Massive Bio publishes trial‑matching study

Most cancer drug trials start with a paperwork problem: every study has a long checklist, and every patient has scans, lab reports, pathology notes, and genetics results scattered across dozens of pages. Matching the two is often still done by hand, one chart at a time. (massivebio.com) A clinical trial match is basically a yes-or-no fit test. A patient has to satisfy inclusion rules like cancer type or mutation, and avoid exclusion rules like prior treatments or organ problems that would make the trial unsafe or invalid. (massivebio.com) That sounds simple until you remember that one cancer center can have hundreds of open studies, and one patient record can run to dozens of documents. Massive Bio says its new study processed more than 157,000 clinical document pages to evaluate 3,804 patients with metastatic cancer in routine practice during 2024. (massivebio.com) The company’s system tries to turn that pile of text into a map. It uses what Massive Bio calls a knowledge graph, which is a structured web linking facts like tumor type, biomarker, drug, trial site, and eligibility rule so the software can compare patient records against trial requirements. (massivebio.com) The new part is not just that software suggested matches. Massive Bio says the study was prospective, which means the system was used on incoming real patients in live workflows, instead of being tested later on old records where the answers are already known. (pharmiweb.com) That distinction matters because many artificial intelligence papers in health care look good on retrospective data and then struggle in the clinic. Massive Bio says its study generated more than 17,000 oncologist-confirmed patient-trial matches across those 3,804 patients. (pharmiweb.com) The study was published in ESMO Real World Data and Digital Oncology, in a special issue on artificial intelligence in clinical oncology. Massive Bio says the paper reports trial matching that was four times faster than conventional methods, with an F1 score of 0.82, a standard measure that combines missed matches and wrong matches into one accuracy number. (massivebio.com) (morningstar.com) Massive Bio describes the engine as neuro-symbolic and multi-agent. In plain English, that means one part reads messy human language from medical records, while another part checks those findings against explicit trial rules instead of treating the whole problem like a black box. (massivebio.com) The bottleneck this is aimed at is enrollment, not drug discovery. Clinical trials can fail to recruit enough patients even when eligible people exist, because the right patient, doctor, site, and study do not line up at the right moment. (massivebio.com) What this paper adds is a real-world count large enough to show the workflow can run at scale: 3,804 patients, metastatic disease, live oncology practice, and more than 17,000 confirmed matches. That does not prove better survival or higher enrollment by itself, but it does show that automated matching can move from demo to operations. (massivebio.com)