OBJECTIVES: In 2016, Medicare banned direct quality-adjusted life-years (QALY), considering biases without considering disability adjustments and blocking simple transformations of cost-effectiveness analysis (CEA) to willingness-to-pay (WTP). Therefore, Lakdawalla and Phelps proposed the Generalized Risk-Adjusted Cost-Effectiveness (GRACE) framework to cope with the barrier, which was yet to be completed in empirical publications.

We adopted the GRACE framework to validate acute ischemic stroke patients' outcomes through artificial intelligence (AI) interventions to test the difference between evaluation methods based on the nature of the high potential of disability.

METHODS: We applied deidentified data from the Johns Hopkins Health System, including 251 acute ischemic stroke admissions, 17.9% under novel AI interventions. We establish a first-order Markov Model to estimate the Modified Rankin Scale (mRS) of stroke survivors within a 10-year timeframe after deriving matrices of annual health status changes by GRG Nonlinear algorithm.

We dynamically run Propensity Score Matching (NN=3) through characteristics and comorbidities to lower heterogeneous errors, then apply a cyclical structure to update annual mRS with the probability of readmissions. Subsequently, we aggregate results with a 5% discount rate to show a health-loss difference. Additionally, we project differences in terms of QALY and WTP for a net benefit estimation as our reference point. Finally, we measure disability discounts (mRS) and risk preferences in two versions of projected results under the GRACE framework, followed by validation through the caregiver's burden over local median wages.

RESULTS: The structural estimation projects 1.4-2.5 times of net benefit for AI intervention compared to the traditional CEA-WTP combo. Meanwhile, the projected caregiver's burden and procedural costs match the results within a 10% difference after controlling the estimation assumptions.

CONCLUSIONS: The GRACE model showed a major evaluation difference compared to traditional CEA with validation through the caregiver's burden, hence showing an empirical application of GRACE that other medical professionals could follow and adapt.