TLDR: A systematic review from 2020-2025 highlights the growing role of large language models (LLMs) in automating clinical trial matching. The review found that LLMs offer a promising solution to the labor-intensive and error-prone manual matching process, improving efficiency and potentially reducing costs. GPT-4 consistently demonstrated strong performance, often outperforming other models. Key challenges include access to real-world data, deployment costs, and mitigating risks like hallucinations and bias. The paper emphasizes the need for standardized metrics, realistic test sets, and collaboration with clinical experts for successful and safe implementation.
Matching patients to the right clinical trials is a crucial step in developing new treatments, especially in areas like oncology where options can be limited. Traditionally, this process is manual, time-consuming, expensive, and prone to errors, often leading to delays in recruitment and issues with patient diversity in trials. Imagine the effort involved in sifting through extensive patient records and complex eligibility criteria for countless trials – it’s a significant bottleneck in medical research.
However, a new era of automated matching systems is emerging, with large language models (LLMs) at the forefront. A recent systematic review, published by Braxton A. Morrison, Madhumita Sushil, and Jacob S. Young, explores the advancements in using LLMs for clinical trial matching. This comprehensive review, which you can read more about in the full paper here, analyzed 31 studies published between 2020 and 2025, highlighting the capabilities, limitations, and best practices in this rapidly evolving field.
How LLMs are Changing the Game
LLMs offer a flexible and scalable alternative to older, rule-based automation systems. They can interpret both structured and unstructured data, extract critical eligibility criteria, and match patients to trials with greater efficiency. The review categorized LLM applications into four main types:
- Patient-criterion matching: Assessing a patient against individual trial criteria.
- Patient-to-trial matching: Identifying a list of suitable trials for a given patient.
- Trial-to-patient matching: Finding eligible patients for a specific trial.
- Binary classification: Determining if a patient-trial pair is a match (yes/no).
Most LLM-based systems follow a similar pipeline: acquiring patient and trial data, pre-processing this data, retrieving relevant information, and finally, matching. Data pre-processing is particularly vital, involving steps like ‘enrichment’ to make raw text machine-readable. This includes extracting key medical concepts (e.g., “Stage IV non-small cell lung cancer”), normalizing terms (e.g., “high blood pressure” to “hypertension”), and detecting negations (e.g., “no history of diabetes”). These steps ensure accurate interpretation and prevent missed matches.
Retrieval-augmented generation (RAG) is a common technique used to extract the most relevant parts of patient records, reducing the amount of data an LLM needs to process. This not only keeps the data within the LLM’s context window but also significantly cuts down computational costs. After pre-processing, data is fed into either prompt-based systems (where LLMs generate responses that are then scored) or embedding-based systems (where patient and trial data are converted into numerical representations and ranked by similarity).
Performance and Cost Insights
While direct comparisons across studies were challenging due to varied tasks and datasets, one consistent finding emerged: the GPT-4 model (and its variant, GPT-4o) generally outperformed other models, including finely-tuned open-source ones, in tasks like eligibility extraction and patient-to-trial matching. This superior performance often comes with higher costs, but studies also showed that LLM-assisted matching can be significantly cheaper than traditional human screening.
For instance, one study reported a per-patient cost of just $0.17 for an OncoLLM pipeline (based on Qwen-1.5-14B), compared to $6.18 for GPT-4. However, advanced strategies like RAG can drastically reduce GPT-4’s cost, bringing it down to as little as 2 cents per patient in some cases. LLMs also offer speed, with studies reporting evaluation times between 1 and 12.4 minutes per patient for a trial.
Also Read:
- Navigating the Future of Healthcare: A Deep Dive into Large Language Models in Medicine
- Understanding Medication Events in EHRs: A Comparative Study of AI Models
Challenges and Future Directions
Despite the promising advancements, several challenges remain. A major hurdle is accessing sufficiently large and diverse real-world datasets, as privacy concerns (PHI) restrict data sharing. Many models are currently trained on synthetic or simplified data, which may not fully capture the complexity of real-world electronic health records (EHR). Deployment challenges also include managing costs, mitigating the risk of ‘hallucinations’ (generating incorrect information), preventing data leakage, and addressing potential biases in the models.
The review highlights several promising strategies for the future: zero-shot prompting with proprietary LLMs like GPT-4o (which requires less annotated data), advanced retrieval methods, and fine-tuning smaller, open-source models for institutions that need to prioritize data privacy and host models within their own infrastructure. Transparency is also key; LLMs that provide rationales for their matching decisions, citing specific sentences from input, can improve accuracy and allow human users to verify and correct mistakes.
Ultimately, the successful implementation of LLM-based trial matching will require standardized evaluation metrics, more realistic test sets, and close collaboration with clinical stakeholders like trial coordinators. These tools are designed to augment, not replace, human decision-making, ensuring that critical clinical details are not overlooked and that these powerful technologies integrate seamlessly into existing healthcare workflows.
