Uncovering Global Trends in Surgical Oncology Trials: Applying Natural Language Processing to ClinicalTrials.gov

Identifying surgical oncology trials within the National Clinical Trial (NCT) database is challenging owing to the absence of medical specialty labels. We developed a machine learning classifier to systematically identify and analyze surgical oncology trials registered in the NCT database.

We analyzed 425,736 trials. A training dataset of 4863 trials were labeled as surgical oncology according to Complex General Surgical Oncology (CGSO) neoplasms. Labels were applied by two independent reviewers, and disagreements resolved by a third, using study descriptions and Medical Subject Headings (MeSH). Neural networks were assessed using stratified 5-fold cross-validation. Models were evaluated using accuracy, balanced accuracy, specificity, sensitivity, and Cohen's kappa. Neural embedding representations were analyzed with principal component analysis (PCA). Trial characteristics, geospatial mapping and social network analyses were evaluated.

The selected model achieved excellent accuracy (94.1% [95%CI 93.3-94.9]). PCA of neural embeddings demonstrated separation of CGSO-relevant terms. We identified 24,345 surgical oncology trials. Trial volume was strongly associated with national GDP (r = 0.96) and global disparities were identified. Network analysis revealed strong collaboration amongst USA and European institutions. Chinese institutions demonstrated high trial leadership but relatively limited external collaboration. Academic centers including Memorial Sloan Kettering and MD Anderson remained prominent, while industry participation, was consistently evident.

This study provides a machine learning approach to identify surgical oncology trials within the NCT database. Our research highlights collaboration networks amongst institutions and global disparities in surgical oncology research. Future work should incorporate funding information to better inform surgical research policy and development.