I was speaking at a primary care meeting in Canada this weekend and heard a talk on the NHS-Galleri trial testing the Grail multicancer blood test. The trial was negative.
Then I remembered that an FDA advisory panel voted 10-0 on safety, 6-4 on efficacy and 7-2 on benefits greater than harms.
More data is being presented this week at the European Society of Medical Oncology meeting.
I searched Sensible Medicine and found that we have not done a formal review. So I went and looked at the September 2026 NEJM publication.
Here is what I learned
NHS Galleri
This was a blinded RCT in 8 English cancer alliance regions. Blood was obtained from adults 50-77 years of age at up to three annual visits. About 71,000 adults were randomly assigned to have the blood tested with the multi-cancer test and another 71,000 had their blood stored.
Trial enrollees had to be cancer-free in the previous 3 years and not be undergoing tests for anything suspicious. Data analysis was conducted independently by Queen Mary University of London and verified by Grail. By design, participants from the most socioeconomically deprived areas were overrepresented.
The primary endpoint was the incidence of stage 3-4 cancer across 12 types of cancer after 3 rounds of tests and at least 12-months of follow-up. A key secondary endpoint was stage 4 cancer in the 12 types. The trial was not powered for mortality. Grail funded the trial and helped draft both papers.
Results:
For the primary endpoint, 706 participants in the intervention group and 688 participants in the control group had received a diagnosis of stage III or IV cancer among the 12 prespecified cancer types, which yielded an incidence rate per 100,000 person-years of 300.5 in the intervention group and 292.5 in the control group. This was not significantly different at an incidence rate ratio, 1.03; 95% CI, 0.92 to 1.14; P=0.63).
For the key secondary endpoint of stage 4 cancers, the incidence rate ratio (0.86 (95% CI 0.74–1.00) favored the intervention arm but had an upper bound of the confidence interval of 1.0. Formal statistical testing was not done because the primary endpoint was not significant.
The authors looked at cancers that did not have a screening test available: 1.05 95% CI 0.96-1.14.
The early-cancer numbers are revealing. The screened group had more early-stage cancers diagnosed: 647 vs. 559 among the 12 prespecified types (RR 1.16, 95% CI 1.03–1.30, a post hoc analysis). Yet there was no reduction in late-stage cancers, the primary endpoint. There are two explanations. One is overdiagnosis: extra cancers found that would never have caused harm. The other is that screening simply moved some diagnoses earlier in time, and the matching cancers in the control group will surface with longer follow-up. Either way, earlier diagnosis has not yet translated into fewer late-stage cancers. The authors say they cannot distinguish these possibilities, though they argue overdiagnosis is biologically unlikely.
About 1% or 1800 adults in the intervention arm had a positive test for cancer. Of these, cancer was found in about half (52%). [I will come back to this.]
A follow-up Nature Medicine paper described the test characteristics.
The test has a strong specificity of about 99.5%, meaning few cancer-free people test positive (≈ 0.45%). But it had a terrible sensitivity of only 30%, meaning it missed a lot of cancers.
A few summary statements from the Nature Medicine paper:
Over three annual visits, the test caught about a quarter of all cancers, and about 45% of the 12 high-mortality ones. But about 1% of cancer-free people had a false positive test.
You may be drawn to the low false positive rate of the test. A specificity of 99+% seems good on the surface.
I like to think of tests with likelihood ratios, which are how much a positive test increases the odds of a disease over baseline and vice-versa, how much a negative test reduces the odds of having the disease.
The positive likelihood ratio is the sensitivity (≈ 0.30) divided by the 1-specificity (.005), which is roughly 69. Anything over 10 is considered strong.
The negative likelihood ratio is 1-sensitivity (0.70) divided by the specificity (.995), which is roughly 0.70.
Now apply these numbers to people age 50-77 who might get the test. A positive result raises the baseline risk of 1.5% of having cancer to 52% (48% would be a false positive). A negative result however does little to the baseline risk of 1.5%, lowering it to only about 1.1%.
This is not good.
It means that if you have a positive test, it’s almost as likely to be a false positive as a true positive. And if you have a negative test, you are not much less likely to have cancer as you were before the test. If a younger person gets the test and their cancer risk at baseline is 0.5% instead of 1.5%, the likelihood ratio only changes it to 26%, so here, there is a 3-fold greater chance that the test is a false positive.
Summary Comments:
I do not understand the FDA advisory committee recommendations. In the randomized controlled trial there was no difference in the primary endpoint of stage 3-4 cancers. The primary endpoint failed to show a benefit. The slightly positive secondary endpoint is not persuasive.
In the Nature Medicine paper, the test characteristics are not great. A positive test leads to as many false positives as true positives, and a negative test is not at all reassuring.
Imagine being a person in the group where your test was positive, but no cancer was found. Now what? How many more tests do you have to get before being reassured?
It would likely be worse in the US. Here, every time a US patient has a test, there is chance of finding an incidental finding, which leads to more tests and more risk of medical harm. Notable is that the trial, as large as it was, did not systematically look at harms once participants were referred for diagnostic workup.
I looked up the Grail test and found that it is available for about a $1000. If FDA approved, insurers (aka you and me), will be asked to pay for it. This is clearly a bad thing.
Let me know if I have anything wrong here.
Great review.
Particularly timely one day after the COI post on SM. So Grail “helped” to write those 2 papers…isn’t that interesting. It sure would be interesting to see the COI disclosures of those panelists. How do 6 people vouch for “efficacy” and 7 for “benefits>risks”, for a product whose landmark trial was negative on the primary endpoint? When people wonder why experts are no longer trusted, that FDA advisory panel could be exhibit A.
This is (yet another) test for the worried well. And with that false positive rate, it will give lots of the worried well plenty of reason to keep on worrying. For them, it will be $ well spent and power to them. But I totally agree there is no way this should receive 1 red cent of public funding.
Not assessing harms of subsequent testing seems unfathomable to me. If they measure and report OS over 5-10 years I suppose that could be a decent proxy but the whole concept of medicalizing otherwise healthy individuals without measuring harms is wild. Anyone run the NNT / NNH on this?