Medical device simulation: choosing the next physical test
Medical device simulation can help engineers reject designs before building prototypes. That benefit depends on evidence that the model answers the specific engineering question. Generating more CAD variants does not supply that evidence, and a structural model cannot answer whether a user misunderstands a control.
In the fictional worksheet below, a team considers 12 cable-tray designs. A deliberately simple screening policy rejects two. Ten still need a physical candidate screen. Those counts describe invented assumptions, not measured savings or an FDA-accepted test plan.
Download the 12-candidate worksheet (CSV). Every numerical value in the file is fictional. The example is for planning nonclinical engineering work, not for choosing a real material, setting a safety limit or releasing a device.
FDA's modeling guidance has a defined scope
FDA's final Assessing the Credibility of Computational Modeling and Simulation in Medical Device Submissions, issued November 16, 2023, covers physics-based, mechanistic and other first-principles models. It does not cover standalone statistical or data-driven models such as machine learning. Its recommendations are nonbinding.
An AI system that proposes geometry and a mechanics solver that evaluates it do different jobs. A credible mechanics model does not validate the geometry generator. Nor does this guidance grant blanket permission to replace clinical evidence with synthetic patients. FDA describes possible uses of modeling, including in silico device testing, but asks sponsors to justify credibility for the proposed context of use.
The MDIC CM&S Reference Hub, updated July 13, 2026, makes the same scope distinction. It is a resource index for physics-based or mechanistic modeling, not an AI validation standard or regulatory approval.
A fictional cable-tray candidate set
The component is an external cable-management tray on an equipment cart. It has no patient contact and makes no diagnostic or treatment decision. Its role here is to illustrate early candidate selection. The example does not determine whether a real tray would be a regulated accessory.
The team combines three rib patterns, two wall options and two material families. That produces 12 candidates. These are categorical labels, not buildable specifications. The worksheet gives no dimensions, loads or material properties.
The team's hypothetical constraints are:
- Keep the tray within a fixed mounting envelope and clear of the cart's controls. Assume every candidate passes this initial geometry check.
- Limit deflection under one defined loading condition. For arithmetic only, express deflection relative to an invented target of 1.0.
- Keep cables accessible without obstructing cleaning or the user's view of the cart. A structural calculation cannot establish these conditions.
- Use a manufacturable design. Molding defects and assembly variation remain outside the toy calculation.
One material family is assumed to lie within the model's supported domain. The other is outside it. A real team would need evidence for that distinction: material characterization, relevant comparison tests and a documented applicability assessment. The worksheet assumes that work exists for the first family. It does not supply it.
The screen uses bounds, not a nominal score alone
For the six in-domain candidates, the worksheet assigns a fictional normalized deflection and an illustrative allowance of 0.2. Lower and upper bounds equal the score minus or plus that allowance. The allowance is not a measured error bound, a confidence interval or a known model accuracy.
The toy policy rejects a candidate only when its lower bound exceeds 1.0. A candidate whose upper bound is at most 1.0 stays on the physical-screen list. A candidate whose interval crosses the target also stays, marked borderline. The out-of-domain family receives no numerical prediction or simulation-based rejection.
| Candidate | Rib pattern / wall | Assumed score | Illustrative interval | Next step in the toy policy |
|---|---|---|---|---|
| CT-01 | Sparse / thin | 1.5 | 1.3 to 1.7 | Reject before prototype screen |
| CT-02 | Sparse / thick | 1.3 | 1.1 to 1.5 | Reject before prototype screen |
| CT-03 | Parallel / thin | 1.1 | 0.9 to 1.3 | Physical screen; borderline |
| CT-04 | Parallel / thick | 0.8 | 0.6 to 1.0 | Retain for physical screen |
| CT-05 | Cross / thin | 0.9 | 0.7 to 1.1 | Physical screen; borderline |
| CT-06 | Cross / thick | 0.6 | 0.4 to 0.8 | Retain for physical screen |
| CT-07 through CT-12 | Same six combinations; other material | Not assessed | No justified bounds | Characterize material and review applicability; physical screen |
CT-03 illustrates a premature rejection. Its nominal score exceeds the target, but its assumed interval straddles it. CT-04 illustrates a different mistake: treating a favorable screen as permission to release. Under this policy, CT-04 still goes to physical testing. The invented target and allowance are not clinical or engineering acceptance criteria.
The script analyze-medical-device-testing.py generates the CSV using decimal arithmetic. Run python3 scripts/analyze-medical-device-testing.py --check to verify the committed file. The script enumerates assumptions; it does not run a finite-element solver or analyze experimental data.
Count the testing work that remains
Without the screen, the hypothetical team would run one initial physical deflection screen for each of 12 candidates. With this policy, it runs ten. That is two fewer candidate screens out of 12, not two fewer regulatory tests. Each screen could require several specimens, conditions and repeats. None of those quantities is specified here.
The count excludes creating the model, checking the solver setup, characterizing materials, collecting comparison data and reviewing discrepancies. It also excludes later verification on the chosen design. The example therefore supports no claim about net cost, calendar time or return on investment. If credibility work costs more than the avoided screens, this small candidate set may be a poor reason to build the model.
More generated designs can increase the remaining work. If a generator adds six candidates using a third unsupported material, this policy would add six physical screens unless the team first justifies extending the model's domain. The existing numerical scores provide no basis for discarding those designs.
Evidence needed before relying on a real screen
FDA's credibility guidance starts with a question of interest and the model's context of use. For this example, the narrow question is which tray candidates merit a prototype deflection screen under a specified loading condition. It excludes fatigue, cleaning, user interaction and clinical performance.
The guidance assesses model risk through model influence and decision consequence. A model used as one input to an early prototype choice has a different role than a model used as the main evidence supporting a device safety claim. The team must assess the actual use and consequences, rather than assign a risk level because the solver is familiar.
For a real structural screen, the credibility record would need to explain the following evidence, as appropriate to that use:
| Evidence question | Record to examine | Finding that would stop reliance on the screen |
|---|---|---|
| Does the implementation solve the intended equations? | Code verification evidence and software version | Unsupported implementation or unresolved numerical defect |
| Is the specific calculation sufficiently resolved? | Mesh or time-step study, convergence checks and numerical error assessment | Candidate ordering changes with unresolved numerical choices |
| Do predictions agree with relevant physical data? | Model validation comparisons and measurement uncertainty | Discrepancy exceeds the justified allowance |
| Does the evidence cover these candidates and conditions? | Material range, geometry range, fixtures, boundary conditions and applicability review | New material or loading condition falls outside the supported domain |
| Could uncertain inputs reverse the decision? | Sensitivity and uncertainty analysis | A candidate crosses the decision boundary under plausible inputs |
These are evidence questions, not a prescribed validation protocol. FDA's guidance distinguishes calibration evidence from validation against data independent of model development. Reusing the same observations to tune a parameter and then claim independent predictive accuracy would misdescribe the evidence.
Physical tests answer different questions
A prototype deflection screen helps compare candidates under a defined fixture and load. Later design verification asks whether the selected design meets its specified inputs. A candidate that looks favorable in the toy table could still fail that verification because of an incorrect boundary condition or an unmodeled failure mode.
Human factors validation asks whether intended users can use the device safely and effectively under expected use conditions. FDA's final Applying Human Factors and Usability Engineering to Medical Devices, issued August 3, 2026 and originally issued in 2016, describes that work. A deflection score cannot establish that a tray leaves controls visible or that users can route cables without a dangerous mistake. Early formative evaluation can identify such problems before the team commits to a design.
Manufacturing evidence has another purpose. A favorable analysis of nominal geometry does not establish that a production process consistently makes acceptable parts. Process validation, where applicable, concerns the production process. It is separate from model validation and design validation. FDA's QMSR overview confirms that the current quality-system regulation took effect February 2, 2026 and incorporates ISO 13485:2016.
Finally, validating a physics model does not validate an AI function embedded in a medical device. That would require evidence suited to the function, intended use and applicable regulatory pathway. This cable-tray example contains no such function and supplies no clinical evidence.
Preserve rejected candidates and unresolved questions
The CSV retains rejected rows and identifies out-of-domain candidates. A working engineering record would also preserve each geometry version, solver configuration, source material data, comparison-test report and reviewer decision. A later material change could invalidate an earlier rejection or ranking.
The design-change worked example follows a change through requirements and risk records. This worksheet covers an earlier decision: choosing candidates for the next experiment. For use-related evidence, see human factors engineering and IEC 62366-1.
Arca Libraries let a team ask questions of uploaded documents and open cited source passages. That can help a reviewer locate the report supporting a candidate decision. It does not establish model credibility, run the mechanical tests or approve a design.
Scope and limitations
Sources were checked November 16, 2026. No engineer or clinician has validated this fictional example. Its candidate scores, domain labels and screening policy are invented for reproducible arithmetic. A real team must define device-specific requirements and justify its evidence plan with qualified reviewers. The worksheet cannot establish conformity, safety, effectiveness or regulatory acceptance.