AI for Evidence Generation · TrialGPT 2.0

Towards AI-Assisted Clinical Trial Matching: Practical Considerations, Multicenter Evaluation, and Real-World Deployment

Yin Fang, Qiao Jin, Shubo Tian, Lauren He, Maya Geer, Noor Naffakh, Ryan Huu-Tuan Nguyen, Zifeng Wang, Jimeng Sun, Charalampos S. Floudas, James L. Gulley, Kamilia Moalem, Catarina Martins Maia, Amanda Nottke, Juan W. Valle, Melinda Bachini, Lourdes Rocha-Nussbaum, Kari Ramage, Nikita Curry, Megan Barnes, Mandy Mansaray, Darlene Gabeau, Craig E. Grossman, Heath Skinner, Michael Burczynski, NIH-TrialBench Consortium, Zhiyong Lu

arXiv preprint September 2026

PDF arXiv

These slides were generated with the help of AI and may contain errors.

01Motivation

Clinical trial recommendations require more than eligibility

Our TrialGPT framework evaluates eligibility and ranks candidate trials.
TrialGPT 2.0 also evaluates treatment fit, clinical priorities, and missing patient information.
Clinicians review the explanations and select trials for further consideration.

Source: Introduction and system design, pp. 2–4.

02Method

TrialGPT 2.0 evaluates trial fit with local clinical priorities

Inputs include patient context, a local trial corpus, and a clinician-defined matching policy.
GPT-4.1 generates structured trial-level assessments; retrieval shortlists large trial pools.

TrialGPT 2.0 architecture: optional retrieval, clinical trial-fit factors, and structured match assessment. — Trial-fit assessment includes eligibility, treatment context, and uncertainty. Fig. 1b.

03Method

TrialGPT 2.0 separates patient-trial fit from missing information

Explicit eligibility conflicts lower the fit score; missing information lowers confidence.
The explanation separates eligible reasons, ineligible reasons, missing information, and rationale.

Output	Interpretation
Fit score	Patient-trial alignment, scored from 0 to 100.
Confidence	Reliability of the assessment given available information.
Recommendation	Highly Recommended, Possible Match, or Low Fit.
Rank	Fit score first, with confidence used to resolve ties.

Local matching policies guide prioritization without overriding explicit eligibility criteria. Source: Methods, p. 16.

04Evaluation

Clinical evaluation includes five retrospective workflows

The retrospective evaluation includes 288 cases and 1,340 reviewed patient-trial pairs.
Four cohorts evaluate top-ranked recommendations; NCI evaluates all 9 trials for each case.

Cohort	Cases	Clinical workflow
NIH Office of Patient Recruitment	19	Call intake and broad trial discovery
Cholangiocarcinoma Foundation	10	Disease-specific trial navigation
University of Illinois Cancer Center	15	Precision oncology tumor-board referral
UPMC Hillman Cancer Center	144	Radiation-oncology consultation
National Cancer Institute	100	Oncology referral and exhaustive trial review

Top-ranked review: 188 cases and 440 pairs. NCI exhaustive review: 100 cases and 900 pairs. Source: Figs. 2a and 3, pp. 5–8.

05Result

TrialGPT 2.0 identifies a recommended trial for 91% of cases

Clinicians evaluate up to 10 Highly Recommended trials for each of 188 cases in four cohorts.
Hit rate measures cases with at least one clinician-recommended trial in the evaluated list.

Evaluated list	Hit rate	Recommended precision	Eligible precision
Top 1	88%	88%	89%
Top 10	91%	83%	87%

Eligible precision counts Recommended and Eligible but Not Recommended labels. These metrics evaluate top-ranked outputs, not recall of all suitable trials. Source: Fig. 3b and Results, pp. 7–8.

06Result

TrialGPT 2.0 reduces clinician screening time by 55.0%

Two clinicians review 100 NCI cases in a counterbalanced experiment.
Mean trial-level decision time per case decreases from 129.8 to 58.4 seconds.

129.8 s

Without assistanceseconds per case

58.4 s

With TrialGPT 2.0seconds per case

Time is summed over 9 trials per case and excludes patient-context review and automated report generation. Source: Fig. 3d and Results, pp. 8–9.

07Result

TrialGPT 2.0 adds trial recommendations for 10 additional cases

The prospective tumor-board evaluation includes 27 cases reviewed from February to July 2026.
Cases with a clinician-selected trial recommendation increase from 11 to 21 (90.9%).

Routine workflowof 27 cases

Routine workflow + TrialGPT 2.0of 27 cases

TrialGPT 2.0 uniquely identifies 37 of 54 final recommendations. This evaluation measures clinician-selected recommendations, not trial enrollment. Source: Fig. 4a–b and Results, pp. 9–10.

08Result

TrialGPT 2.0 improves target-trial recovery on NIH-TrialBench

NIH-TrialBench contains 126 clinician-authored synthetic vignettes from 11 NIH Institutes and Centers.
Target-trial Recall@10 measures recovery of the intended trial within the 10 highest-ranked results.

Method	Target-trial Recall@10
Gemini-3.5 Flash Extended Thinking	21%
ChatGPT-5.5 Thinking	43%
TrialGPT 1.0	54%
TrialGPT 2.0	70%

Macro-average over the 11 contributing Institutes and Centers, using 1,373 candidate trials. Source: Fig. 5c and Results, pp. 11–12.

09Limitations

Clinical benefit requires evaluation beyond trial recommendations

The study evaluates recommendations, agreement, and screening time; enrollment and patient outcomes are not measured.
The prospective analysis includes 27 cases in one precision oncology tumor board.
Clinical evaluation primarily concerns U.S. oncology workflows, and some recommendations miss specialist interpretations.

Source: Discussion, pp. 14–15.

10Summary

TrialGPT 2.0 supports clinical trial recommendation workflows

Background	Clinical teams need suitable trial options for individual patients.
Problem	Eligibility alone does not determine clinical priority.
Approach	Local matching policies guide explained trial recommendations.
Results	55.0% less screening time; 10 additional cases receive final recommendations.
Conclusion	TrialGPT 2.0 assists clinician review; enrollment benefit remains unmeasured.