomous trial-matching decision-maker.
How Should Pharma Companies Audit AI Vendors?
“This is one of the riskiest industries there is because you're dealing with human lives, you're dealing with experimental protocols, and you're dealing with regulatory bodies where you might not get another shot at that clinical trial."
Types of LLMs in Clinical Trial Screening: Zero-Shot, RAG, and Fine-Tuned LLMs Have Different Roles
There are different ways to deploy LLMs.
A zero-shot LLM operates without model-specific training on trial-matching data. This offers a relatively simple route into experimentation. However, there is a significant limitation, especially for oncology where eligibility criteria can become sufficiently specific that performance falls as the complexity of the matching problem increases.
Zero-shot models may therefore be more useful for high-volume preliminary screening than final eligibility decisions.
Retrieval-augmented generation (RAG) provides another route.
Instead of asking an LLM to rely only on what it learned during training, RAG systems retrieve relevant information from external knowledge sources and provide that information to the model during the task.
For clinical trial matching, this could mean giving the model current trial criteria, structured biomarker information, or other relevant data before it generates a screening output.
There are also few-shot approaches that take a different route. They provide carefully selected examples within the prompt to guide the model towards the desired screening behaviour.
Any of the different approaches can improve performance, but they introduce infrastructure requirements.
Good LLM screening depends on good retrieval, good prompts, current trial information, and reliable underlying patient data.
Fine-tuned models are designed specifically for oncology trial criteria. An example would be o1 LLM and Synergy AI OS that use curated oncology clinical documents to improve domain-specific matching.
Fine-tuning can help a model handle specialised terminology, staging concepts, biomarker definitions, and other oncology-specific information. But sponsors must first consider how much model customisation is justified for a screening workflow that must remain auditable, current, and scalable across trials?
The Biggest LLM for Screening Risk Is Not Speed. It’s Trust.
The attraction of LLM-based screening is obvious. So is the risk.
An LLM can produce an apparently plausible interpretation that is wrong.
A small error can have a large operational consequence. Misreading a biomarker threshold could exclude an eligible patient or misinterpreting previous treatment could produce a false match, while missing a safety criterion could send an inappropriate candidate into the next stage of screening.
There are several LLM limitations, including fragmented clinical data, complex eligibility criteria, computational requirements, explainability, privacy, and algorithmic bias.
The black-box problem is particularly important. For a conventional database query, the logic can usually be inspected. A rules engine can show which condition was met or failed.
An LLM-generated recommendation is harder to interrogate.
Luckily, there are several ways of addressing this. Retrieval-grounded systems can require the model to connect outputs to specific eligibility criteria; rules-based or neurosymbolic layers can encode explicit clinical thresholds, and structured rationale outputs can show why a patient was included or excluded.
These approaches change the role of the LLM. A safe workflow considers which eligibility criteria can be supported by the available evidence, which cannot, and what requires clinical review?”
Pharmatica has previously examined the same issue in EHR patient recruitment, where data quality, interoperability, workflow integration, and AI validation remain major barriers.
Technology cannot compensate for poor source data.
If clinical information is incomplete, inaccessible, inconsistently coded, or buried across multiple systems, a more sophisticated model does not automatically solve the underlying problem.
AI in Pharma: Why the Future of Healthcare Starts With Patients, Not Tech
Kate O’Reilly, President & Chair of the Healthcare Businesswomen’s Association (HBA) Dublin-Ireland Chapter & Healthcare Transformation Partner at Roche, discusses AI in pharma, patient engagement, and the future of healthcare innovation.
Human-AI Screening Is a More Credible Clinical Model
Ultimately, a hybrid model, in which automated screening performs the high-volume work and clinical professionals provide oversight and validation, is probably safest.
The goal of LLMs in clinical trial screening is not to automate clinical responsibility. It should be to reduce the amount of repetitive information processing that prevents clinical teams from applying their expertise where it matters most.
A well-designed workflow could therefore look like this:
- Clinical and trial data are collected from relevant systems.
- An LLM extracts potentially relevant information from unstructured records.
- Retrieval systems provide current trial criteria and supporting information.
- Rules-based logic applies deterministic eligibility requirements.
- The system generates a structured screening recommendation.
- A clinician or research professional reviews the evidence.
- The final screening decision remains attributable to an appropriately qualified human.
This model also fits with the broader movement towards interoperable clinical data.
ClinicalTrials.gov now supports study data in HL7 FHIR format, with the stated potential to support applications that identify relevant trials and patients matching eligibility criteria.
The U.S. National Cancer Institute is also working on standardised cancer clinical trial matching data elements to improve exchange between EHR systems and trial-matching services.
The LLM may be the intelligence layer, but interoperability is the infrastructure underneath it.
That matters for those considering investment, as clinical data, trial information, eligibility logic, governance, and human review has to be built into a workflow that can produce reliable screening decisions.
LLMs in Clinical Trial Screening: The Way Forward
LLMs in clinical trial screening have moved clinical trial matching beyond simple keyword search, but they have not removed the need for clinical judgement.
LLMs are particularly attractive because they can work with the unstructured information that conventional screening systems often struggle to process.
But the technology creates a new dependency on data quality, retrieval architecture, validation, explainability, interoperability, and governance.
Sponsors should therefore assess LLM screening as an end-to-end clinical development capability, not an isolated AI project.
That means asking whether the organisation can demonstrate:
- Where the patient information came from
- Which trial criteria were assessed
- What evidence supported each recommendation
- Which information was missing
- When a human reviewer intervened
- How the model was evaluated
- How performance changes across populations and trial types
The equity question also matters. If AI screening works best in health systems with highly structured records, patients receiving care in fragmented or resource-constrained environments could benefit less from the technology.
That would risk reinforcing the same recruitment disparities that better screening is supposed to address.
U.S. Food and Drug Administration (FDA) guidance now explicitly focuses on increasing participation among representative clinical trial populations, including demographic and non-demographic characteristics.
Better matching should not narrow access.
The Opportunity for LLMs in Clinical Trial Screening
LLMs in clinical trial screening do not replace screening teams.
It does give those teams a faster way to interrogate complex clinical information, surface potential candidates, identify missing evidence, and focus expert attention where it has the greatest value.
That is a more measured proposition than autonomous trial matching. It also creates a clearer path towards responsible deployment.
At Pharmatica, we analyse the technologies and operating models reshaping Clinical Development, from AI-enabled recruitment and EHR data to adaptive trial design and decentralised trial strategy. Our Insights focus on where technology moves beyond technical promise and where it creates measurable impact across pharmaceutical development.
Pharmatica: Insight. Connection. Impact.
Frequently Asked Questions
What are LLMs used for in clinical trial screening?
LLMs can analyse unstructured clinical information, extract relevant patient characteristics, interpret trial eligibility criteria, and generate structured recommendations to support clinical trial pre-screening.
How can LLMs improve clinical trial recruitment?
LLMs can process information that may be difficult to identify using conventional structured searches, including narrative clinical notes, treatment histories, pathology information, and other unstructured records. This could help research teams identify potential candidates more efficiently.
What is RAG in clinical trial matching?
Retrieval-augmented generation (RAG) combines an LLM with external information sources. In clinical trial matching, RAG can provide the model with relevant and current trial criteria or patient information rather than relying only on the model's existing training.
Can LLMs replace clinical professionals in trial screening?
Hybrid approaches can assist with clinical trial screening. They combine automated screening with clinician oversight. LLMs can assist with information processing and candidate identification, but complex eligibility decisions still require appropriate human review.
What are the risks of using LLMs for clinical trial matching?
Key risks include inaccurate interpretation of eligibility criteria, incomplete or fragmented clinical data, bias, privacy concerns, limited explainability, and difficulty maintaining reliable performance across different trial designs and patient populations.
About the author
Did you enjoy the content?
Why not support Nicole Dale by giving this content a like
Comments (0)