A/B'ing Trust Signals: How MediPrompt Optimized Clinical AI Prompts for Patient Safety

AI · 7 min read

MediPrompt integrates generative models into clinical workflows to suggest documentation text and coding recommendations. Early rollouts saw clinicians either over-relying on suggestions or reflexively rejecting them due to safety concerns. The product team framed the problem as a trust calibration issue: clinicians needed signals to judge when the AI was helpful and when to be cautious.

The team tested variations of trust signals: model confidence bands, provenance badges showing data sources, explainable snippets that highlighted why a suggestion matched the chart, and “suggestion lock” actions that required a second confirmation for high-risk content. Designs were vetted in shadow-mode before being surfaced in live workflows. Privacy and compliance demanded that all provenance data be generated without exposing PHI to third-party services and that audit logs be immutable.

A controlled rollout across three hospital systems measured clinician behavior: the version with concise explainable snippets plus a confidence band reduced high-risk overrides by 42% and increased accepted-suggestion accuracy by 16%. The “suggestion lock” pattern decreased erroneous sign-offs but introduced a modest 6% increase in documentation time; clinicians preferred the balanced explainability approach as it preserved speed while improving safety.

MediPrompt’s experiments highlight that in regulated domains, UX is safety engineering. Designers must treat trust signals as first-class features and measure both clinical accuracy and workflow impact. The winning pattern combined clear, brief explanations with immutable audit trails—supporting clinician decision-making without creating unnecessary friction.