Journal Club: PATHFINDER (GRAIL)

Study Design

PATHFINDER was a prospective cohort study conducted at seven US health networks, including both oncology and primary care outpatient clinics. Between December 2019 and December 2020, the investigators enrolled a "convenience sample" (a non-probabilistic sampling method where participants are chosen based on accessibility and willingness to participate, in contrast to random selection) of 6,662 adults aged 50 or older without signs or symptoms of cancer, of whom 6,621 had analyzable MCED results. Fifty-six percent (3,681) had at least one cancer risk factor, including a history of smoking 100 or more cigarettes, genetic predisposition based on established guidelines, or prior cancer history with definitive treatment at least three years before enrollment, while 44% (2,940) did not have any risk factor.

The study did not mandate any specific diagnostic workup. If the test returned a "cancer signal detected" result, the predicted cancer signal origin was shared with the participant's physician, and the physician decided what to do next. Participants were followed for 12 months, at which point cancer status was confirmed through electronic health record review.

Participants were overwhelmingly white (91.7%), highly educated (64.6% with college degrees), and already adherent to standard screening (92% were up to date on colorectal cancer screening, and 80% of women were up to date on mammography screening). Nearly a quarter (24.5%) had a prior cancer history.

PATHFINDER used an earlier version of GRAIL’s MCED test, not the refined Galleri version that is currently available. After the study launched, GRAIL developed a refined test with a higher specificity threshold for hematologic signals, no "indeterminate" cancer signal origin category, and predictions limited to the two most likely tumor types. The study assessed both versions (the original was returned to physicians, and the refined version was analyzed retrospectively, and the results were displayed separately in the publication).

Results

Of 6,621 participants, 92 (1.4%) received a positive test result. Of those 92, 35 (38%) were true positives, and 57 (62%) were false positives. Seventy-three percent of true positives achieved diagnostic resolution within three months. No adverse events resulted from diagnostic workups. There were adverse events reported for four participants: two with events related to phlebotomy (anxiety related to phlebotomy and bruising at the venipuncture site where blood was drawn from their arm) and two with anxiety that was reported before results were available.

Regarding the diagnostic workup in the 92 participants with a cancer signal detected (n = 90 which excluded two participants whose workups started before the test results were available), 79% of true positives and 88% of false positives underwent laboratory tests, 91% of true positives and 93% of false positives underwent imaging (most commonly PET-CT at 61%, CT at 39%, MRI at 21%), 82% of true positives and 30% of false positives underwent diagnostic procedures, and 3 true positives and 1 false positive underwent a surgical procedure. The 3 true positives who underwent a surgical procedure had a mesenteric lymph node excision, mesenteric mass and lymph node biopsy, and tongue biopsy. The patient with a false positive who underwent a surgical procedure had an orchiectomy, which was prompted by abnormal imaging. Notably, as it relates to imaging, 53% of the 83 patients who underwent an imaging study underwent more than one imaging study.

PPV was 38% (43% among those with risk factors (24/56) and 31% among those without risk factors (11/36)). Negative predictive value was 98.6% (6,235/6,321). Specificity was 99.1%. The number needed to screen to detect cancer was 189. Cancer signal origin accuracy was 97% for the first or second prediction (33/34, excluding 1 indeterminate). With the refined test version analyzed retrospectively and reported separately in the paper, PPV improved to 43.1% and specificity to 99.5%.

Across the entire study, 122 cancers were diagnosed in 121 participants over the 12-month follow-up. Of those, 35 (29%) were detected by the MCED test, and 86 (71%) were not. The cancers not detected by the MCED test were detected by standard screening (29), nonstandard screening (9), or clinical presentation with symptoms. Among the 35 true positives, 36 cancers were found (one participant had two cancers), which included 19 solid tumors and 17 hematologic malignancies. Seventy-four percent of the true positives detected by the MCED test were cancer types that lack United States Preventive Services Task Force (USPSTF)-recommended screening.

What I Think

The good news:

1. The specificity seen in CCGA held up in PATHFINDER's prospective population, confirming that the false positive rate stays low when the test is moved to an outpatient screening population.
2. The fact that 74% of the cancers detected by the MCED test were types without a USPSTF-recommended screening test is important. When the test worked, it found cancers that existing approved screening measures don't catch.
3. The test had high cancer signal origin accuracy (97%), which should've theoretically helped guide the diagnostic workup (more on this below in the limitations).
4. No serious adverse events were reported related to diagnostic workup. Invasive procedures, like biopsies, are not harmless, and it is reassuring that no patient had a reportable complication.

Limitations of the study are worth noting.

1. Depending on who you ask, the arbitrary "three-month" threshold to confirm the diagnosis in true positives could seem like a long time. In an affluent community composed of individuals with ample resources, which I assume was a non-trivial number of patients in PATHFINDER based on the demographics of the study population, the time to obtain imaging and biopsy should be shorter than this. A caveat is that this study was conducted during the COVID-19 pandemic. Enrollment happened between December 2019 and December 2020, which likely affected both willingness to participate and speed of diagnostic workups. The 162-day median resolution for false positives may partly reflect pandemic-era delays in scheduling imaging and procedures.
2. The cancer signal origin prediction, accurate 97% of the time for the first or second prediction, while great, should have seemed to give clinicians a clear starting point for the workup. However, PET-CT (a specialized scan of the entire body, which is arguably the least-focused type of imaging study) was the imaging study used most often, and a significant number of patients underwent multiple imaging studies. This raises questions about cost-effectiveness.
3. Of the 92 positive results, 57 were false positives, resulting in a PPV of 38%, meaning roughly 3 out of 5 people who received a "cancer signal detected" result did not have cancer. They went through imaging, lab work, and invasive procedures before learning they didn't have cancer.
4. Sixty-one percent of false positives had a hematologic cancer signal origin prediction, and a third of those (12/35) had precursor conditions like monoclonal gammopathy of undetermined significance (MGUS). MGUS is present in roughly 3% of adults over 50 and is, by definition, not cancer (but it can theoretically lead to cancer). The methylation assay picked it up as a cancer signal. The refined test addressed this by raising the detection threshold for hematologic signals, which is likely why specificity improved.
5. Only 29% of all cancers diagnosed in the study population during the 12-month follow-up were detected by the MCED test. Seventy-one percent were found by other means, including standard screening, nonstandard screening, or symptom presentation.
6. A cohort that is 91.7% white, mostly college-educated, and already compliant with existing screening recommendations provides limited insight into how this test performs in the populations that stand to benefit most, which are those with limited healthcare access and poor screening adherence. Therefore, generalizability remains a question.