A new AI framework lets oncologists query clinical trial evidence tables in plain language, instead of wrestling with database syntax.
Researchers built FD-SCoPE, a language-model system that answers clinicians' questions about trial evidence tables, including the kind of question the table doesn't actually contain an answer for, like a drug's target class or a harmonized endpoint. Tested on an oncology table of 159 immune checkpoint inhibitor trials, the system completed all 140 clinician-style tasks it was given, with accuracy on individual task types ranging from 90.7% to 97.9%. For questions requiring derived attributes not present in the table, it retrieved 99.3% of relevant trial records at a positive predictive value of 89.8%, beating four other approaches on a combined accuracy score (77.7% versus 64.8-73.4%). After clinicians corrected 299 of its answers, its accuracy on 1,201 new, unseen questions rose from 77.9% to 84.9%.
Evidence tables are supposed to make trial data easy to compare, but in practice they only answer questions someone thought to encode as columns ahead of time. This system shows every query it ran, which trials it pulled from, and the rule it used to derive an answer, so a clinician can check the reasoning rather than just trusting a summary.
It's a narrow test on one cancer drug class and one table, not a general medical oracle, but that audit trail is exactly the feature most chatbot summaries of research skip.