Matching patients to clinical trials using LLM

Innovative solution to the patients matching to clinical trials problem.


The main objective of the project was to develop an effective large language model that enhances the accuracy and speed of processing textual medical data in the patient-clinical trial matching process.

The results obtained during the comprehensive evaluation of the developed model indicate that the tasks set on the path to achieving the goal have been successfully completed, and the key objective has been met. The GPT-TrialMatch model, implemented based on GPT-3.5, demonstrates undeniable leadership and high efficiency at all stages of the conducted analysis.

The GPT-TrialMatch model can effectively predict a research participant's compliance with the inclusion/exclusion criteria, providing not only a final eligibility conclusion for each criterion but also a substantive justification for the decision with references to relevant sentences.

Moreover, predictions at the criterion level can be used for generalized eligibility assessment. The GPT-TrialMatch model is capable of clearly identifying patients who meet trial criteria, patients who do not meet the requirements, and patients who lack necessary information.

It is worth noting that the conducted study had several limitations. Firstly, the manually synthesized datasets for training and testing the model do not fully reflect real scenarios, as they were created using conclusions from the GPT-4 model. Secondly, the analysis of the error characteristics of the developed model indicates issues with missing important details and the inability to process implicit information.

These described limitations provide a basis for further research. To overcome these challenges, it is suggested to utilize real clinical data and adapt the implemented methods to include more representative instances. Additionally, it would be advisable to consider further fine-tuning the model on a larger dataset, ensuring that examples addressing the issues of missing important details and the inability to process implicit information are included.

In summary, it can be stated that the implementation of such models can significantly improve the efficiency of medical research, contributing to enhanced quality of medical services through the automation of medical record processing and, consequently, the faster adoption of new medical technologies and improved patient care.