How many fewer cancer deaths will occur within 10 years per 100,000 people offered annual Galleri?

TL;DR

My probability-weighted estimate is 135 fewer cancer deaths per 100,000 people offered annual Galleri over ten years; the median is 85. I assign a 14% chance of no benefit or net harm, with an 80% interval from 15 more deaths to 330 fewer deaths. The main brake is that NHS-Galleri did not reduce combined stage III/IV cancer incidence, while its favorable 14% stage IV result was a secondary endpoint and stage IV alone has performed worse than combined stage III/IV as a mortality surrogate in randomized screening trials (ASCO abstract, JAMA meta-analysis).

Context

As of September 20, 2026, Galleri has no randomized cancer-mortality result. NHS-Galleri enrolled about 142,000 adults and offered three annual tests. The primary result covered cancers diagnosed through roughly 18 months after the third appointment: combined stage III/IV incidence was 706 versus 688, an incidence-rate ratio of 1.03 with a 95% confidence interval of 0.92–1.14. The trial therefore missed its main endpoint (official NHS results, ASCO abstract).

The target here is harder. It covers ten annual offers, ten years of deaths, and everyone offered screening, including people who decline. NHS-Galleri randomized people after they had consented and supplied a baseline blood sample, so its randomized comparison does not measure the large initial uptake loss in a population programme. Mortality follow-up is planned, but the current registry contains no mortality result (trial registry, trial design).

Evidence

The historical screening record says that stage reduction is useful evidence, but not a mortality result. A 2024 review covered 41 randomized screening trials. Of the 13 trials that significantly reduced stage III/IV incidence, eight did not significantly reduce cancer mortality. The estimated mortality response to a given late-stage reduction varied sharply: it was strongest for lung cancer, weaker for colorectal and breast cancer, and inconsistent for prostate cancer. When the review used stage IV alone, correlations with mortality weakened for breast, colorectal, lung, and ovarian cancer (JAMA review, full analysis).

UKCTOCS is the sharpest warning. It randomized 202,638 women and followed them for a median 16.3 years. Multimodal ovarian screening produced 24.5% fewer stage IV cancers and 39.2% more stage I/II cancers, but no significant reduction in ovarian or tubal-cancer deaths. Combined stage III/IV incidence fell by only 10.2%. This resembles the Galleri pattern in one key way: moving cases out of stage IV is not enough when many move only into another high-mortality stage (Lancet trial report).

The reference class is not wholly negative. A separate meta-regression using cancer-specific stage definitions and longer follow-up found much tighter stage-to-mortality relationships for lung screening and metastatic colorectal cancer. Galleri targets both. This supports a positive tail, but not a universal conversion factor across dozens of cancers (Journal of Medical Screening meta-regression).

NHS-Galleri's strongest favorable result was 342 versus 397 stage IV cancers in the 12 prespecified cancers: 55 fewer cases, or about 77 per 100,000 intervention participants. The overall stage IV rate ratio was 0.86, with the upper confidence bound at 0.998. The round-specific reductions were 9%, 22%, and 26%. Yet combined stage III/IV moved from 19% higher in the prevalence round to 5% and 12% lower in the two incident rounds; neither incident-round estimate excluded no effect (ASCO abstract, trial factsheet).

Other randomized signals point in the right direction. Stage I/II diagnoses in the 12 cancers increased from 559 to 647. Across all cancers, screen-detected diagnoses rose from 290 to 1,173, clinically detected diagnoses fell by 21%, and emergency presentations fell by about 21% using the raw counts. But total cancer diagnoses were 3,637 versus 3,400, an excess of 237. That excess may shrink with longer follow-up if it is mainly lead-time advancement, but it may also contain overdiagnosis. Current data cannot separate the two (ASCO abstract).

Test performance also limits the ceiling. Three-round episode sensitivity was 30.7% across all cancers and 54.7% for the 12 selected cancers. Specificity was 99.55% and positive predictive value was 52.0%. Galleri therefore finds a meaningful subset of cancers, but misses most cancer episodes across all sites and sends nearly half of positive episodes into work-up without a cancer diagnosis during the defined follow-up (trial factsheet, official accuracy summary).

Pretrial models now look too optimistic. The Galleri trial-planning model projected a 37%–46% stage IV reduction, a 9%–24% combined stage III/IV reduction, and a 13%–16% mortality reduction after three annual rounds. The randomized results were 14%, negative 3%, and unknown, respectively (trial-planning model, ASCO result). An independent 2025 model estimated 6%–9% lower five-year mortality for the target cancers when three-year late-stage incidence fell by 6%–23%; Galleri's primary point estimate did not fall at all (JAMA Network Open model). A methodological review found that existing MCED models rely on uncertain stage-shift and natural-history assumptions that were not fully propagated through their results (model review).

A competing-risk life-table check using pooled 2018–2020 US age-specific cancer and all-cause death rates gives about 4,470 usual-care cancer deaths per 100,000 over ten years for a population-weighted cohort initially aged 50–79. I use 4,500 as the operational baseline, with a practical range of about 4,000–5,200 because the result depends on age mix, country, future mortality trends, healthy-participant selection, and prior cancer exclusions (CDC WONDER cancer-mortality data).

A bottom-up check centers the result near the distribution median. The observed three-round difference was about 77 fewer stage IV cancers per 100,000. The control arm had roughly 186 stage IV diagnoses per 100,000 per year across the 12 cancers. If the later-round 24% reduction persists for seven further rounds, then applying 70% effective population exposure and assuming 30% of avoided stage IV diagnoses become deaths prevented within the ten-year window gives approximately [77 + 7 × 186 × 0.24] × 0.70 × 0.30 = 82. The 70% factor is consistent with current NHS bowel-screening uptake of 65.2% and breast-screening coverage of 71.8%, while allowing annual re-offering after a missed round (bowel-screening statistics, breast-screening statistics).

I represent the remaining structural uncertainty with four regimes:

• 25% null or slight-harm regime: normal distribution with mean −5 and standard deviation 45.
• 45% small-benefit regime: lognormal distribution with median 80 and log-standard-deviation 0.60.
• 25% moderate-benefit regime: lognormal distribution with median 220 and log-standard-deviation 0.50.
• 5% high-benefit regime: lognormal distribution with median 550 and log-standard-deviation 0.50.

This mixture has a mean of 135.3 and a median of 86.3 deaths prevented per 100,000. Its 25th and 75th percentiles are 37 and 182. Its 10th and 90th percentiles are −16 and 328, and its 5th and 95th percentiles are −43 and 461. The high-benefit tail allows repeated screening to perform better than the first three rounds, but gives little weight to benefits near the older vendor-model projections.

What's non-obvious

The headline result is often described as more than 20% fewer stage IV cancers. That is true only for the second and third rounds. The overall reduction was 14%. More subtly, stage IV alone is not the better validated mortality surrogate: in the 2024 randomized-trial review, it performed worse than combined stage III/IV for every cancer type with enough data. Galleri's favorable result is therefore on the weaker secondary surrogate, while the stronger prespecified primary surrogate was null (trial factsheet, JAMA analysis).

The other hidden issue is the word offered. NHS-Galleri invited 1,496,311 people and enrolled 142,924, but that 9.6% research-enrollment rate is too low as a proxy for a free routine programme. Conversely, 93.8% of respondents in an English survey said they would probably or definitely accept MCED screening, which is too high as a behavioral forecast. Actual NHS screening participation near 65%–72% is the better anchor. Ignoring this initial uptake step overstates the intention-to-screen effect by roughly one-third (enrollment study, acceptability survey, NHS screening statistics).

Uncertainties

• No randomized cancer-death counts have been published. The present evidence is a conference abstract and official summary, not a complete peer-reviewed primary-results manuscript (ASCO abstract, trial registry).
• Public results do not give enough cancer-site-specific stage, treatment, recurrence, or mortality data. The answer changes sharply depending on whether the stage IV reduction is concentrated in lung and colorectal cancer or in cancers where stage III remains highly lethal.
• There are no observed rounds four through ten. Long-run attendance, diagnostic capacity, treatment delays, and the persistence of the later-round stage effect are assumptions.
• The 237 excess diagnoses have not yet been divided into lead-time advancement and overdiagnosis. Longer cumulative-incidence and treatment-harm data would narrow both the negative tail and the upper-benefit tail (ASCO abstract, 2026 systematic review).
• The target does not specify a country or exact starting-age distribution. This creates roughly 10%–20% uncertainty in baseline cancer mortality before uncertainty about Galleri's relative effect is considered.