Uncertainty Communication Design in AI-Augmented Clinical Risk Stratification Tools

Abstract

This study examines how uncertainty communication design affects clinician interpretation, confidence calibration, and decision quality in AI-augmented clinical risk stratification tools. The study developed a design framework that integrates predicted risk probability, risk tier, predictive uncertainty, threshold proximity, and action-oriented review cues into a unified clinical interface. A structured evaluation compared numerical, categorical, and hybrid uncertainty communication formats across standardized clinical risk cases. The risk stratification model produced 184 low-risk cases, 156 moderate-risk cases, and 137 high-risk cases, with stable prediction rates of 92.7%, 76.9%, and 89.8%, respectively. Moderate-risk cases showed the highest review-needed proportion at 23.1%, confirming that uncertainty communication is most critical near decision boundaries. The hybrid format produced the strongest clinician outcomes, achieving 89.7% interpretation accuracy, 91.3% uncertainty comprehension, and 88.5% decision consistency. It also reduced mean response time to 39.8 seconds, compared with 52.4 seconds in the numerical format and 43.7 seconds in the categorical format. Workflow indicators further showed that the hybrid format reached 96.9% case completion and 86.2% review trigger use. Confidence calibration improved substantially under the hybrid format, with the lowest calibration gap of 3.9% and the highest appropriate decision rate of 90.1%. Over-trust events declined from 18 in the numerical format to 5 in the hybrid format, while under-triage events decreased from 14 to 6. These findings indicate that uncertainty-aware interface design can improve clinical risk interpretation without increasing cognitive burden. The study contributes a clinically grounded design model for communicating AI uncertainty as part of patient safety, decision accountability, and human-AI collaboration.