Clinical technology
AI Trial Alerts Did Not Increase Enrollment in One Randomized Study. What Should Teams Ask Next?
Read the 2025 randomized AI-alert study accurately and use its limits to evaluate the complete trial enrollment workflow.
By Lampgen Clinical Research

3 min read · estimated
What the source reported
A 2025 JAMA Network Open randomized study included 20,707 patients with genomically characterized solid tumors at one academic center. When AI analysis indicated progression and likely treatment change, the intervention notified oncologists about genomically matched trials. Enrollment was 2.20% versus 2.03% in the control arm, with no statistically significant difference (P=.41). [1]
The finding concerns this notification intervention in this setting. It does not demonstrate that every AI matching approach fails, that the two strategies are equivalent, or which downstream barrier caused the result.
What this means for trial teams
Lampgen’s analysis: a matching system should be evaluated against the problem a team intends to solve. Finding a plausible study, sending an alert, obtaining a clinician review and enrolling a participant are different outcomes. A strong result on an upstream measure should not be relabeled as evidence of improvement at every later step.
For a vendor evaluation, write the intended use before the demonstration. Which patients and trial types are in scope? What information does the system need? Who receives the output? What decision should follow? Ask whether the supporting evidence tests that same population, workflow and outcome, rather than relying on an impressive overall matching count.
Next, map the action pathway. Someone has to review the candidate match, confirm trial availability and assess whether a patient discussion is appropriate. Screening may depend on updated results, slots or visits. These are proposed workflow considerations, not mechanisms established by the randomized study. They should be tested locally instead of treated as an explanation already proven.
Define ownership for exceptions. An alert can be inaccurate, incomplete, outdated or clinically irrelevant. The process needs a way to record that assessment and route questions without overwhelming clinicians. For operational tracking, distinguish an alert delivered from a review completed; distinguish a patient discussion from consent; and distinguish consent from completed eligibility and enrollment.
Choose denominators carefully. If a system flags many people but only a smaller group is clinically ready to consider a new treatment, its performance depends on which population is being counted. Compare the evaluation group with the intended users and patients. Consider whether the evidence includes an appropriate comparator and enough follow-up for the claimed outcome.
Patient access belongs in the assessment too. A technically valid match may lead to a trial that is geographically difficult, unavailable locally or impractical for the person’s circumstances. The system’s role and the care team’s role should be explained plainly. Automated output does not determine eligibility or replace an investigator’s clinical assessment.
The potential impact of this research is a better purchasing and implementation question: what improvement has been demonstrated, under what conditions, and what work remains after the alert? A negative result is useful when it sharpens the next test. It becomes misleading when converted into a verdict on an entire technology category.
Lampgen’s contribution is the operations perspective. Before introducing another stream of alerts, identify who can act, what information they need and how the team will evaluate the complete pathway. Technology and staffing should be assessed against the same clearly stated objective.