Clinical AI Trust Half-Life: Time-Decaying Prediction Assurance Under Data and Fairness Drift

Abstract

Clinical artificial intelligence (AI) models are commonly validated at a fixed point, although the conditions that support their predictions can change after deployment. This paper introduces Clinical AI Trust Half-Life, a theoretical measure of the time required for overall prediction assurance to decline to one-half of its initially validated level. Assurance is represented as a normalized composite of predictive performance, calibration, distribution stability, and subgroup fairness, while workflow change and monitoring delay influence the rate and practical consequences of decay. A reproducible simulation trains logistic regression on synthetic baseline data for two patient groups and then evaluates the unchanged model across thirteen periods with gradual covariate drift, increasing outcome prevalence, changing predictor–outcome relationships, and faster degradation in one subgroup. In the executed simulation, assurance decreased from 0.858 at baseline to 0.334 at period 12. The empirical half-life occurred at period 9, when assurance first fell below 0.429; an exponential approximation yielded 8.78 periods. The simulation demonstrates that aggregate discrimination can remain superficially usable while calibration and subgroup sensitivity deteriorate substantially. Trust half-life is proposed as a governance summary, not a substitute for component-level surveillance or clinical validation. It can support risk-adjusted monitoring intervals, early-warning thresholds, subgroup audits, revalidation, recalibration, and explicit assignment of maintenance responsibilities.