We explore large-scale training of generative models on clinical trial registry data. First, we use clustering to group clinical studies based on key characteristics—such as duration, indications, and interventions—so that each cluster reflects studies with similar design parameters. For each cluster, we train a transformer-based model that leverages structured data (with unique NCTID codes) to learn context-specific design patterns. Our system, Neuroute, is then able to generate new study protocols that mirror successful randomization schemes and other design features observed in historical data. Our preliminary results suggest that this clustering-driven, generative modeling approach could streamline clinical trial design, paving the way for more efficient, semi-autonomous clinical development processes.
This technical report focuses on two main aspects: (1) our method for clustering diverse clinical trial data into a unified representation that facilitates large-scale training of generative models, and (2) a qualitative evaluation of Neuroute’s capabilities and limitations. Note that model architecture and implementation details are not included in this report.
Much prior work has studied the use of AI in clinical development using a variety of methods, including recurrent neural networks [1–3], generative adversarial networks [4–7], autoregressive transformers [8-10], and diffusion models [11–13]. These approaches often focus on specific types of clinical data or fixed study designs. In contrast, Neuroute is designed as a generalist model for clinical development, capable of generating protocols that span diverse therapeutic areas, study durations, and patient populations.
Turning study data into clusters
Our clustering approach mirrors the selective “cutting” process of RNA polymerase [13, 14]. We begin by compressing clinical study data into a knowledge graph that captures each study’s core characteristics. Next, we apply clustering algorithms to “cut” and group characteristics into distinct clusters for inclusion and exclusion parameters, outcome measures, site, and sponsor information. Clusters are grouped together by similar therapeutic areas, durations, and design features. This process filters out extraneous details and distills the data into coherent, representative clusters, enabling our generative model, Neuroute, to efficiently learn and synthesize innovative clinical trial protocols [15, 16, 17].
At a high level, we distill clinical study data into its fundamental components—much like breaking down a complex structure into individual Lego blocks. We then cluster these elements based on their core characteristics, enabling us to systematically evaluate performance and fine-tune protocols for optimal outcomes.
Defining patient population
Balancing the statistical significance of clinical trial outcomes with patient convenience is a critical aspect of trial design. Ensuring that trials are accessible and considerate of patient needs can enhance recruitment and retention, thereby improving the generalizability and applicability of the results [18]. Designing trials with patient convenience in mind can lead to higher participation rates, which in turn can provide more robust data and enhance the statistical power of the study [19].
To build a patient population on Neuroute, users start by inputting their protocol parameters. This enables tailoring of the study population to those most likely to benefit from the intervention which can increase the likelihood of demonstrating treatment efficacy. Neuroute ranks clusters by patient preference and correlated randomization rates for users to select patient population and outcomes parameters that are more favourable to better participation and retention.
Risk mitigation network
We train a network that reduces the risk of study failure. This network compares clusters in failed and successful studies. Clusters that are semantically unique in failed studies are then tagged. This has been found to be particularly useful in deciphering patient population parameters where automatically detecting and tagging restrictive eligibility criteria acts to improve patient inclusion. This targeted tagging enables researchers to identify potential barriers in trial design and make informed adjustments, enabling improved efficiency and efficacy in clinical development.
Quantifying failures
To best mitigate failure, we first quantified the cause of failure in the cases of failed studies. This involved multi-layer sub-grouping. The first sub-groups by failure due to 1) inefficiency or 2) inefficacy. Within sub-groups 1 and 2, we created a targeted identification and tagging system to label all the reasons for study failure. We used a multi-agentic system connected by LLMs to compile all existing media, research papers and articles related to the study failure to train and provide a deeper understanding of how to mitigate against the failure type.
Understanding strategies by sponsor
We use the same targeted tagging system to identify and tag every unique sponsor across clinical research registries, building a profile of study clusters for each sponsor. This enables Neuroute to evaluate trends and patterns in sponsor decision making across different indication areas. This is particularly interesting when evaluating patient population sub-grouping for different sponsors to forecast future market asset share.
Site selection based on study design
Neuroute systematically ranks and filters sites based on protocol parameters. Granular semantic mapping of criteria enables the selection of sites with a significantly greater feasibility match score for a particular study. Our feasibility match score is a function of i) the similarity of past patient cohorts randomized into the site, ii) site availability, iii) site speciality, and iv) historical site performance - this is the randomization (RAND) rate.
Forecasting randomization rates
To improve study design, Neuroute forecasts randomization rates based on historical data. By analyzing past trials with similar eligibility criteria, therapeutic areas, and study sites, Neuroute predicts the likelihood of successful patient randomization. This is done by training a model on stratified patient cohorts, assessing dropout risks, and identifying potential bottlenecks in patient recruitment. The result is an adaptive, data-driven approach to trial planning that ensures adequate enrollment while reducing the likelihood of underpowered studies or excessive delays. Users can input real-time values as the study progresses enabling them to quickly shift to an alternative scenario (plan) under specific parameters e.g. if drop-out >30%.
Utilizing AI agents for site support
Neuroute AI is on the first multi-agentic platforms in the clinical development space. Our agents help the user save time in the initial stages of site feasibility. This allows clinical development staff to engage and visit sites only where relevant rather than spending a significant amount of time on exploratory activities. Additionally, AI agents enhance adaptive trial methodologies by integrating real-time study data to refine ongoing scenarios, ensuring the model remains relevant throughout the trial lifecycle. Updates are then shared across the team allowing for greater visibility across the organization.
Adaptive planning
Clinical trial timelines are often disrupted by unforeseen setbacks such as slow recruitment, site unavailability, or regulatory delays [20,21,22]. Neuroute employs adaptive planning techniques to mitigate these risks. By continuously monitoring real-world data, Neuroute dynamically adjusts study timelines, resource allocation, and recruitment strategies. This allows sponsors to proactively respond to deviations rather than react to failures. Users can build models that act as contingency strategies based on prior trials, identifying optimal adjustments to keep studies on track without compromising scientific or regulatory standards.
Discussion
By structuring trial data into modular components, Neuroute effectively captures design patterns and failure modes, leading to more efficient and informed protocol generation. Forecasting randomization rates and optimizing site selection enhance feasibility assessments, reducing delays in patient recruitment and site availability. Additionally, the AI-driven risk mitigation network identifies and mitigates common failure points, offering a proactive framework to improve trial success rates.
The integration of multiple advanced clustering techniques including frameworks like TCBPMA (Text Clustering Based on Pre-trained Models and Autoencoders) into clinical trial planning bridges the gap between academic research and real-world applications. Neuroute employs a similar approach, leveraging deep clustering and AI-driven synthesis to refine study designs dynamically. Future research will focus on forecasting opportunity areas and further incorporating patient-generated data into patient population modelling.
References
Recurrent Neural Networks in Clinical Development:
Rasmy, L., et al. (2021). Simple Recurrent Neural Networks is all we need for clinical events predictions using EHR data. arXiv:2110.00998.
Choi, E., et al. (2016). Doctor AI: Predicting Clinical Events via Recurrent Neural Networks. Proceedings of Machine Learning Research, 56, 301–318.
FDA. (2019). A New Approach in Pharmacometrics: Recurrent Neural Networks for Modeling Drug Exposure and Drug Response.
Generative Adversarial Networks in Clinical Development:
Kuo, N. I., et al. (2022). Generating Synthetic Clinical Data that Capture Class Imbalanced Distributions with Generative Adversarial Networks: Example using Antiretroviral Therapy for HIV. arXiv:2208.08655.
Shi, Y., et al. (2023). 3DGAUnet: 3D Generative Adversarial Networks with a 3D U-Net Based Generator to Achieve the Accurate and Effective Synthesis of Clinical Tumor Image Data for Pancreatic Cancer. arXiv:2311.05697.
Jeong, J. J., et al. (2022). Systematic Review of Generative Adversarial Networks (GANs) for Medical Image Classification and Segmentation. Journal of Digital Imaging, 35, 137–152.
Wang, R., et al. (2022). "Applications of Generative Adversarial Networks in Neuroimaging and Clinical Neuroscience." arXiv:2206.07081.
Autoregressive Transformers in Clinical Development:
Rasmy, L., et al. (2021). Simple Recurrent Neural Networks is all we need for clinical events predictions using EHR data. arXiv:2110.00998.
Choi, E., et al. (2016). Doctor AI: Predicting Clinical Events via Recurrent Neural Networks. Proceedings of Machine Learning Research, 56, 301–318.
Hügle, M., et al. (2020). A Dynamic Deep Neural Network for Multimodal Clinical Data Analysis. arXiv:2008.06294.
Diffusion Models in Clinical Development:
Shi, Y., et al. (2023). 3DGAUnet: 3D Generative Adversarial Networks with a 3D U-Net Based Generator to Achieve the Accurate and Effective Synthesis of Clinical Tumor Image Data for Pancreatic Cancer. arXiv:2311.05697.
Jeong, J. J., et al. (2022). Systematic Review of Generative Adversarial Networks (GANs) for Medical Image Classification and Segmentation. Journal of Digital Imaging, 35, 137–152.
Kuo, N. I., et al. (2022). Generating Synthetic Clinical Data that Capture Class Imbalanced Distributions with Generative Adversarial Networks: Example using Antiretroviral Therapy for HIV. arXiv:2208.08655.
RNA Polymerase II as a Selective Sensor:
Xu, Y., & Li, L. (2016). RNA Polymerase II Acts as a Selective Sensor for DNA Lesions and Modifications. Transcription, 7(5), 227–234.
Selective Constraint and Evolution of RNA Polymerase II CTD:
Eickbush, M. T., & Laird, C. D. (2013). Selective Constraint and the Evolution of the RNA Polymerase II C-Terminal Domain. Transcription, 4(4), 185–192.
Clustering Algorithms in Clinical Trial Design:
Hügle, M., et al. (2020). A Dynamic Deep Neural Network for Multimodal Clinical Data Analysis. arXiv:2008.06294.
Clinical Trial Design:
Briel, M., et al. (2021). Exploring Reasons for Recruitment Failure in Clinical Trials: A Qualitative Study with Clinical Trial Stakeholders in Switzerland, Germany, and Canada. Trials, 22, 844.
Fogel, D. B. (2018). Factors Associated with Clinical Trials That Fail and Opportunities for Improving the Likelihood of Success: A Review. Contemporary Clinical Trials Communications, 11, 156–164.
Carlson, J. J., et al. (2021). Barriers to effective clinical trial completion: A systematic review of delays and solutions. Trials, 22, 387.
Kearney, A., et al. (2018). Identifying reasons for slow recruitment in clinical trials: A review of experiences from multicenter trials. Trials, 19, 557.
Hwang, T. J., et al. (2016). Failure of Investigational Drugs in Late-Stage Clinical Development and Publication of Trial Results. *J
Livia Ng, Darsh Jadhav, Arunavo Ray, Qin Yu, Udit Kumar, Pejvak Moghimi