Every healthcare AI dashboard we inherit from a previous vendor has the same problem: too many metrics, no thresholds, and no narrative. The clinical lead opens it twice and then never again.
The dashboards we build have three numbers on the front page. Everything else lives one click deep. The three numbers, in order, are:
1. Acceptance rate
What fraction of suggestions are taken without modification. This is your top-of-funnel signal. If acceptance is below 30% in week three, the model is wrong or the surface is wrong; either way, you don’t scale.
2. Modification rate
What fraction of suggestions are taken with edits. This is the most underrated metric in clinical AI. A high modification rate is good news — it means clinicians are using the system as a starting point, and the diff between suggestion and final tells you exactly how to improve the model.
We instrument this at the field level. “Modified the medication” is a different signal than “modified the dose.”
3. Override rate, by reason code
What fraction of suggestions are rejected, and crucially — why. The reason codes are the only signal that tells you whether to retrain, retune the prompt, or pull the integration entirely.
The four reason codes we ship by default:
- Wrong for this patient. Population mismatch — usually a retraining signal.
- Already addressed. Latency or context-window issue — usually an integration fix.
- Disagree, see note. Clinical judgment edge — usually a prompt or threshold tuning.
- Not applicable. Workflow mismatch — usually a routing fix, sometimes a sunset signal.
What we leave off
Latency, throughput, model version, deployment region. All of it matters; none of it goes on the page the clinical lead opens. A dashboard that tries to inform two audiences usually informs neither.