Fig. 4: Data extraction experiment results.
From: Accelerating clinical evidence synthesis with large language models
a In the TrialMind platform, users can define the name and description of target data fields and create a structured list. Upon clicking the “Extraction” button, TrialMind processes the selected studies and populates the study table with new columns representing the specified fields. Users can also click on extracted values to view the source text referenced by TrialMind. b Accuracy of data extraction across four therapy topics, with results further stratified by data type. c Confusion matrix illustrating hallucination and missing rates in data extraction across three data types. d Comparison of trial result extraction accuracy between TrialMind and baseline methods across four topics. e Comparison of trial result extraction accuracy between TrialMind and baselines across the most frequent clinical endpoint types. f, Error analysis of result extraction: Inaccurate—incorrect data extraction; Extraction Failure—TrialMind failed to extract data and returned null; Unavailable Data - target data was absent in the input document, resulting in null output; Hallucination—TrialMind generated data that did not exist in the input document. g In TrialMind's result extraction step, users can define the target clinical endpoint and cohort based on conditions and treatments. Clicking the extraction button triggers the extraction process for all selected studies, with results exportable in tabular format for further analysis.
