Background:
Patient recruitment is a major challenge in clinical trials (CT) and the most common cause of trial discontinuation. Over half (55%) of publicly funded CTs fail to reach their recruitment targets [4], leading to incomplete studies, increased costs when trials are extended, or wasted research resources when discontinued. In neurology related trials, recruitment is particularly challenging as eligibility requires integrating information across structured EHRs, clinical notes, radiology impression reports, and laboratory results. Eligibility often depends on specific imaging findings (lesion characteristics, tumour location) described in impression text, combined with clinical and laboratory evidence. Current approaches process each modality independently, demanding manual review of documentation by a clinician which ultimately becomes a bottleneck. This project addresses these challenges by exploring multi-agent architectures that coordinate across data modalities to automate cohort identification.
Evaluation framework:
● Individual agent performance: Accuracy, precision, recall for structured queries, text extraction, and information retrieval
● System-level metrics: End-to-end cohort identification accuracy against gold-standard recruitment lists
● Ablation studies: Contribution of each modality and agent to overall performance
● Error analysis: Characterizing failure modes and identifying where multimodal integration succeeds or fails
● Efficiency metrics: Time savings compared to manual review, query complexity handling
Expected Contributions
- Benchmarks and baseline results for text-to-SQL, information extraction, and multimodal integration in neurology trial recruitment
- Analysis of where multi-agent coordination provides value over single-model approaches
- Open-source tools and evaluation frameworks for cohort identification research
- Real-world deployment insights from UK neurology trial settings
References:
[1] Julia Ive, et al. Clean & Clear: Feasibility of Safe LLM Clinical Guidance. arXiv:2503.20953, 2025. https://arxiv.org/abs/2503.20953
[2] Gyubok Lee, et al. EHRSQL: A Practical Text-to-SQL Benchmark for Electronic Health Records. arXiv:2301.07695, 2023. https://arxiv.org/abs/2301.07695
[3] Kawsar Noor, et al. Multimodal clinical trial outcome prediction using large language models and transformers. Brain Communications, 2025. https://www.sciencedirect.com/science/article/pii/S1878875025002633
[4] Sully BGO, Julious SA, Nicholl J. A reinvestigation of recruitment to randomised, controlled, multicenter trials: a review of trials funded by two UK funding agencies. Trials. 2013;14:166. https://doi.org/10.1186/1745-6215-14-166
[5] Divya Gopinath, et al. Fast, Structured Clinical Documentation via Contextual Autocomplete. arXiv:2007.15153, 2020. https://arxiv.org/abs/2007.15153
[6] https://www.stopem-trial.org.uk/