From principles to measurement: a risk-proportionate, evidence-weighted framework for assessing bias and explainability readiness of healthcare AI in the United Kingdom

Abstract Ethical guidance for healthcare artificial intelligence is abundant, but it is expressed as principles rather than as measurements. Organisations are therefore left to judge whether a given system is ready for deployment without an instrument that states what is being measured, how responses are combined, and how much confidence the resulting judgement warrants. Existing self-assessment tools inherit three measurement defects that are rarely made explicit: they treat unevidenced assertions as equivalent to audited evidence, they weight ethical dimensions equally irrespective of how the system will be used, and they aggregate additively, so that strength in one area can silently offset a disqualifying weakness in another. This paper specifies an ethical readiness assessment framework for United Kingdom healthcare AI that addresses these three defects directly. The framework retains a twenty-item instrument covering Bias and Fairness and Transparency and Explainability, and adds an explicit measurement model: each response is discounted by the tier of evidence supporting it; dimension weights are derived from the system’s intended purpose using the two axes of the international risk-categorisation framework that the national regulator has committed to align with; responses are combined by a weighted power mean whose compensability is a declared parameter rather than an unstated assumption; and four items anchored to United Kingdom statutory duties act as gates that cap the reported band when the corresponding safeguard is absent. Every item is mapped to the specific legal duty or assurance standard it evidences. The framework is demonstrated on three constructed, fully specified system profiles, with all item-level responses published so that every reported figure is reproducible. The demonstration shows that a system with strong documentation but weak fairness safeguards is classified as Moderate readiness under equal-weight additive scoring and as Low readiness under the specified model, with the two new mechanisms catching it independently. A Monte Carlo study over response and weight uncertainty (20,000 replicates) reproduces the nominal band in 84 and 100 per cent of replicates for the two profiles whose scores lie away from a band boundary, and in 82 per cent for the profile sitting half a percentage point below one, which is reported rather than concealed. Sensitivity analyses over the aggregation parameter, over departures from the equal-interval assumption, over all twenty-seven codings of the intended-use risk profile, and over the full range of the two stipulated parameters leave the deficient system’s classification unchanged in every case, while showing that the mature system’s band turns on whichever of those choices is moved. The framework is presented as a governance-support instrument with a stated measurement model and a stated uncertainty, not as an assurance or certification mechanism, and its psychometric validation with United Kingdom stakeholders remains outstanding.