Designing Trustworthy AI Visualizations in High-Stakes Clinical Decision-Making: A Human-Centered Study of Visualization, Infrastructure, and Workflow

AI systems are increasingly deployed in health settings to assist with diagnosis, screening, triage, and physiological monitoring, yet little attention has been paid to how their outputs are visually communicated to users. This paper presents a qualitative, interface-centered comparative analysis of six publicly accessible AI health tools - SkinVision, Infermedica, Aysa, Qure.ai, Anura, and PMcardio evaluated through a Trust and Infrastructure Coding Frame derived from the Microsoft HAX guidelines and grounded in three user personas. Four cross-cutting themes emerge: visual clarity creates an illusion of certainty, explanation is consistently weaker than presentation, different modalities produce different trust dynamics, and infrastructure signals are routinely hidden. The paper introduces the High-Stakes AI Visualization Framework, a set of six design principles, and demonstrates its application through critical redesigns of three failure cases. The framework argues that trust calibration, not trust maximization, should guide health AI interface design.