In May 2026, Dexcom recalled versions 2.9.0 to 2.11.2 of its G7 continuous glucose monitor iOS app in a Class II action. The defect was not a wrong calculation. It was a scheduling problem, and it is exactly the kind of thing medical device software testing tends to miss. A low-priority file input/output check was allowed to block the higher-priority thread that processes estimated glucose values. Notifications queued and replayed, and the display flickered through stale readings on a device people trust with insulin decisions. You can read the recall in the FDA CDRH database at https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfres/res.cfm and a good technical write-up at https://www.softwarecpr.com/2026/05/fda-recall-software-defect-responsible-for-several-errors/ .
The logic was fine. The glucose math was fine. What failed was the order in which the operating system got to run the code, and no amount of checking the arithmetic would have found it.
What actually broke
This is a textbook priority inversion. A cheap, unimportant task grabbed a resource or a scheduling slot and starved the task that mattered. On a quiet phone with nothing else going on, it never surfaces. On a busy phone, mid-day, with other apps competing for the same file system and the same CPU, the low-priority file check gets in the way of the thread that turns sensor data into a number and an alert. The reading you see is a few minutes old. The alert you needed fires late, or replays after it is no longer true.
Sit with why this passed testing, because the reason is the whole lesson. Functional testing asks: given this input, do I get the right output? Feed the app a glucose value, confirm it displays the value, confirm the alert threshold triggers. Green. Every functional case can pass and the bug still ships, because the bug does not live in the input-output mapping. It lives in timing, in thread scheduling, in what happens when two pieces of code want the same thing at the same moment.
Medical device software testing is two things different
Standard testing craft still applies here. The equivalence classes, the boundary values, the traceability from requirement to test, all of it carries over. But medical device software testing adds two things on top of ordinary software testing, and this recall sits squarely inside both.
- Documentation and auditability. Reproducibility and traceability stop being nice-to-haves. You write your test cases in a consistent way, you review them, you keep records that an auditor can follow from a hazard to the specific test that exercises it. Under the FDA QMSR, in effect since February 2, 2026 (21 CFR 820 now incorporates ISO 13485:2016 by reference), those records are under more scrutiny, not less.
- Risk management. This is the part that would have caught the Dexcom defect. Delayed or stale glucose readings are a named hazard. Once you have named it, the question is not "does the app usually show the right number." The question is "can the sequence that produces a stale reading happen at all, under load," and you have to design tests that try to make it happen.
Notice that the second one is not a documentation exercise. It is a test-design decision. It changes what you go looking for.
Testing a risk control for effectiveness, not presence
Here is the move most V&V plans skip. When you add a risk control, whether it is a watchdog, a priority guarantee, or a timeout, the point of testing it is not to confirm the control exists in the code. The point is to prove the hazardous sequence can no longer occur.
Those are different tests. Confirming the control exists is a functional check: the watchdog is present, it compiles, it fires in a demo. Proving effectiveness means you recreate the conditions that caused the hazard, a loaded device, contended file access, competing threads, and you show that the stale-reading sequence does not reproduce. If you cannot make it happen under the worst conditions you can construct, the control works. If you never tried to make it happen, you tested nothing.
The Boston Scientific pacemaker recall earlier in 2026 is the grim version of the same lesson. A software and battery defect, addressed by a software update, was tied to deaths and injuries as reported in the recall roundups. When a risk control stands between a patient and a hazard, "we checked that the control is there" is not evidence it works. Proving the hazard no longer occurs is.
This is a pattern, not a one-off
It would be comfortable to file the Dexcom recall as bad luck. The trend says otherwise. One 2026 analysis of recall data (MedDeviceGuide, citing FDA figures) puts software defects at 8.2% of recall events and growing 31% year over year, as more of what a device does moves into software. The more the device is software, the more of its failures are software failures, and a rising share of those are the timing and concurrency defects that functional testing does not see.
The standards are moving to match. IEC 62304 Edition 2 is expected to publish in August 2026, reworking the software safety classes and mapping the software process to risk management, cybersecurity, and AI. The old Class A/B/C ambiguity that let teams argue their way into less rigor is the thing being tightened. If your V&V plan only exercises functional correctness today, Edition 2 is not going to make that easier to defend.
Where Nopmark fits
Nopmark's Medical Device Bundle is built for exactly this gap: software validation under IEC 62304, risk management under ISO 14971, and the V&V thinking that ties a hazard to the test that proves it cannot recur. If your team writes test plans that pass functional cases and still ships scheduling bugs, this is the training that moves your effort to where the real defects hide. It is taught by accredited instructors with hands-on labs, not slides about compliance in the abstract.
FAQ
Why did functional testing pass when the app had a serious defect?
Because the defect was not in the input-output logic. The glucose math was correct. The failure was in thread scheduling under load, which functional cases do not exercise. You have to design tests that create contention on purpose.
What is concurrency testing in medical device software testing?
It is testing that deliberately exercises timing, thread priority, and shared-resource contention: running the software on a loaded device, forcing competing threads, and checking that a high-priority safety task is never starved by a low-priority one.
How do I test a risk control for effectiveness?
Recreate the conditions that produce the hazard and show the hazardous sequence does not occur. Confirming the control is present in the code is not enough. You have to try to make the failure happen and fail to reproduce it.
Does IEC 62304 Edition 2 change how I test?
Edition 2 is expected in August 2026 and reworks the software safety classes and their mapping to risk, cybersecurity, and AI. It is best read as tightening the rigor you can justify, not as a new set of test techniques. Watch for the published text before you rewrite procedures.
Is this only a concern for life-critical devices?
The stakes are highest there, but the testing gap is general. Any software with concurrent tasks and a task that must not be delayed has the same exposure. Life-critical devices just make the cost of missing it unforgivable.
Ready to write the V&V plan that catches the bug that only shows up on a busy device? The Medical Device Bundle from Nopmark covers software validation, risk management, and effectiveness testing, taught by accredited instructors with hands-on labs, and you can see the full bundle at https://www.nopmark.com/bundles/medical-device-bundle