TrialGPT 2.0: trial matching that faces real oncology workflows

Multicenter retrospective and prospective evaluation of an AI-assisted clinical trial recommender: ~91% top-10 hit rate vs clinician picks, 55% less screening time, and more trial access in a live tumor board.

Primary source: arXiv:2609.01202: https://arxiv.org/abs/2609.01202

What’s new: Most “AI for clinical trials” demos stop at eligibility classification. TrialGPT 2.0 (arXiv:2609.01202, submitted 1 Sep 2026) is framed as a deployment-oriented recommender: not only whether a patient might qualify, but which trials warrant further look given clinical need and local workflow priorities, with structured, inspectable explanations for expert review.

The authors evaluate retrospectively and prospectively across oncology-focused settings (government, academic cancer center, patient advocacy, NIH referral). In retrospective multicenter cohorts (288 cases), TrialGPT 2.0 retrieved at least one clinician-recommended trial in its top 10 for about 91% of cases and cut clinician screening time by 55.0%. In a six-month prospective evaluation inside an active precision-oncology tumor board, it surfaced additional trial opportunities missed by routine workflow, expanding patient access to trial participation by 90.9% (authors’ reported figure). They also release NIH-TrialBench: 126 clinician-authored synthetic vignettes and matching scenarios from 11 NIH Institutes and Centers for reproducibility.

Why it matters: Accrual is the quiet failure mode of oncology trials. A system that sits in a tumor board, explains itself, and finds missed opportunities is closer to “AI that does something” than another eligibility F1 on a static snapshot. The dual retrospective/prospective design is the eval signal Bitware cares about, workflow time and missed options, not only ranking metrics.

Caveats: Preprint. The 90.9% “expanded access” figure is relative to what the routine board missed in that prospective window, read the paper for denominators and how “access” was counted before citing it as a population-level accrual miracle. Retrospective hit rates depend on what clinicians marked as recommended. Synthetic NIH-TrialBench helps reproducibility but is not live EHR matching. Inspectable explanations still need human gatekeeping for eligibility and ethics. Treat this as evidence that assistive matching can save screening time and catch overlooked trials in specific oncology workflows, not as autonomous enrollment.