Biological validation, human factors and responsible translation

Verification and validation ask different questions

Verification checks whether the chosen specification or mathematical model was implemented correctly, including numerical errors where relevant. Validation evaluates whether the representation or system is adequate for its intended real-world use. Correct arithmetic cannot establish that a biological model contains the mechanisms needed for a particular prediction.

Plan evidence around a defined question, output, population and conditions. Bench mechanics, biological response, manufacturing consistency, software performance and usability may each contribute. Independent validation cases should be distinct from data used to choose model parameters. A change in material, processing, software or intended use can change which earlier evidence remains relevant.

A constructed validation comparison

An imaginary response model predicts 10 arbitrary response units for each of four donor-derived systems. Their independently supplied reference means are 8, 10, 12 and 14. Define residual as prediction minus reference: +2, 0, −2 and −4. The mean residual is −1 unit, but that average hides errors in both directions and larger underprediction for the last case.

Mean absolute error is (2+0+2+4)/4=2 units. The figure retains each paired result. Suppose each reference mean came from five technical readings of one sample: there are twenty readings but only four donor-derived systems in this comparison. If each donor was processed in a separate batch, donor and batch effects cannot be separated from these data alone.

Usability and ethics are part of the evidence

People perceive information, interpret it, act and receive feedback. The interface includes controls, connections, displays, packaging, instructions and maintenance interactions. Study the intended users and environments, including foreseeable difficulties and consequences of use error. A correct sensor output is of little use if a display obscures its units or makes a critical state easy to misinterpret.

Responsible translation also considers informed voluntary participation, protection from harm, fair distribution of burdens and potential benefits, privacy and access. The Belmont Report provides a named ethical framework of respect for persons, beneficence and justice. It is not a complete statement of every jurisdiction’s current requirements, and no prototype result grants permission to expose people to an unreviewed intervention.

Follow the selected record all the way to export

Consider an invented, nonclinical viewer for the four donor-model comparisons above. Assign viewer records R1–R4 and donor labels D1–D4 to those cases in their original order. Its limited claim is that a trained bench reviewer can identify a requested donor record and export that same record with its prediction, reference, units and residual convention. This is a person–interface–software claim. The correct residual formula alone does not establish it.

For a positive walkthrough, request record R4, donor D4, with prediction 10 and reference 14 arbitrary response units. A proposed display explicitly shows those identities, both values, “prediction minus reference” and residual −4, together with calculation version V1. The reviewer selects R4, sees that identity retained in the export preview, and obtains the same labelled pair. This constructed sequence specifies the observation the claim needs; it is not a report that real users succeeded. The earlier four-case error statistics still describe model agreement, while this trace checks the record attached to that agreement.

Now suppose interface version V2 sorts the table after selection but leaves a highlight on the same screen position. That position can now represent R3, donor D3, whose reference is 12 and residual −2. Exporting the highlighted row could produce mathematically correct values for the wrong requested donor. The causal gap is between the user’s intended identity, the visible selection and the export state. A persistent record label and matching preview are proposed controls whose usefulness still requires observations of intended users doing that task in the relevant configuration.

Keep evidence boundaries specific. A deterministic export check can verify matching record identifiers for supplied inputs. Representative task observations must also show whether users notice and interpret the identity and residual labels, including likely interruptions or mistakes. Expert use of V1 on a wide display does not establish novice use of V2 on a crowded display. Changing a production batch for the biological reference raises a different question about model evidence; it cannot be resolved by a clearer selection highlight.

Explain what this small comparison can support

The constructed residuals have squared-error mean (4+0+4+16)/4=6 units² and root-mean-square error √6≈2.45 units. Mean bias −1, absolute error 2 and root-mean-square error 2.45 describe different aspects of these same four cases. None is a pre-established acceptance criterion or a confidence bound for future users.

A useful next evidence plan would specify the acceptable prediction error for the intended decision, relevant biological variation, batch coverage and independent comparisons. It would also address critical user tasks and the remaining ethical and oversight questions. Preserve unsuccessful results, deviations and versions so that the scope of the conclusion can be audited rather than enlarged by selective reporting.

In the changed-selection example, an export containing R3’s residual −2 instead of requested R4’s −4 does not add a new model residual to the four-case dataset. It is a task/identity mismatch. Keep the original prediction-error calculation separate from the interaction record and its interface version so that the remedy addresses the observed failure.

Try the reasoning

1. Does verification of a simulation establish biological validity?

   Show the answer and reasoning

   No. It checks implementation; adequacy for a biological prediction needs relevant real-world evidence and an explicit context of use.

2. Why are the twenty technical readings not twenty donor systems?

   Show the answer and reasoning

   The readings are nested within four biological identities; repeating a measurement does not create additional biological independence.

3. Can a small mean residual establish acceptable individual errors?

   Show the answer and reasoning

   No. Positive and negative errors can cancel, and acceptability depends on the intended decision, distribution of errors and evidence coverage.

4. A scripted V2 export test preserves IDs, and experienced reviewers complete the task on a large display. The proposed users are now novices on a smaller display with the same calculation code. Which evidence transfers, and what remains untested?

   Show the answer and reasoning

   The verified identifier mapping and unchanged calculation can remain relevant within their tested inputs/configuration. They do not show that novices can find, interpret and retain the intended record on the smaller display. Task observations with those users and that presentation must address selection, identity confirmation, units/convention and the export outcome. This is distinct from adding independent donor/batch comparisons for biological prediction validity.

What to carry forward

Tie each claim to relevant independent evidence, preserve biological and user variation, and make uncertainty and ethical responsibility visible before translation.