AI-assisted life sciences evidence work: define the context before judging a model

Name the decision and model role

Start by writing one sentence that connects the model's output to a human decision. “The model predicts a value” is incomplete. Who receives the value, what action might follow, what other evidence is available, and what happens when the model is wrong? In this exercise, a fictional research team is deciding whether a model is sufficiently characterized for a further controlled study. The model does not diagnose, select treatment, release a batch or establish safety or effectiveness.

The US Food and Drug Administration's January 2025 draft guidance on AI used to support regulatory decision-making describes a risk-based credibility assessment framework tied to a particular context of use. The page identifies the document as draft guidance, not for implementation, with non-binding recommendations. The European Medicines Agency's reflection paper on AI in the medicinal-product lifecycle discusses principles across the lifecycle. These primary sources inform the distinction between model performance and fitness for a defined use. They do not approve the fictional model or prescribe one universal evidence package.

Sources: [1] [2]

Build a context-of-use evidence card

Imagine a synthetic model that flags microscope images for a researcher's second review. The development summary reports strong aggregate performance on images from Laboratory A. The proposed pilot will use images from Laboratory B, which has a different scanner and staining workflow. A slide labels the model “externally validated,” but the attached table contains no Laboratory B data. A later email mentions that low-quality images were removed before evaluation, without defining the exclusion rule.

Create a card with fields for intended use, user, decision influenced, input provenance, development population, evaluation population, operating conditions, performance measures, failure costs, human review, exclusions, known gaps and evidence owner. Each field links to an exact source. The card records the reported Laboratory A results without converting them into a claim about Laboratory B. It flags the scanner and staining shift, asks how excluded images would be handled in practice, and identifies who must decide whether a bridging study is needed.

AI may help extract candidate facts from the approved synthetic pack. The worker checks each item against the original table, protocol and email. If the model summary says “works across laboratories,” reject that wording because only one laboratory appears in the evidence. If an average metric hides a subgroup with sparse examples, keep the denominator and uncertainty visible. If the evaluation set overlaps with development material, record the possible leakage rather than treating the headline metric as independent confirmation.

Test the card against a changed use

Run an adversarial review by changing only the downstream decision. First, the model merely prioritizes images for a second look, with every image still reviewed. Next, someone proposes using the same output to omit low-priority images from review. The model and performance table are unchanged, but the consequence of error and necessary evidence are not. The card should show that the second use is materially different and cannot inherit the first use's assessment automatically.

Then add five poor-quality images that the original evaluation excluded. A strong worker asks whether these inputs should be rejected, routed to manual review or studied separately. They do not fill the gap with a plausible AI explanation. The final internal note states which evidence belongs to the defined use, which facts remain uncertain, and which owner must authorize any next study.

Preserve scientific, regulatory and assessment boundaries

Real work must follow applicable research, quality, privacy, clinical and regulatory requirements with qualified specialists. This exercise does not establish model credibility for an actual regulated purpose and does not assess patient care. Two Prune could use a synthetic case to observe whether a participant defines the context of use, detects dataset shift and keeps exclusions visible. Its report would describe task-bounded assessment evidence for human interpretation, not certify a model or automate a scientific, regulatory, clinical or employment decision.