Healthcare AI evaluations often report predictive performance and changes in outcomes following deployment. These are useful measures, but they answer different questions.
Predictive performance describes how well a model estimates an outcome. Post-deployment outcome changes describe what happened after implementation. Neither, on its own, establishes that the AI system caused the change.
Consider a hospital that introduces AI-assisted triage and subsequently observes shorter waiting times. Staffing levels may also have increased, workflows may have changed, or patient volume may have declined.
A simple before-and-after comparison cannot reliably isolate the AI contribution.
Impact evaluation therefore requires a design suited to the intervention and clinical setting. Depending on feasibility, this may involve randomized trials, controlled implementation studies, interrupted time-series analysis or carefully designed observational comparisons.
Evaluation should also distinguish between the effect of making predictions available and the effect of acting on them. A predictive model may have little impact if clinicians cannot act on its outputs. Conversely, an agentic system may change outcomes through the actions it initiates, creating additional pathways to assess.
Researchers should consider confounding, selection effects, implementation fidelity, unintended consequences and differences across patient groups.
For agentic systems, the evaluation should examine the complete intervention pathway, from the model's output through execution and clinical response to relevant outcomes.
The core principle is simple: temporal association is not causal evidence.
Healthcare AI needs credible evidence that deployment improves care compared with an appropriate alternative, not merely evidence that favorable outcomes occurred afterward.
Top comments (0)