The Raw Numbers You Need for MRD Statistics

My manual answer at top.  An AI re-written answer at bottom.

##

Dear BRUCE,

How many patients do you need for a successful MRD study at Moldx?

##

Dear (JOHN),

Good questions.  This area is confusing to many people, but I think I've developed a good sense of what MOLDX is looking for, and some insights into the "moats" which are bigger than one might guess.   Total population size (55, 79, 101) is not the only driver.

01 LCD

Start with the LCD.  Basically, you can get coverage if you have data for:

1) Post op check for completeness.  You have a breast or colon resection, and at 6 weeks you check for ongoing tumor DNA.

2) If the patient is negative post op, surveillance, e.g. quarterly for several years.    For small resected cancers the recurrence rate is likely around 10% per year.

3) Drug response  Serial testing during drug therapy (esp. immuno-oncology).  Rising DNA is bad, low DNA is good.

The LCD says that the MRD test must detect earlier than a standard of care test, and, with at least similar accuracy.   Earlier doesn't mean 2 days; it's not defined but is 2+ months, often 4 months or so.

02   Comparison for mental framing

Before going into MRD, consider colon cancer screening.   These are usually 10,000 patient studies.  The reason is that the cancer rate is around 1%, and FDA wants you to accumulate circa 100 cases (60+ let's say).   This is because they want precise statistics around sensitivity and specificity of the test (assuming a colonoscopy gold standard).   They want to know that sensitivity is (say) 80% plus or minus 2%, not 80% plus or minus 15% due to small numbers and noise.

With colon screening, the pivotal data is the 100 cancer patients, not the 9900 excessive no-cancer patients.  Similar for MRD, the "relapsed" headcount is key to the performance statistics, and you might need 200 MRD patients to get 20 relapses.

03  Data for MRD success

Typically, you want something like 15 or more relapsed patients.   If the relapse rate is 10% per year, then you need 150 patients for a 10% relapse of 15 patients in the first year.   

Also, you want reasonable error bars around the sensitivity.   That means you need enough relapsed cases (say, maybe 15, but more is better).  And, crucially, you need fairly frequent checks against a gold standard (say, CT).   Let's say you want 5-6 blood tests and 4-5 or more CT tests over the observation period.   

If you have 1 blood test and 1 CT test a year later, it's a statisical disaster as there's no way to say the lead time for detection is, say, 3 months plus or minus one month with so little data.

So:  #1 Total number of patients, and #2 Relapse patients (e.g. at a 10% rate), and #3 frequency of blood sampling, AND ALSO, #4 frequency of standard of care CT test.   This is why data is often summarized as "125 patients and 550 total blood samples".   They are flagging there were about 5 blood samples per patient.  If you are missing any one of those parameters (#1 #2 #3 #4) the statistics flop.

04  MOAT

The moat is wider than you might guess.  If you use NEW data, you must start the trial, enroll patients, and follow up with minimal study loss over two years.  While getting them back for serial testing every couple months going into a second year in the future.   Then allow another year for publication and moldx review.   

Let's say you use BIOARCHIVE data.  Then you are at the mercy of how many patients the archive collected, how many frozen plasma samples, how many timed and matched CT scans for the gold standard.   

05 A GOOD EXAMPLE

MolDx recently approved a brast cancer MRD which is PATHLIGHT from SAGA DIAGNOSTICS (Univ. Lund).

https://www.discoveriesinhealthpolicy.com/2025/08/moldx-greenlights-pathlight-new-mrd.html

See the article by ELLIOTT 2025, which summarizes all their data in one place.   

See also the ASCO announcement from NATERA that they had perfected a genome-based MRD test.  i don't have it in front of me, but i think they maybe had something like 500 patients and 3000 blood samples.   That second number contributes to the precision of the statistical data of the MRD test.

##

##

##

AI RE WRITE

Here’s a polished rewrite in a more professional style while keeping the structure and logic intact:

---

The Raw Numbers Behind MRD Study Design

Dear (John),

You’ve raised an excellent question. This area can be confusing, but over time I’ve developed a clear sense of what MolDx expects for minimal residual disease (MRD) coverage decisions. Importantly, the barriers—or “moats”—to entry are broader and deeper than they first appear. Patient sample size (e.g., 55, 79, or 101) is not the sole determinant of success.

---

1. Coverage Criteria from the LCD

MolDx’s local coverage determination (LCD) outlines three main use cases:

1. Postoperative clearance – e.g., testing for circulating tumor DNA six weeks after resection of breast or colon cancer.

2. Surveillance in MRD-negative patients – serial testing (e.g., quarterly) to monitor for recurrence, with recurrence rates in small resected cancers often around 10% per year.

3. Therapy response monitoring – repeated testing during treatment, particularly in immuno-oncology; declining DNA levels indicate response, rising levels suggest resistance.

The LCD requires that the MRD test detect recurrence earlier than standard-of-care methods, with comparable or superior accuracy. “Earlier” is not defined as a few days, but rather on the scale of 2–4 months ahead of conventional imaging.

---

2. Colon Cancer Screening as a Mental Model

A useful comparison is colon cancer screening trials. These require ~10,000 patients because cancer prevalence is ~1%. To evaluate sensitivity and specificity with narrow confidence intervals, FDA generally requires ~100 cancer cases. With fewer cases, confidence intervals become too wide (e.g., 80% sensitivity ±15% vs. ±2%).

The same principle applies to MRD: the number of relapse cases, not total enrollment, drives statistical precision. For example, with a 10% relapse rate, a study of 200 MRD patients may yield only 20 relapse events—just enough to anchor meaningful performance statistics.

---

3. Data Elements for MRD Success

A credible MRD study design typically requires:

• Adequate relapsed patients: ~15 as a minimum, ideally more. With a 10% annual relapse rate, 150 patients are needed to achieve 15 events in year one.

• Reasonable error bounds: sufficient relapse cases are necessary to narrow sensitivity and specificity estimates.

• Frequent blood sampling: ideally 5–6 blood draws per patient over the study period.

• Frequent gold-standard imaging: 4–5 CT scans (or equivalent) to enable valid comparisons.

Without this frequency, the statistics collapse. For example, one blood draw and one CT scan a year later cannot support meaningful lead-time estimates. This is why results are often reported as “125 patients with 550 blood draws”—highlighting the density of longitudinal data as much as cohort size.

---

4. Why the “Moat” is Wider Than It Appears

Two common approaches illustrate the barriers:

• Prospective trials: require patient enrollment, repeated blood draws and imaging over two years, high retention, and then an additional year for publication and MolDx review.

• Use of biobanked specimens: depends on what the archive actually contains (number of patients, matched plasma, and contemporaneous CT scans). Missing elements can cripple statistical validity.

Thus, both pathways demand substantial time, resources, and infrastructure—much more than a simple headcount might suggest.

---

5. Examples

• Saga Diagnostics (PathLight): Recently, MolDx approved a breast cancer MRD assay from Saga Diagnostics (University of Lund). See Elliott, 2025 for a concise summary of their dataset.

• Natera: In its ASCO announcement, Natera highlighted development of a genome-based MRD test supported by ~500 patients and ~3,000 blood samples. The large number of serial blood draws directly strengthens the precision of their statistical claims.

---

In summary: Success depends not only on total patient numbers but also on relapse-event counts, sampling frequency, and the robustness of standard-of-care comparisons. Missing any of these dimensions risks undermining the statistical case for coverage.

---

Would you like me to also condense this into a one-paragraph “payor-facing” version (like an executive summary) so you could drop it into decks or client communications?