Deep learning-based risk prediction for interventional clinical trials based on protocol design: A retrospective...

Area of Science:

• Artificial Intelligence
• Clinical Trial Management
• Biomedical Informatics

Background:

• Clinical trial (CT) success rates are low, with protocol design identified as a significant risk factor.
• Predictive modeling of CT risks is crucial for optimizing trial design and resource allocation.

Purpose of the Study:

• To investigate the application of deep learning methods for predicting the risk of clinical trials based on their protocols.
• To develop and evaluate an ensemble model combining transformer and graph neural networks for risk prediction.

Main Methods:

• A retrospective risk assignment method was used to label CTs as low, medium, or high risk, considering protocol changes and final status.
• Transformer and graph neural networks were designed and ensembled to infer ternary risk categories from CT protocols.
• Performance was evaluated using the area under the receiving operator characteristic curve (AUROC).

Main Results:

• The ensemble deep learning model achieved a robust AUROC of 0.8453.
• Individual transformer and graph neural network architectures showed comparable performance to the ensemble.
• The proposed model significantly outperformed a baseline model using bag-of-words features (0.7548 AUROC).

Conclusions:

• Deep learning models show significant potential for predicting clinical trial risks directly from protocol data.
• This predictive capability can inform customized risk mitigation strategies during the clinical trial protocol design phase.
• The findings pave the way for data-driven improvements in clinical trial success rates.