Clinical Trial Risk Tool: software application using natural language processing to identify the risk of trial uninformativeness

Background: : A large proportion of clinical trials end without delivering results that are useful for clinical, policy, or research decisions. This problem is called “uninformativeness”. Some high-risk indicators of uninformativeness can be identified at the stage of drafting the protocol, however the necessary information can be hard to find in unstructured text documents. Methods: : We have developed a browser-based tool which uses natural language processing to identify and quantify the risk of uninformativeness. The tool reads and parses the text of trial protocols and identifies key features of the trial design, which are fed into a risk model. The application runs in a browser and features a graphical user interface that allows a user to drag and drop the PDF of the trial protocol and visualize the risk indicators and their locations in the text. The user can correct inaccuracies in the tool’s parsing of the text. The tool outputs a PDF report listing the key features extracted. The tool is focused HIV and tuberculosis trials but could be extended to more pathologies in future. Results: : On a manually tagged dataset of 300 protocols, the tool was able to identify the condition of a trial with 100% area under curve (AUC), presence or absence of statistical analysis plan with 87% AUC, presence or absence of effect estimate with 95% AUC, number of subjects with 69% accuracy, and simulation with 98% AUC. On a dataset of 11,925 protocols downloaded from ClinicalTrials.gov, the tool was able to identify trial phase with 75% accuracy, number of arms with 58% accuracy, and the countries of investigation with 87% AUC. Conclusion: : We have developed and validated a natural language processing tool for identifying and quantifying risks of uninformativeness in clinical trial protocols. The software is open-source and can be accessed at the following link: https://app.clinicaltrialrisk.org