The Evidence Base Was Built on Patients Who Don't Represent the Real World.
The clinical evidence powering oncology decisions was generated from trial populations that don't reflect the patients actually receiving treatment.
The Evidence Base Was Built on Patients Who Don't Represent the Real World.
One Idea Worth Acting On.
An AI imaging algorithm is trained to screen for breast cancer.
It performs well. The validation dataset shows strong sensitivity and specificity.
Then it is deployed at a community clinic serving a predominantly Asian patient population.
The algorithm flags significantly fewer eligible patients than expected.
Not because the disease is less prevalent. Because the algorithm was trained on Caucasian breast tissue — and Asian women have different breast density profiles.
The algorithm was not biased. The training dataset was.
What's Actually Happening
Diversity in clinical trials has become a compliance conversation. Sponsors add DEI language to their protocols, file diversity action plans with the FDA, and track enrollment demographics against target percentages.
But the deeper problem is not whether the enrolled population looks diverse on paper. It is whether the evidence generated from that population is valid for the patients who will ultimately receive the therapy — including the patients who were systematically excluded from the trials that generated the evidence.
As Sarah Clark, Chief Digital Officer and contributor to Voices of Oncology, describes it: an AI algorithm must be adjusted to reflect real-world population differences — but that adjustment is only possible if the training data included those populations in the first place. Aggregate performance metrics hide exactly the subgroup failures the DEI commitment is supposed to surface.
The same principle applies to clinical trial evidence. A drug approved on a trial that enrolled predominantly white, male, academically-affiliated patients may perform differently — in efficacy, in toxicity, in the practical management of side effects — in the real-world populations that trial never enrolled.
Why This Matters
The evidence-to-practice gap is not just a communication problem. It is partly an evidence problem. When the trial population does not reflect the treatment population, the data generated in the trial does not fully describe the drug's behavior in the clinic.
Prescribers notice this after approval — in unexpected toxicity patterns, in efficacy that does not replicate the trial results, in patient subgroups whose outcomes diverge from the label. By then, the trial is closed. The data cannot be recovered. The gap between who was enrolled and who is being treated becomes a permanent feature of the evidence base.
And for the patients excluded from those trials — by language barriers, by geographic distance from academic centers, by eligibility criteria calibrated to healthier reference populations — the therapy that was approved for their disease was never tested in someone like them.
Where It Breaks in the Real World
A trial site is activated at a network serving a predominantly non-English-speaking catchment population. No translation materials are produced. Screening yields almost no enrollment from the majority population the site was supposed to reach.
As Dr. Jedd Wolchok, academic clinical investigator and contributor to Voices of Oncology, describes from his work as a cancer center director: ensuring that trial offerings match clinical demographics requires bringing satellite campuses to parity with the flagship — which means building enrollment infrastructure where the patients actually are, not where the trial infrastructure already exists. The patients being served at the satellite are often the patients the DEI commitment most needs to reach. They are also the patients who are least likely to appear in the enrollment data.
The enrollment metric shows the site as low-performing. The DEI report shows the trial as diverse by aggregate percentage. Neither metric reveals the specific population that was unreachable because nobody built the infrastructure to reach them.
What Needs to Change
DEI in clinical trials requires three things that most diversity action plans do not address: enrollment infrastructure designed for the populations being targeted, eligibility criteria stress-tested against those populations' actual clinical profiles, and subgroup evidence reporting that makes the gaps visible rather than hiding them in aggregate metrics.
The goal is not a diverse enrollment number. The goal is an evidence base that is valid for the patients who will receive the therapy — all of them, not just the ones who happened to be accessible when the trial opened.
The Bottom Line
Diversity in a trial is not achieved when the enrollment percentage looks right. It is achieved when the evidence generated is valid for every population that will use the drug.
Until that is the standard, the evidence base will keep reflecting the patients who were easiest to enroll — and the gap between the trial and the clinic will keep widening.
Add Your Voice — Where are you seeing this? What's working — and what's not?