Frameworks

AI Readiness Framework

A structured instrument for one question: when is a clinical prediction model ready for deployment? The framework organizes that judgment into six gated questions, assessed in order, so that fundamental failures are surfaced before secondary ones.

GitHub Zenodo DOI Published in Viruses 2026 (Supplementary S3)

Six gated questions

Each question gates the next: a model that fails an earlier gate does not earn a readiness verdict until the failure is addressed.

1

Data Availability

Is sufficient high-quality data available for the intended use?

2

External Validity

Does the model generalize, or does precision create false confidence?

3

Explainability

Can clinicians trust the parameters and their behavior?

4

Interpretability

Does transparency enable meaningful oversight?

5

Equity

Do aggregate averages mask subgroup disparities and harms?

6

Readiness

Given the above, should implementation proceed — and under what monitoring?

Design principles

Gated, not additive

Fundamental failures are surfaced first. Readiness is not an average of independently scored dimensions.

Transparent & contestable

Scoring is explicit and open to challenge; a verdict states what would change it.

Reporting-aligned

Assessment is aligned with TRIPOD+AI reporting expectations for prediction-model studies.

Readiness ≠ proven

Rigorous validation exposes uncertainty rather than eliminating it. Computationally ready is not the same as clinically proven.

Equity as a gate

Subgroup harm is treated as a first-class failure mode, not a footnote to aggregate performance.

Context-bound

Readiness is judged for a specific deployment context, not asserted in the abstract.