Use remnant samples sparingly.

Remnant samples are tempting for the same reasons they are dangerous. They are cheap; they are fast; they are scientifically real human samples; they are already in a freezer somewhere; and the analytics teams have learned to extract enough metadata from claims and EHR adjacencies that, on a slide, the remnant cohort looks indistinguishable from a prospectively collected one. Sponsors with tight timelines and tighter budgets ask us about remnants on every other engagement. Sometimes we recommend them. More often we do not, and the reasons are worth being explicit about, because the conventional case for remnants overstates what the data they come with can carry.

The first issue is what the remnant is. It is a sample drawn from a patient who was being seen for clinical reasons, processed by a clinical lab according to that lab's procedures, and stored according to that lab's protocols — none of which were designed around the IVD validation it is now being asked to support. The collection conditions were not the conditions specified in the IFU of the assay being validated. The processing latencies were whatever the clinical workflow produced that day, not the windows the validation needs to defend. The storage temperature, the freeze-thaw history, the time on the bench between collection and centrifugation — none of these are documented in the kind of detail an IVD validation requires.

This matters because the validation has to defend the assay's performance in the conditions of its intended use. An assay whose IFU specifies serum collected in a tiger-top tube, processed within two hours, stored at minus eighty degrees Celsius, with no more than one freeze-thaw cycle, has to demonstrate performance under those conditions. A remnant cohort cannot demonstrate that, because the remnant collection did not adhere to those conditions and there is no record that proves whether they were met or violated. Performance on remnants tells you what the assay does in a population of samples whose collection conditions were unknown — which is not the question the validation is required to answer.

The second issue is what the remnant does not have. The conventional response to the documentation gap is to mine ICD codes, claims data, and EHR adjacencies to reconstruct the clinical context of each remnant. The reconstructions are sophisticated. They are also systematically wrong on the dimensions that matter most for an IVD claim. ICD codes are billing artifacts, not diagnostic ones. A patient billed under an ICD code for a condition may have a presumptive diagnosis, a confirmed diagnosis, or a working diagnosis that was later ruled out. The code does not distinguish. EHR adjacencies fare somewhat better but still produce systematic confounds: the remnant's clinical context reflects the population a clinical lab serves, which is not the population the assay's intended-use claim references.

Specifically, remnants are skewed toward patients sick enough to have ordered the test. The standard-of-care patient population most IVD claims target — the patient about whom a clinician will use the assay to inform a real decision — is not the same population whose remnant samples are sitting in the freezer. The remnant population is enriched for severe presentations, for patients with multiple comorbidities, for patients whose clinical complexity drove additional testing. An assay validated on remnants demonstrates performance in that enriched population. It does not demonstrate performance in the standard-of-care population the IFU is going to claim.

The third issue is harder to name and more important. The kind of validation the FDA is willing to accept depends on what the agency thinks the validation is asking it to believe. An assay whose validation rests on remnants is asking the agency to extrapolate from the documented remnant population to the undocumented intended-use population, on the basis of analytics the agency has not validated. Some claims survive that ask. Most do not. The pre-submission discussion that anchors a validation strategy is the right place to discover which is which — and we have found, repeatedly, that the agency's tolerance for remnant-anchored validation is lower than sponsors hope and that the discovery happens earlier when the question is asked early.

None of this is an argument against remnants in all cases. For analytical validation work — precision, linearity, interfering substances, lot-to-lot reproducibility — remnants are often appropriate, and the documentation gap is less binding because the analytical claim does not depend on standard-of-care representativeness. For early signal-finding work, before a validation strategy has been chosen, remnants can productively narrow the candidate-claim space. The argument is against using remnants as the pivotal evidence for a clinical performance claim, which is the use case sponsors most often propose and for which they are most often wrong.

What this means in practice. When a sponsor asks us about remnants for a clinical-performance validation, we ask three questions back. Does the assay's IFU specify collection conditions that the remnant cohort can document? Does the intended-use population match the population whose remnants are available, or is the match an extrapolation supported by ICD code mining? And has the agency given pre-submission feedback on the use of remnants for this specific claim? If any of the three answers is uncertain, the remnant shortcut is not a shortcut — it is a longer path made of cheaper data.

The right default is prospective collection for clinical-performance claims, with remnants reserved for analytical validation and early signal-finding. The default is more expensive at the validation stage. It is materially less expensive across the full pivotal-and-submission lifecycle, because the validation it produces is the one the agency can evaluate.