How many people per 100,000 screened will undergo an invasive diagnostic procedure because of a false-positive Galleri result?

TL;DR

Conditional on numeric resolution, my median forecast is 161.3 people per 100,000 screened, corresponding to 57 people in the current 35,335-person safety cohort (ASCO presentation). The modal bucket is greater than 160 through 165, with 53.8% probability; the modeled mean is 159.4 and the central 90% interval is 138.7–167.0. I assign a separate 10% chance of annulment because the public full-cohort materials still do not print the required false-positive-by-invasive-procedure cell.

Context

The current evidence is the full-cohort PATHFINDER 2 analysis presented on May 31, 2026, using a February 11, 2026 data cutoff. It reports 35,878 enrolled people, 35,335 in the safety analysis, 32,007 with a 12-month cancer-status assessment, 287 positive tests, 173 cancers after a positive test, and 213 safety participants with at least one invasive procedure after a positive result (ASCO abstract; full presentation).

The trial registry defines the relevant primary safety endpoint as invasive procedures among positive-test participants with no cancer diagnosis at diagnostic resolution, and follows participants for up to three years (ClinicalTrials.gov). The presentation labels the current release as 12-month data, with planned 24-month and 36-month analyses; the registry lists estimated study completion on April 30, 2028 (full presentation; ClinicalTrials.gov).

Evidence

The reference-class history is short. The peer-reviewed first PATHFINDER study reported that 17 of 57 false-positive participants underwent a procedure; that is 29.8% of false positives and 256.8 people per 100,000 among 6,621 analyzable tests (Lancet study in PMC). The prespecified initial PATHFINDER 2 analysis, with a December 31, 2024 cutoff, displayed invasive procedures in 36 of 77 false-positive participants, or 46.8%, versus 111 of 126 true positives; the displayed subgroup was narrower than the headline safety endpoint because its counts total 147 while the report says 159 of 25,114 safety participants had an invasive procedure (ESMO 2025 presentation).

The full-cohort evidence points higher than the first PATHFINDER study and close to the initial PATHFINDER 2 rate. The 12-month performance table contains 114 positive participants without cancer, calculated as 287 minus 173. The ASCO abstract says invasive procedures were 1.8 times as likely among participants with cancer, while a GRAIL presentation filed with the SEC summarizes the group rates as about 91% with cancer and 50% without cancer (ASCO abstract; SEC-filed presentation).

Let $y$ be false-positive participants with an invasive procedure. Using the published 173/114 split as the closest approximation to the safety strata gives:

$$y = 213 × 114 / (1.8 × 173 + 114) = 57.08.$$

The nearest integer is 57 people. It gives 57/114 = 50.0%, leaves 156 invasive-procedure participants on the cancer side, and produces a relative risk of 1.803. The target rate is:

$$100000 × 57 / 35335 = 161.313.$$

This is a reconstruction, not a published table cell. It is unusually well constrained. A coherent safety-specific table is 156 of 172 cancer participants and 57 of 114 no-cancer participants: those counts sum to 213, the rates round to 91% and 50%, and the relative risk rounds to 1.8. The one-person shift from 173 to 172 is plausible because the presentation explicitly uses different safety and performance populations; it also reports 290 positive-result diagnostic evaluations versus 287 positives in the performance table (full presentation).

My numeric model puts 78% weight on a final numerator near 57 with only small data-cleaning changes, 15% on later classification or reporting changes that usually reduce the count, and 7% on a wider analysis-set or definition surprise. This yields a mean of 159.4, a median of 161.3, 53.8% in the 160–165 bucket, and a central 90% interval of 138.7–167.0. The distribution is conditional on a numeric resolution; the 10% annulment estimate is outside the array.

What's non-obvious

The obvious calculation, 213/35,335 = 602.8 per 100,000, answers the wrong question. It counts invasive procedures after all positive tests, most of which were followed by a cancer diagnosis. The hidden false-positive cell is near 57, so the requested population rate is about one quarter of that headline safety rate (ASCO presentation).

Longer follow-up creates less downside than it first appears. The registered safety endpoint is anchored to cancer status at diagnostic resolution, not to never developing cancer during three years of observation (ClinicalTrials.gov). NHS-Galleri found that 54 of 303 first-round false positives were diagnosed with cancer in later rounds, but that repeated-screening result is not the PATHFINDER 2 endpoint and should not be subtracted mechanically (GRAIL analyst presentation).

Uncertainties

The main gap is the missing cross-tab. The public report mixes a 35,335-person safety set with a 32,007-person performance set, uses rounded 91% and 50% group rates, and does not disclose the exact safety-set cancer/no-cancer denominators (ASCO presentation; SEC-filed presentation). A full protocol or statistical analysis plan would also clarify whether years 2–3 can revise the safety classification, rather than merely adding longitudinal cancer outcomes.

I found no peer-reviewed full-cohort PATHFINDER 2 results paper by the cutoff; GRAIL's evidence page lists the ESMO 2025 and ASCO 2026 results as presentations rather than a full manuscript (GRAIL clinical evidence). The original PATHFINDER paper did publish the false-positive procedure count, and the PATHFINDER 2 registry makes this a primary safety endpoint, so I expect the exact cell to appear eventually; I still retain 10% annulment risk because the resolution rules require sufficient stratification rather than inference.